Add Batch b35e522b-30ff-4ef6-8663-ecc0a4272e68

retrievalaugmentedmachinetranslationwithunstructuredknowledge/b54ffe29-db93-4126-a191-a7e1bcc795e6_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53068255eebb998ff99c62d845fcd717562372b801ef85474923c10fe000ddf9
+size 117044

retrievalaugmentedmachinetranslationwithunstructuredknowledge/b54ffe29-db93-4126-a191-a7e1bcc795e6_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b7725f4c61921f2ff6f7835e42e97079d190837f980d7924e43d82614529e84
+size 136268

retrievalaugmentedmachinetranslationwithunstructuredknowledge/b54ffe29-db93-4126-a191-a7e1bcc795e6_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b954c7bb8e941e6de2ed55cffb64585e7c906adf846012a2d15f92143c397f6f
+size 769662

retrievalaugmentedmachinetranslationwithunstructuredknowledge/full.md

@@ -0,0 +1,425 @@
+# Retrieval-Augmented Machine Translation with Unstructured Knowledge
+
+Jiaan Wang, Fandong Meng*, Yingxue Zhang, Jie Zhou
+
+Pattern Recognition Center, WeChat AI, Tencent Inc
+
+{torchwang, fandongmeng, yxuezhang, withtomzhou}@tencent.com
+
+# Abstract
+
+Retrieval-augmented generation (RAG) introduces additional information to enhance large language models (LLMs). In machine translation (MT), previous work typically retrieves in-context examples from paired MT corpora, or domain-specific knowledge from knowledge graphs, to enhance MT models. However, a large amount of world knowledge is organized in unstructured documents, and might not be fully paired across different languages. In this paper, we study retrieval-augmented MT using unstructured documents. Specifically, we build RAGtrans, the first benchmark to train and evaluate LLMs' retrieval-augmented MT ability. RAGtrans contains 169K MT samples collected via GPT-4o and human translators. Besides, documents from various languages are also provided to supply the knowledge to these samples. Based on RAGtrans, we further propose a multi-task training method to teach LLMs how to use information from multilingual documents during their translation. The method uses existing multilingual corpora to create auxiliary training objectives without additional labeling requirements. Extensive experiments show that the method improves LLMs by  $1.6\sim 3.1$  BLEU and  $1.0\sim 2.0$  COMET scores in  $\mathrm{En}\Rightarrow \mathrm{Zh}$ , and  $1.7\sim 2.9$  BLEU and  $2.1\sim 2.7$  COMET scores in  $\mathrm{En}\Rightarrow \mathrm{De}$ . We also conclude the critical difficulties that current LLMs face with this task. $^1$
+
+# 1 Introduction
+
+Retrieval-augmented generation (RAG) has grown into a practical paradigm in the development of large language models (LLMs). With the help of retrieved information, LLMs could generate more accurate and knowledge-enrich responses (Li et al., 2022; Gao et al., 2023).
+
+Previous work brought RAG into machine translation (MT), and could be mainly classified into the
+
+following two streams: (1) Retrieving in-context examples (also known as "translation memory"): for a source sentence, a few studies retrieve the relevant paired sentences from bilingual corpora to enhance MT models (Zhang et al., 2018; Bulte and Tezcan, 2019; Khandelwal et al., 2021; He et al., 2021; Hoang et al., 2023). Further, Cai et al. (2021) relax the bilingualism limitation, and try to directly retrieve similar target-language translations to enhance models. (2) Retrieving knowledge triplets: the others retrieve relevant information from knowledge graphs to let the models know domain or cultural knowledge w.r.t. the source sentences (Conia et al., 2024; Chen et al., 2024b). Despite the great success that has been achieved, a large amount of world knowledge is organized in unstructured documents, and might not be fully paired across different languages. This unstructured knowledge is neglected by previous work. For example, Wikipedia serves as an encyclopedia of world knowledge. Most of its information is listed in documents. Besides, for a piece of specific knowledge, Wikipedia does not always provide it in all languages. Though multilingual information of some general knowledge is provided, their content might be differentiated among different languages (Perez-Beltrachini and Lapata, 2021).
+
+In this paper, we study retrieval-augmented MT using unstructured documents. Since we are the first to study this topic and previous datasets do not support the research, we first build a benchmark dataset, named RAGtrans. In detail, RAGtrans is collected based on Wikipedia with three key features: (i) Knowledge-intensive sentences: RAGtrans randomly selects 169K English sentences from Wikipedia as the source sentences, which generally come from the lead paragraphs of different Wikipedia pages, containing knowledge-intensive semantics. Thus, understanding these source sentences tends to require additional knowledge. (ii) Useful relevant documents: To achieve retrieval-
+
+augmented MT, for each source sentence, its following content on the Wikipedia page (in English) could serve as its relevant document. (iii) Transferability to multilingual RAG: Wikipedia also provides multilingual parallel content. Therefore, for a source sentence, its relevant knowledge in different languages can also serve as relevant documents. As a result, MT models can leverage knowledge from multilingual documents beyond the source and the target languages. In this work, we choose Chinese, German, French and Czech. After collecting the source English sentences and relevant documents, we randomly split them into training, validation and testing sets. For training and validation samples, we employ GPT-4o (OpenAI, 2024) to collect the Chinese or German translation; while we employ professional human translators to perform the same process for the testing samples. Finally, RAG-trans involves 79K retrieval-augment  $\mathrm{En} \Rightarrow \mathrm{Zh}$  and  $90\mathrm{K}\mathrm{En} \Rightarrow \mathrm{De}$  MT samples<sup>2</sup>, each of which contains an English source sentence, a document in English, Chinese, German, French or Czech, and the corresponding translation.
+
+Based on RAGtrans, we train LLMs and evaluate their retrieval-augmented MT performance from the following settings: (1) Golden evaluation: providing LLMs with the golden relevant documents during data collection, and testing the translation performance. (2) Robustness evaluation: providing irrelevant documents to test the LLMs' robustness. (3) Full Wiki evaluation: Equipping LLMs with a (multilingual) retriever to first retrieve relevant documents from the whole Wikipedia, and then evaluate their retrieval-augmented MT ability.
+
+Furthermore, during the application phase of a retrieval-augmented MT model, the model might receive multiple documents from various languages. These multilingual documents are not restricted to parallel documents and can convey diverse meanings. In light of this, we propose a multitask training method to enhance LLMs' ability to leverage multilingual knowledge. Specifically, we design three training objectives, i.e., cross-lingual information completion, self-knowledge-enhanced translation and cross-lingual relevance discrimination. Among them, cross-lingual information completion and cross-lingual relevance discrimination train LLMs to refine and judge information from multilingual documents. Self-knowledge-enhanced
+
+translation lets LLMs generate relevant knowledge in various languages for the source sentences, and then perform MT with the help of its multilingual self-knowledge. The multi-task training samples of these objectives can be automatically created from existing multilingual corpora, and do not need any additional labeling costs. Experiments on RAGtrans show that the multi-task training method improves LLMs' ability to leverage relevant knowledge. Using Qwen2.5-7B (Yang et al., 2024) as the backbone, compared with simply instruction-tuning on RAGtrans, the retrieval-augmented MT performance is improved by  $1.6 \sim 3.1$  BLEU and  $1.0 \sim 2.0$  COMET scores in  $\mathrm{En} \Rightarrow \mathrm{Zh}$ , and  $1.7 \sim 2.9$  BLEU and  $2.1 \sim 2.7$  COMET scores in  $\mathrm{En} \Rightarrow \mathrm{De}$ . Finally, we discuss specific challenges that current approaches faced with this task and give multiple promising directions for future research.
+
+Our main contributions are concluded as follows:
+
+- To the best of our knowledge, we are the first to study retrieval-augmented MT using unstructured knowledge. To this end, we construct the first corresponding benchmark dataset, i.e., RAGtrans, containing 169K translation samples collected via GPT-40 and human translators.  
+- We propose a multi-task training method with three designed training objectives to improve LLMs' retrieval-augmented MT ability. The multi-task training samples are low-cost, and do not require additional labeling costs.  
+- In-depth analyses of the retrieval-augmented MT results on automatic evaluation and human evaluation provide a deeper understanding of this research direction.
+
+# 2 RAGtrans
+
+In this section, we first discuss how we select English source sentences and their relevant documents from Wikipedia (§ 2.1). Then, we introduce the details of the data translation via GPT-4o and human translators (§ 2.2). Finally, we give statistical analyses of RAGtrans (§ 2.3), and provide the details of benchmark settings (§ 2.4).
+
+# 2.1 Data Selection
+
+When deciding the source sentences we focus on, there are three requirements that should be met: (1) The source sentences should involve knowledge-intensive semantics, otherwise, they might be trivial to translate and do not need additional knowledge. (2) It should be convenient to collect their
+
+relevant documents from existing resources, otherwise, annotating relevant documents is labor-intensive. (3) It should also be possible to collect relevant documents in other languages. This is because world knowledge is recorded in multilingual form. If we restrict the language of the retrieved documents, the practicality will decrease.
+
+After carefully comparing existing open-source resources, we decide to select both source sentences and relevant documents from Wikipedia. Formally, we denote an English document on a Wikipedia page as  $D^{\mathrm{en}} = \{p_1^{\mathrm{en}}, p_2^{\mathrm{en}}, \dots, p_{|D|}^{\mathrm{en}}\}$ , where  $p_i^{\mathrm{en}}$  indicates the  $i$ -th paragraph in  $D^{\mathrm{en}}$ . Inspired by Perez-Beltrachini and Lapata (2021), the lead paragraph of a Wikipedia page contains knowledge-intensive semantics. Thus, we use  $p_1^{\mathrm{en}}$  from each randomly selected Wikipedia page as a source sentence to meet the requirement (1). In view of the paragraphs on the same Wikipedia page are generally highly related, to meet the requirement (2) for  $p_1^{\mathrm{en}}$ , we randomly select its consecutive paragraphs, i.e.,  $D^{\mathrm{en}} \setminus p_1^{\mathrm{en}}$ , as its relevant document. To further collect relevant documents beyond English, i.e., the requirement (3), we exploit the parallel documents in other languages of  $D^{\mathrm{en}}$  provided by Wikipedia. In this work, we choose Chinese, German, French and Czech, and denote the corresponding parallel documents as  $D^{\mathrm{zh}}$ ,  $D^{\mathrm{de}}$ ,  $D^{\mathrm{fr}}$  and  $D^{\mathrm{cs}}$ , respectively. Given this, the consecutive paragraphs from  $D^l \setminus p_1^l (l \in \{\mathrm{zh}, \mathrm{de}, \mathrm{fr}, \mathrm{cs}\})$  form as the relevant document in other languages.
+
+To ensure robustness, for a small number of source sentences, we randomly select documents from the whole Wikipedia to serve as noisy documents. After the above process, we obtain 169K English source sentences. For each sentence, a (relevant or noisy) document in English, Chinese, German, French or Czech is also provided.
+
+# 2.2 Translation Annotation
+
+For a given source sentence, we next collect its translation in the target language conditioned on the corresponding document. In this work, we focus on  $\mathrm{En} \Rightarrow \mathrm{Zh}$  and  $\mathrm{En} \Rightarrow \mathrm{De}$  translation, and we collect the Chinese translation for the 79K English sentences, while German translation for the remaining 90K English sentences. Considering the trade-off between quality and cost, we decide to translate the source sentences of the training and validation sets via GPT-4o (OpenAI, 2024), while those of the testing set are translated via human translators. GPT-4o translation. Given a source sentence and
+
+![](images/48049676038d9db40819616c1e4ab3b90d5a2e83c08f882e4475ad76e3697cce.jpg)  
+Figure 1: The overview of GPT-4o translation.
+
+the corresponding document, we prompt GPT-4o to perform retrieval-augmented MT to collect its Chinese or German translation. To achieve better translation, we let GPT-4o first judge the relevance of the given document to respond with a judgment and a 5-point rating, and then translate the sentence in a chain-of-thought (CoT) manner. Figure 1 gives a brief overview of the process. We also provide an example of the complete prompt, quality analysis and other details in Appendix A.
+
+Human translation. For source sentences in the testing set, we employ 17 professional human translators to collect the Chinese or German translations. Among them, 10 translators are  $\mathrm{En} \Rightarrow \mathrm{Zh}$  translators while 7 translators are  $\mathrm{En} \Rightarrow \mathrm{De}$  translators. All translators are native Chinese. For  $\mathrm{En} \Rightarrow \mathrm{Zh}$  translators, they major in English, and have passed the English translator qualification. For  $\mathrm{En} \Rightarrow \mathrm{De}$  translators, they major in German, and have passed the German translator qualification. We only provide the source English sentences to the annotators, and encourage them to search for the information they need from Wikipedia. In addition, there are five data reviewers with rich experience in checking translation quality, and  $20\%$  of the sentences translated by each translator are checked by a reviewer. If the translation accuracy is lower than  $95\%$ , the translator needs to modify all his/her translations under the guidance of the reviewer.
+
+Finally, we obtain  $79\mathrm{K}\mathrm{En}\Rightarrow \mathrm{Zh}$  and  $90\mathrm{K}$ $\mathrm{En}\Rightarrow \mathrm{De}$  retrieval-augmented MT samples. Among them,  $77\mathrm{K}\mathrm{En}\Rightarrow \mathrm{Zh}$  and  $88\mathrm{K}\mathrm{En}\Rightarrow \mathrm{De}$  samples from the training and validation sets are translated by the GPT-4o translator. Each sample can be formulated as a triplet  $\langle s,d^l,t\rangle$  , where  $s$  and  $t$  indicate the source English sentence and its translation, respectively.  $d^{l}$  indicates the given document for  $s$  and  $l\in \{\mathrm{en},\mathrm{zh},\mathrm{de},\mathrm{fr},\mathrm{cs}\}$  represents its language. In addition,  $2\mathrm{K}\mathrm{En}\Rightarrow \mathrm{Zh}$  and  $2\mathrm{K}\mathrm{En}\Rightarrow \mathrm{De}$  samples
+
+<table><tr><td colspan="2">Document</td><td colspan="3">En⇒Zh</td><td colspan="3">En⇒De</td></tr><tr><td>Type</td><td>Lang.</td><td>Train</td><td>Valid</td><td>Test</td><td>Train</td><td>Valid</td><td>Test</td></tr><tr><td rowspan="6">Relevant</td><td>En</td><td>19,500</td><td>500</td><td></td><td>19500</td><td>500</td><td></td></tr><tr><td>Zh</td><td>19,500</td><td>500</td><td></td><td>19500</td><td>500</td><td></td></tr><tr><td>De</td><td>9,700</td><td>300</td><td></td><td>19500</td><td>500</td><td></td></tr><tr><td>Fr</td><td>9,700</td><td>300</td><td></td><td>9700</td><td>300</td><td></td></tr><tr><td>Cs</td><td>9,700</td><td>300</td><td rowspan="2">2,000</td><td>9700</td><td>300</td><td rowspan="2">2,000</td></tr><tr><td>En</td><td>1,850</td><td>150</td><td>1850</td><td>150</td></tr><tr><td rowspan="4">Noisy</td><td>Zh</td><td>1,850</td><td>150</td><td></td><td>1850</td><td>150</td><td></td></tr><tr><td>De</td><td>900</td><td>100</td><td></td><td>1850</td><td>150</td><td></td></tr><tr><td>Fr</td><td>900</td><td>100</td><td></td><td>900</td><td>100</td><td></td></tr><tr><td>Cs</td><td>900</td><td>100</td><td></td><td>900</td><td>100</td><td></td></tr><tr><td colspan="2">Total</td><td>74,500</td><td>2,500</td><td>2,000</td><td>85250</td><td>2750</td><td>2000</td></tr></table>
+
+![](images/6939eb546dc01b4885805a302dc6b1e0c5b0bc1e78ab771625a3898f9326c86c.jpg)  
+Figure 2: The distribution of relevance scores.
+
+from the testing set are translated by human translators. For each test sample, we provide the relevant English, Chinese and German documents (derived from the corresponding and parallel Wikipedia documents) in RAGtrans. Thus, a testing sample could be formulated as a quintuple  $\langle s,d^{\mathrm{en}},d^{\mathrm{zh}},d^{\mathrm{de}},t\rangle$
+
+# 2.3 Data Statistics
+
+Table 1 shows the number of samples w.r.t. different types (relevant or noisy) and different languages of the given documents. In the  $\mathrm{En}\Rightarrow \mathrm{Zh}$  training and validation sets,  $8.59\%$  and  $24\%$  of samples are associated with noisy documents. We emphasize the ratio of noisy documents in the validation set since robustness is vital in real applications.  $\mathrm{En}\Rightarrow \mathrm{De}$  samples also show the same tendency. Moreover, for the training and validation samples, GPT-4o also outputs a 5-point rating (named relevance score) w.r.t. the given documents (c.f. right middle section in Figure 1). Taking  $\mathrm{En}\Rightarrow \mathrm{Zh}$  as an example, for samples with relevant documents, we also calculate the distribution of their relevance scores. As shown in Figure 2, more than  $92\%$  of the documents are regarded as "relevant"  $(\geq 3)$  by GPT-4o. For sam
+
+Table 1: The number of retrieval-augment MT samples in RAGtrans w.r.t. different types and different languages (Lang.) of documents.  
+
+<table><tr><td></td><td></td><td>Min.</td><td>Max.</td><td>Avg.</td><td>95th ptcl.</td></tr><tr><td rowspan="7">En⇒Zh</td><td>Source (En)</td><td>5</td><td>526</td><td>85.83</td><td>173</td></tr><tr><td>Target (Zh)</td><td>6</td><td>669</td><td>100.26</td><td>202</td></tr><tr><td>Document (En)</td><td>2</td><td>3,254</td><td>326.08</td><td>874</td></tr><tr><td>Document (Zh)</td><td>2</td><td>4,456</td><td>349.65</td><td>925</td></tr><tr><td>Document (De)</td><td>33</td><td>1,065</td><td>367.43</td><td>791</td></tr><tr><td>Document (Fr)</td><td>29</td><td>3,481</td><td>365.09</td><td>902</td></tr><tr><td>Document (Cs)</td><td>38</td><td>962</td><td>369.84</td><td>769</td></tr><tr><td rowspan="7">En⇒De</td><td>Source (En)</td><td>6</td><td>559</td><td>83.19</td><td>168</td></tr><tr><td>Target (De)</td><td>8</td><td>679</td><td>101.27</td><td>206</td></tr><tr><td>Document (En)</td><td>26</td><td>1155</td><td>322.85</td><td>786</td></tr><tr><td>Document (Zh)</td><td>11</td><td>1446</td><td>339.36</td><td>854</td></tr><tr><td>Document (De)</td><td>31</td><td>1026</td><td>360.68</td><td>784</td></tr><tr><td>Document (Fr)</td><td>29</td><td>3265</td><td>362.31</td><td>888</td></tr><tr><td>Document (Cs)</td><td>41</td><td>943</td><td>369.05</td><td>779</td></tr></table>
+
+Table 2: The minimum (Min.), maximum (Max.), average (Avg.) and 95th percentile (ptcl.) of tokens in the source sentence, target translation, and documents.
+
+plies with noisy documents,  $99.93\%$  (6,995/7,000) of samples are judged as "1", while the remaining  $0.07\%$  (5/7000) are "2".
+
+As for the length of source sentences, target sentences and documents, we use tiktoken<sup>3</sup> to calculate their token-level length. Table 2 shows the minimum, maximum, and average length. 95th percentile of length is also provided. We find that an extremely small number of documents only have single-digit tokens, which should be considered as noises, and we reserve these samples under the robustness consideration.
+
+# 2.4 Benchmark Settings
+
+We design three benchmark settings to evaluate the retrieval-augmented MT models: (1) Golden Evaluation: For each testing sample  $\langle s, d^{\mathrm{en}}, d^{\mathrm{zh}}, d^{\mathrm{de}}, t \rangle$ , we give the source sentence  $(s)$  and a golden relevant document  $(d^{\mathrm{en}} / d^{\mathrm{zh}} / d^{\mathrm{de}})$  to the model, and evaluate models' translation. (2) Robustness Evaluation: We give  $s$  and an irrelevant document (randomly selected from Wikipedia) to the model, and evaluate its translation. (3) Full Wiki Evaluation: This setting equips the MT models with a retriever, and truly tests models' retrieval-augmented MT ability. For a given  $s$ , a retriever should first retrieve relevant documents from the whole Wikipedia, and then input both  $s$  and retrieved documents to the MT model to get translation.
+
+# 3 Multi-Task Training
+
+To further enhance LLMs' retrieval-augmented MT ability, we propose a multi-task training method, named CSC, which involves three designed train
+
+ing objectives, i.e., Cross-lingual information completion, Self-knowledge-enhanced translation and Cross-lingual relevance discrimination. In this section, we first introduce these objectives (§ 3.1) and then discuss how to create their training samples from existing corpora (§ 3.2).
+
+# 3.1 Multi-Task Training Objectives
+
+When developing a retrieval-augment MT model in real applications, it is possible to retrieve information from multilingual knowledge bases for a given source sentence. As a result, the model might receive multiple documents from various languages, extending beyond both the source and target languages. In such a situation, the challenge of effectively refining knowledge from these multilingual documents becomes increasingly significant. To this end, we design three training objectives:
+
+(1) Cross-lingual information completion. Given a multilingual document  $d^{\mathrm{mix}}$  whose paragraphs might be in different languages, and its truncated summary  $\hat{y}$  in one language (e.g., English), we require LLMs to expand  $\hat{y}$  to a complete summary  $y$ . Formally, this objective can be formulated as  $\Theta(y|d^{\mathrm{mix}}, \hat{y})$ , where  $\Theta$  denotes the LLMs.  
+(2) Self-knowledge-enhanced translation. As revealed by recent RAG studies (Wang et al., 2023b; Liu et al., 2024; Asai et al., 2024), RAG models can achieve better performance with the help of their own knowledge. Inspired by this idea, we design self-knowledge-enhanced translation. Specifically, given a source sentence  $s$ , LLMs first generate its relevant document  $\tilde{d}^l$  in a specific language  $l \in \{\text{en}, \text{zh}, \text{de}, \text{fr}, \text{cs}\}$  and then incorporate the document to translate  $s$  to  $t$ , denoted as  $\Theta(t|\tilde{d}^l|s)$ .  
+(3) Cross-lingual relevance discrimination. Given that the retrieved documents may be in various languages, a crucial capability is to assess the relevance between two texts in different languages. To this end, given a document pair  $\langle d^{l_1},d^{l_2}\rangle$  ( $l_{1} \neq l_{2}$ ),  $l_{1}$  and  $l_{2}$  denote the languages of the documents, the model is required to generate the relevance between  $d^{l_1}$  and  $d^{l_2}$ , denotes as  $\mathrm{r}(d^{l_1},d^{l_2})$ . The object can be formulated as  $\Theta (r|d^{l_1},d^{l_2})$
+
+# 3.2 Multi-Task Training Samples
+
+To create the samples for these training objectives, a principle is to reformulate existing corpora instead of labeling new data to ensure scalability.
+
+(1) Cross-lingual information completion. To create the multilingual document  $d^{\mathrm{mix}}$  and its summary  $y$ , we reformulate the Wikipedia corpus. As
+
+revealed by Perez-Beltrachini and Lapata (2021), the lead paragraph in a Wikipedia page could be regarded as its summary. Given this, for an English Wikipedia page  $D^{\mathrm{en}}$ , we extract its lead paragraph (i.e.,  $p_1^{\mathrm{en}}$ ) as  $y$ , and randomly truncate  $y$  to  $\hat{y}$ . We next construct  $d^{\mathrm{mix}}$  from the remaining paragraphs  $\hat{D}^{\mathrm{en}} = \{p_i^{\mathrm{en}} | i \geq 2\}$ , and the parallel counterparts in other languages, i.e.,  $\hat{D}^{\mathrm{zh}}$ ,  $\hat{D}^{\mathrm{de}}$ ,  $\hat{D}^{\mathrm{fr}}$  and  $\hat{D}^{\mathrm{cs}}$  in this work. Since there might be redundant information across parallel paragraphs, we use MMR algorithm (Carbonell and Goldstein, 1998) to select paragraphs from these multilingual paragraphs, i.e.,  $\bigcup_{l} \hat{D}^{l}$ , to form  $d^{\mathrm{mix}}$ . MMR is a statistical algorithm that iteratively selects key paragraphs from the given document, at each selection step, it evaluates the relevance and redundancy of the unselected paragraphs in relation to the selected ones to determine which paragraph to select in that step.
+
+(2) Self-knowledge-enhanced translation. We reformulate previous multilingual MT corpora to create samples. In detail, we use TED talk corpus (Aharoni et al., 2019), where each sentence is provided with multilingual parallel sentences. For an English sentence  $s^{\mathrm{en}}$ , we input the sentence or its parallel sentences in other languages (i.e.,  $s^l$ ) to a LLM  $\Theta$ , and prompt  $\Theta$  to generate its relevant knowledge in the corresponding languages, i.e.,  $\tilde{d}^l$ . In this way,  $\tilde{d}^l$  could be used as a relevant document to translate  $s^{\mathrm{en}}$  to other languages.  
+(3) Cross-lingual relevance discrimination. We reformulate the parallel Wikipedia documents to create the samples. Intuitively, randomly selected paragraphs from two parallel documents are relevant; while those from different documents are irrelevant. In this way, we create the document pair and the corresponding boolean relevance.
+
+# 4 Experiments
+
+# 4.1 Experimental Setup
+
+Metrics. Following previous work, we adopt BLEU (Papineni et al., 2002) and reference-based COMET score (Rei et al., 2022). BLEU evaluates n-grams overlap between the generated translations and corresponding references, while COMET evaluates the semantic similarity of translations against references. Besides, recent studies (Kocmi and Federmann, 2023; Wang et al., 2023a) also show the strong ability of LLMs in NLP evaluation. Thus, we use evaluators implemented using GPT-4o in reference-based and reference-free styles, which we refer to as GRB and GRF, respectively.
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="3">English⇒Chinese (En⇒Zh)</td><td colspan="3">English⇒German (En⇒De)</td></tr><tr><td>Zero-Shot LLMs</td><td>SFT LLMs</td><td>SFT+CSC LLMs</td><td>Zero-Shot LLMs</td><td>SFT LLMs</td><td>SFT+CSC LLMs</td></tr><tr><td colspan="8">w/ Empty Document</td></tr><tr><td>1</td><td>Qwen2.5-7B</td><td>50.1 / 84.6 / 85.5 / 87.0</td><td>54.6 / 86.7 / 88.5 / 89.1</td><td>56.8 / 87.7 / 91.2 / 91.8</td><td>56.1 / 66.7 / 74.7 / 78.1</td><td>59.2 / 69.7 / 81.5 / 83.4</td><td>60.9 / 71.8 / 84.4 / 86.9</td></tr><tr><td>2</td><td>Qwen2.5-14B</td><td>51.3 / 84.7 / 86.2 / 87.7</td><td>55.4 / 87.0 / 89.1 / 89.4</td><td>57.6 / 87.9 / 91.6 / 92.0</td><td>57.6 / 67.3 / 79.1 / 81.3</td><td>60.2 / 70.6 / 82.1 / 84.9</td><td>61.6 / 72.0 / 85.0 / 87.5</td></tr><tr><td>3</td><td>LLama-3-8B</td><td>40.9 / 77.5 / 80.4 / 78.1</td><td>54.4 / 86.6 / 88.1 / 88.2</td><td>56.1 / 87.6 / 90.3 / 91.1</td><td>59.1 / 67.4 / 81.4 / 85.0</td><td>60.1 / 70.6 / 82.6 / 85.2</td><td>61.6 / 72.0 / 85.0 / 87.8</td></tr><tr><td>4</td><td>Mistral-7B</td><td>38.1 / 76.4 / 79.8 / 77.6</td><td>53.6 / 86.3 / 87.7 / 88.1</td><td>54.7 / 87.2 / 90.1 / 90.5</td><td>46.8 / 60.1 / 67.1 / 72.2</td><td>59.7 / 70.2 / 81.7 / 84.0</td><td>61.3 / 71.9 / 82.2 / 85.4</td></tr><tr><td colspan="8">w/ Noisy Document</td></tr><tr><td>5</td><td>Qwen2.5-7B</td><td>48.6 / 83.4 / 84.6 / 86.1</td><td>54.7 / 86.7 / 88.5 / 89.1</td><td>56.5 / 87.7 / 91.1 / 91.8</td><td>55.2 / 66.0 / 74.0 / 76.9</td><td>59.0 / 69.7 / 81.2 / 83.0</td><td>61.1 / 71.9 / 84.2 / 86.1</td></tr><tr><td>6</td><td>Qwen2.5-14B</td><td>49.4 / 83.7 / 84.9 / 86.5</td><td>55.5 / 87.0 / 89.2 / 89.4</td><td>57.7 / 87.9 / 91.6 / 92.1</td><td>56.3 / 66.2 / 77.8 / 80.2</td><td>60.2 / 70.6 / 81.9 / 84.6</td><td>61.6 / 72.2 / 85.1 / 87.8</td></tr><tr><td>7</td><td>LLama-3-8B</td><td>26.3 / 74.6 / 78.2 / 78.9</td><td>54.3 / 86.6 / 88.0 / 88.2</td><td>56.1 / 87.6 / 90.3 / 91.1</td><td>57.9 / 66.5 / 79.7 / 81.5</td><td>60.2 / 70.7 / 82.8 / 85.2</td><td>61.5 / 72.0 / 84.9 / 87.5</td></tr><tr><td>8</td><td>Mistral-7B</td><td>24.3 / 73.9 / 77.7 / 78.0</td><td>53.5 / 86.3 / 87.7 / 88.1</td><td>54.5 / 87.1 / 90.0 / 90.5</td><td>41.9 / 59.3 / 62.4 / 65.7</td><td>59.5 / 70.1 / 81.5 / 83.9</td><td>61.5 / 72.0 / 82.4 / 85.6</td></tr><tr><td colspan="8">w/ Golden English Document</td></tr><tr><td>9</td><td>Qwen2.5-7B</td><td>49.6 / 84.3 / 85.1 / 86.5</td><td>57.1 / 87.7 / 90.6 / 91.3</td><td>58.7 / 88.7 / 92.5 / 93.1</td><td>55.4 / 66.2 / 73.5 / 77.4</td><td>61.2 / 71.3 / 83.1 / 85.6</td><td>63.5 / 73.4 / 84.6 / 87.9</td></tr><tr><td>10</td><td>Qwen2.5-14B</td><td>50.6 / 84.5 / 85.5 / 87.1</td><td>57.0 / 87.9 / 91.1 / 91.8</td><td>59.4 / 88.9 / 92.7 / 93.3</td><td>57.2 / 67.3 / 78.2 / 81.0</td><td>61.5 / 71.7 / 83.9 / 86.0</td><td>63.8 / 73.7 / 87.0 / 88.2</td></tr><tr><td>11</td><td>LLama-3-8B</td><td>29.4 / 75.3 / 80.2 / 80.9</td><td>55.8 / 87.6 / 89.7 / 90.4</td><td>58.0 / 88.6 / 91.4 / 92.0</td><td>58.7 / 67.2 / 80.9 / 85.3</td><td>61.7 / 71.7 / 84.0 / 86.5</td><td>64.2 / 73.6 / 87.8 / 89.1</td></tr><tr><td>12</td><td>Mistral-7B</td><td>26.6 / 74.8 / 79.8 / 80.3</td><td>54.9 / 87.3 / 89.5 / 89.8</td><td>56.6 / 88.1 / 91.1 / 91.9</td><td>42.3 / 60.2 / 66.7 / 69.6</td><td>61.4 / 71.5 / 83.5 / 85.8</td><td>63.4 / 72.9 / 85.8 / 87.9</td></tr><tr><td colspan="8">w/ Golden Chinese Document</td></tr><tr><td>13</td><td>Qwen2.5-7B</td><td>49.9 / 84.4 / 85.4 / 86.7</td><td>57.3 / 87.1 / 91.2 / 91.5</td><td>60.0 / 89.1 / 93.1 / 93.6</td><td>54.5 / 65.2 / 72.2 / 74.9</td><td>60.3 / 70.4 / 82.3 / 84.2</td><td>63.2 / 73.1 / 84.3 / 87.3</td></tr><tr><td>14</td><td>Qwen2.5-14B</td><td>50.7 / 84.7 / 85.7 / 87.0</td><td>58.1 / 87.2 / 91.7 / 92.0</td><td>60.5 / 89.2 / 93.3 / 93.7</td><td>55.8 / 66.2 / 77.4 / 80.3</td><td>61.2 / 71.2 / 83.0 / 85.3</td><td>63.8 / 73.5 / 87.1 / 88.3</td></tr><tr><td>15</td><td>LLama-3-8B</td><td>35.7 / 76.4 / 80.8 / 81.0</td><td>57.1 / 87.0 / 90.2 / 90.8</td><td>59.5 / 89.0 / 92.8 / 93.2</td><td>57.0 / 66.7 / 79.3 / 83.1</td><td>61.1 / 71.0 / 82.7 / 84.9</td><td>63.8 / 73.2 / 87.5 / 89.3</td></tr><tr><td>16</td><td>Mistral-7B</td><td>34.9 / 75.2 / 79.8 / 80.4</td><td>56.4 / 87.8 / 89.9 / 90.3</td><td>58.8 / 88.8 / 92.6 / 93.0</td><td>41.8 / 59.0 / 65.8 / 68.2</td><td>60.7 / 70.8 / 82.5 / 84.6</td><td>63.3 / 73.1 / 85.9 / 88.2</td></tr><tr><td colspan="8">w/ Golden German Document</td></tr><tr><td>17</td><td>Qwen2.5-7B</td><td>44.2 / 83.2 / 84.3 / 85.1</td><td>55.6 / 87.2 / 88.9 / 89.6</td><td>58.7 / 88.7 / 92.4 / 93.1</td><td>55.8 / 66.4 / 76.9 / 77.8</td><td>62.0 / 72.3 / 83.6 / 85.2</td><td>64.5 / 74.5 / 85.9 / 90.0</td></tr><tr><td>18</td><td>Qwen2.5-14B</td><td>45.2 / 82.9 / 84.8 / 85.6</td><td>56.3 / 87.5 / 89.4 / 89.7</td><td>59.4 / 88.9 / 92.7 / 93.1</td><td>57.4 / 67.5 / 78.6 / 80.9</td><td>62.6 / 72.7 / 84.2 / 85.7</td><td>65.1 / 74.7 / 88.2 / 90.4</td></tr><tr><td>19</td><td>LLama-3-8B</td><td>35.3 / 74.8 / 78.5 / 78.9</td><td>55.1 / 87.1 / 88.6 / 89.1</td><td>57.7 / 88.6 / 91.3 / 91.7</td><td>60.0 / 67.5 / 81.7 / 84.9</td><td>62.6 / 72.5 / 84.5 / 85.8</td><td>65.0 / 74.6 / 88.4 / 90.0</td></tr><tr><td>20</td><td>Mistral-7B</td><td>35.0 / 73.3 / 77.3 / 76.5</td><td>54.3 / 86.8 / 88.3 / 88.8</td><td>56.6 / 88.1 / 91.0 / 91.7</td><td>46.1 / 59.6 / 66.4 / 71.8</td><td>61.9 / 72.0 / 83.5 / 84.8</td><td>64.6 / 74.0 / 87.8 / 88.8</td></tr></table>
+
+Table 3: Experimental results (BLEU / COMET / GRB / GRF) on RAGtrans. "SFT LLMs" denotes the LLMs are instruction-tuned on the training data of RAGtrans, while "SFT+CSC LLMs" denotes the LLMs are instruction-tuned on both RAGtrans and CSC multi-task training. All results of SFT+CSC LLMs are statistically significantly better than those of the corresponding SFT LLMs with t-test  $p < 0.01$ .
+
+Backbones. We adopt four LLMs in the experiments: (1) Qwen2.5-7B-Instruct and (2) Qwen2.5-14B-Instruct (Yang et al., 2024) are two cutting-edge Qwen-series LLMs. (3) Llama-3-8B (Dubey et al., 2024) is the latest llama-series LLM. (4) Mistral-7B (Jiang et al., 2023) also shows great performance among the same-scale LLMs.
+
+Retriever. To support the full Wiki evaluation in RAGtrans, we implement two retrievers in the experiments: (1) BM25 (Robertson et al., 2009) is a traditional lexical search method that matches keywords efficiently with an inverted index. For a given source sentence, BM25 can retrieve its relevant documents only in the same language. (2) BGE-m3 (Chen et al., 2024a) is a multilingual sentence embedding model that supports dense retrievals across different languages.
+
+Implementation Details. For details about training hyper-parameters, SFT prompts, model checkpoints, training costs, CSC training samples and metric implementation, please refer to Appendix B.
+
+# 4.2 Main Results
+
+Table 3 shows the main results of the golden and robustness evaluation settings. For each LLM, we evaluate its retrieval-augmented MT performance when giving empty, noisy or golden documents.
+
+Zero-Shot Performance. Among all backbones,
+
+Qwen2.5-14B typically performs best in  $\mathrm{En} \Rightarrow \mathrm{Zh}$  while Llama-3-8B performs best in  $\mathrm{En} \Rightarrow \mathrm{De}$ . When giving noisy documents to LLMs, the MT performance of all LLMs decreases compared with those of giving empty documents. For example, Qwen2.5-7B (w/ empty document) achieves 50.1 BLEU and 84.6 COMET in  $\mathrm{En} \Rightarrow \mathrm{Zh}$ , while the counterparts of Qwen2.5-7B (w/ noisy document) are 48.6 and 83.4. This observation indicates the low robustness of zero-shot LLMs when faced with irrelevant documents. Moreover, when zero-shot LLMs use golden relevant documents as inputs, their MT performances do not increase (compared with those using empty documents) as expected. Specifically, the performance either decreased or remained relatively stable (rows 9-20 vs. rows 1-4). Besides, when given German documents in  $\mathrm{En} \Rightarrow \mathrm{Zh}$  translation, or given Chinese documents in  $\mathrm{En} \Rightarrow \mathrm{De}$  translation, the model performance decreases significantly. Thus, the retrieval-augmented MT ability is limited in zero-shot LLMs, especially when the retrieved documents are in a language beyond the source and the target languages (named a third language).
+
+Instruction-Tuning Performance. After we tune LLMs on RAGtrans, their MT performance generally increases by a large margin. For example, when giving empty documents in  $\mathrm{En} \Rightarrow \mathrm{Zh}$ , SFT
+
+Qwen2.5-7B outperforms zero-shot Qwen2.5-7B by 4.5 BLEU, 2.1 COMET, 3.0 GRB and 2.1 GRF. Similarly, in  $\mathrm{En}\Rightarrow \mathrm{De}$ , SFT Qwen2.5-7B outperforms zero-shot Qwen2.5-7B by 3.1 BLEU, 3.0 COMET, 6.8 GRB and 5.3 GRF. In addition, we find that when using golden documents, SFT LLMs achieve better performance than those using empty documents, indicating that instruct-tuning on RAGtrans improves LLMs' retrieval-augmented MT ability. Even when giving the relevant documents in a third language, it can bring improvement. The model robustness is also enhanced, and the given noisy documents do not significantly perturb the model performance. This is because a small number of training samples in RAGtrans consist of irrelevant documents as inputs, thus, LLMs can learn to translate source sentences conditioned on both relevant and irrelevant documents.
+
+CSC Training Performance. After instruction-tuning on RAGtrans and CSC multi-task training, the model performance is further improved. When giving empty documents, CSC brings  $1.1\sim 2.2$  BLEU and  $0.9\sim 1.0$  COMET improvements compared with SFT LLMs in  $\mathrm{En}\Rightarrow \mathrm{Zh}$ , while bringing  $1.4\sim 1.7$  BLEU and  $1.4\sim 2.1$ COMET improvements compared with SFT LLMs in  $\mathrm{En}\Rightarrow \mathrm{De}$  (rows 1-4). This observation verifies the effectiveness of CSC, and LLMs' retrieval-augmented MT ability can be enhanced by the designed training objectives. Besides, taking Qwen2.5-14B as an example, when giving the relevant documents in a third language to SFT Qwen2.5-14B, it brings 0.9 BLEU and 0.5 COMET improvements in  $\mathrm{En}\Rightarrow \mathrm{Zh}$  (row 18 vs. row 2) and 1.0 BLEU and 0.6 COMET in  $\mathrm{En}\Rightarrow \mathrm{De}$  (row 14 vs. row 2) compared with when giving empty documents. The counterpart improvements in SFT+CSC Qwen2.5-14B are 1.8 BLEU and 1.0 COMET in  $\mathrm{En}\Rightarrow \mathrm{Zh}$ , and 2.2 BLEU and 1.5 COMET in  $\mathrm{En}\Rightarrow \mathrm{De}$ . Therefore, CSC enhances LLMs' ability to leverage relevant knowledge in a third language. Moreover, we discuss the scalability of CSC in Appendix C.
+
+# 4.3 Full Wiki Evaluation
+
+Table 4 shows the experimental results on full Wiki evaluation. We use the Wikipedia dumps in different languages as the knowledge sources for retrievers to retrieve relevant documents, and then leverage SFT+CSC LLMs to translate the source sentences (please refer to Appendix D for more details). Compared with using empty documents, retrieving documents from knowledge sources gen
+
+<table><tr><td>#</td><td>Knowledge</td><td>Method</td><td>En⇒Zh</td><td>En⇒De</td></tr><tr><td>1</td><td>Empty</td><td>Qwen2.5-7B</td><td>56.8 / 87.7</td><td>60.9 / 71.8</td></tr><tr><td>2</td><td>Document</td><td>Qwen2.5-14B</td><td>57.6 / 87.9</td><td>61.6 / 72.0</td></tr><tr><td>3</td><td>English</td><td>Qwen2.5-7B (+BM25)</td><td>57.4 / 88.0</td><td>61.9 / 72.2</td></tr><tr><td>4</td><td>Wikipedia</td><td>Qwen2.5-14B (+BM25)</td><td>58.0 / 88.2</td><td>62.3 / 72.4</td></tr><tr><td>5</td><td>English</td><td>Qwen2.5-7B (+BGEm3)</td><td>57.9 / 88.1</td><td>62.6 / 72.5</td></tr><tr><td>6</td><td>Wikipedia</td><td>Qwen2.5-14B (+BGEm3)</td><td>58.7 / 88.4</td><td>63.0 / 72.7</td></tr><tr><td>7</td><td>Chinese</td><td>Qwen2.5-7B (+BGEm3)</td><td>58.3 / 88.3</td><td>62.4 / 72.5</td></tr><tr><td>8</td><td>Wikipedia</td><td>Qwen2.5-14B (+BGEm3)</td><td>59.0 / 88.5</td><td>62.8 / 72.6</td></tr><tr><td>9</td><td>German</td><td>Qwen2.5-7B (+BGEm3)</td><td>57.9 / 88.1</td><td>63.2 / 73.1</td></tr><tr><td>10</td><td>Wikipedia</td><td>Qwen2.5-14B (+BGEm3)</td><td>58.7 / 88.3</td><td>63.6 / 73.4</td></tr><tr><td>11</td><td>French</td><td>Qwen2.5-7B (+BGEm3)</td><td>57.6 / 88.0</td><td>62.5 / 72.4</td></tr><tr><td>12</td><td>Wikipedia</td><td>Qwen2.5-14B (+BGEm3)</td><td>58.3 / 88.3</td><td>62.7 / 72.6</td></tr><tr><td>13</td><td>Czech</td><td>Qwen2.5-7B (+BGEm3)</td><td>57.3 / 87.9</td><td>61.8 / 72.2</td></tr><tr><td>14</td><td>Wikipedia</td><td>Qwen2.5-14B (+BGEm3)</td><td>57.9 / 88.1</td><td>62.4 / 72.5</td></tr></table>
+
+Table 4: Experimental results of full wiki evaluation on SFT+CSC LLMs (BLEU / COMET).  
+
+<table><tr><td></td><td colspan="2">BLEU w/ Empty Doc.</td><td colspan="2">BLEU w/ Golden En.</td></tr><tr><td>Qwen2.5-7B (SFT+CSC)</td><td>56.8</td><td>87.7</td><td>58.7</td><td>88.7</td></tr><tr><td>- CLIC</td><td>56.5</td><td>87.6</td><td>57.9</td><td>88.3</td></tr><tr><td>- SKET</td><td>56.1</td><td>87.5</td><td>57.5</td><td>88.1</td></tr><tr><td>- CLRD</td><td>56.5</td><td>87.7</td><td>58.3</td><td>88.5</td></tr><tr><td></td><td colspan="2">w/ Golden Zh.</td><td colspan="2">w/ Golden De.</td></tr><tr><td>Qwen2.5-7B (SFT+CSC)</td><td>60.0</td><td>89.1</td><td>58.7</td><td>88.7</td></tr><tr><td>- CLIC</td><td>59.5</td><td>88.9</td><td>57.0</td><td>88.2</td></tr><tr><td>- SKET</td><td>58.7</td><td>88.5</td><td>57.9</td><td>88.5</td></tr><tr><td>- CLRD</td><td>59.7</td><td>89.0</td><td>56.7</td><td>87.9</td></tr></table>
+
+Table 5: Ablation study on  $\mathrm{En}\Rightarrow \mathrm{Zh}$  golden evaluation. Doc.: Document; "En.", "Zh." and "De." indicate English, Chinese and German documents, respectively.
+
+erally brings improvement, indicating the usability of retrieved knowledge. Compared with the BM25 retriever, BGEm3 retriever helps LLMs achieve better performance (rows 5-6 vs. rows 3-4). Besides, BGEm3 could retrieve knowledge from other languages, and the retrieved knowledge from a third language could also enhance model performance, verifying the SFT+CSC LLMs could leverage multilingual knowledge in retrieval-augmented MT.
+
+# 4.4 Ablations
+
+As shown in Table 5, we conduct ablation studies on  $\mathrm{En} \Rightarrow \mathrm{Zh}$  golden evaluation<sup>4</sup> to figure out the contributions of each training objective in CSC. Specifically, we remove each objective, and evaluate the model performance accordingly. In each case, the performance is lower than using all training objectives, indicating the effectiveness of every objective. When giving empty documents or documents in source/target language, the most important objective is self-knowledge-enhanced transla
+
+<table><tr><td colspan="3">Method</td><td colspan="5">Error Type (%)</td></tr><tr><td>#</td><td>Model</td><td>Document</td><td>Ref.</td><td>Word</td><td>Phrase</td><td>Fluency</td><td>Other</td></tr><tr><td>1</td><td>Qwen2.5-7B (SFT)</td><td>Empty</td><td>0.50</td><td>7.00</td><td>3.33</td><td>1.50</td><td>0.50</td></tr><tr><td>2</td><td>Qwen2.5-14B (SFT)</td><td>Empty</td><td>0.33</td><td>5.50</td><td>2.67</td><td>1.33</td><td>0.50</td></tr><tr><td>3</td><td>Qwen2.5-7B (SFT+CSC)</td><td>Empty</td><td>0.50</td><td>5.50</td><td>2.67</td><td>1.50</td><td>0.50</td></tr><tr><td>4</td><td>Qwen2.5-14B (SFT+CSC)</td><td>Empty</td><td>0.50</td><td>4.67</td><td>2.33</td><td>1.00</td><td>0.17</td></tr><tr><td>5</td><td>Qwen2.5-7B (SFT)</td><td>Golden Zh.</td><td>0.33</td><td>5.17</td><td>2.50</td><td>1.33</td><td>0.50</td></tr><tr><td>6</td><td>Qwen2.5-14B (SFT)</td><td>Golden Zh.</td><td>0.17</td><td>4.67</td><td>2.33</td><td>1.33</td><td>0.00</td></tr><tr><td>7</td><td>Qwen2.5-7B (SFT+CSC)</td><td>Golden Zh.</td><td>0.17</td><td>5.00</td><td>2.17</td><td>1.50</td><td>0.33</td></tr><tr><td>8</td><td>Qwen2.5-14B (SFT+CSC)</td><td>Golden Zh.</td><td>0.17</td><td>3.50</td><td>1.50</td><td>1.33</td><td>0.33</td></tr></table>
+
+tion (SKET) since it enhances models' MT ability. When giving documents in a third language, cross-lingual information completion (CLIC) and cross-lingual relevance discrimination (CLRD) become more important than SKET, since these two objectives train LLMs to refine and judge information from multilingual knowledge.
+
+# 4.5 Human Evaluation
+
+Retrieval-Augmented MT Ability. We employ human evaluation to further study the MT performance of SFT LLMs and SFT+CSC LLMs. Specifically, human evaluators judge whether the translations include the following flaws: reference errors, word-level errors, phrase-level errors, fluency flaws and other errors (more details are given in Appendix F). As shown in Table 6, the two most common error types are word-level and phrase-level errors. In these two types, the CSC multi-task training method and the golden documents enhance the model performance, verifying the effectiveness of CSC and the usability of the golden documents.
+
+The Effects of Documents. To study the effects of documents, we provide the different documents to Qwen2.5-7B (SFT+CSC), including empty, noisy, golden English, golden Chinese and golden German documents, to evaluate if the corresponding translations involve flaws. As shown in Table 7, the noisy documents still increase the number of translation flaws (row 2 vs. row 1). Robustness is a crucial factor for the deployment of LLMs in real applications, thus future work could pay more attention to model robustness. Besides, we also find that when providing the model with golden Chinese or English documents, the number of translation flaws typically decreases. However, when providing golden German documents, the number of word-level errors significantly increases (row 5 vs. row 1). We further observe the cases, and find this is because the German documents might encourage MT models to translate some entities to German if the entities listed in the documents, thus
+
+Table 6: Human evaluation (En  $\Rightarrow$  Zh) of the retrieval-augmented MT ability. Ref.: Reference.  
+
+<table><tr><td rowspan="2">#</td><td rowspan="2">Document</td><td colspan="5">Error Type (%)</td></tr><tr><td>Ref.</td><td>Word</td><td>Phrase</td><td>Fluency</td><td>Other</td></tr><tr><td>1</td><td>Empty</td><td>0.50</td><td>5.50</td><td>2.67</td><td>1.50</td><td>0.50</td></tr><tr><td>2</td><td>Noisy</td><td>1.83</td><td>6.67</td><td>5.33</td><td>1.83</td><td>1.50</td></tr><tr><td>3</td><td>Golden Chinese</td><td>0.17</td><td>5.00</td><td>2.17</td><td>1.50</td><td>0.33</td></tr><tr><td>4</td><td>Golden English</td><td>0.17</td><td>3.17</td><td>1.33</td><td>0.83</td><td>0.17</td></tr><tr><td>5</td><td>Golden German</td><td>0.50</td><td>8.17</td><td>1.67</td><td>1.33</td><td>0.33</td></tr></table>
+
+Table 7: Human evaluation (En  $\Rightarrow$  Zh) of the effects of documents, using Qwen2.5-7B (SFT+CSC) as the MT model.
+
+raising word-level errors. This issue should also be noticed in future work, since the retrieval-argument MT models might receive documents from various languages in real applications.
+
+# 5 Related Work
+
+To leverage additional knowledge to enhance MT performance, previous literature typically explores paired sentences (also known as "translation memory") or structured knowledge graphs as the knowledge sources: (1) Paired Sentences: Zhang et al. (2018) utilize a search engine to retrieve sentence pairs whose source sides are similar to the input sentences. Bulte and Tezcan (2019) design a fuzzy retriever to enhance the model performance. He et al. (2021) design a fast and accurate method to improve the robustness of pair-sentence-enhanced MT models. Cai et al. (2021) relax the bilingualism limitation in retrieving paired sentences, and they try to retrieve similar target-language sentences to enhance MT models. (2) Knowledge Graphs: A few studies leverage relevant information from structured knowledge graphs to enhance MT models. Conia et al. (2024) use Wikidata (Vrandecic and Krötzsch, 2014), a multilingual knowledge graphs, to enhance MT models. Chen et al. (2024b) build an internal knowledge graph based on context, and then use it to enhance translation. Different from previous work, we aim to utilize unstructured documents to provide knowledge to MT models.
+
+# 6 Conclusion
+
+In this paper, we explore the retrieval-augmented MT with unstructured knowledge. To this end, we build RAGtrans dataset with 169K retrieval-augment MT samples to train and evaluate LLMs' retrieval-augmented MT ability. Further, we propose CSC multi-task training method with three designed objectives to teach LLMs to leverage multilingual knowledge in retrieval-augmented MT. Extensive experiments demonstrate the usability of
+
+RAGtrans and the effectiveness of CSC.
+
+# Limitations
+
+While we show LLMs' retrieval-augmented MT ability and the effectiveness of CSC multi-task training method, there are some limitations worth noting: (1) For multilingual knowledge bases, we use Wikipedia in some specific languages (e.g., Chinese, English, German, French, and Czech). Future work could extend the multilingual sources to other languages or other sources. (2) During data collection of RAGtrans, a CoT prompt is used in the GPT-4o translation (c.f. Figure 1). However, in the SFT process, we do not use the CoT prompt to train LLMs, and future work could explore the effect of CoT in retrieval-augmented MT.
+
+# Ethical Considerations
+
+We discuss the main ethical considerations of RAGtrans as follows: (1) Licenses. The source sentences and documents are derived from Wikipedia, whose texts are under CC BY-SA and GFDL licenses. We will release the RAGtrans dataset under CC-BY-SA 4.0 license. (2) Compensation. During the translation annotation, the salary for translating each sentence is determined by the average time of annotation and local labor compensation standard.
+
+# References
+
+Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-RAG: Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations.  
+Bram Bulte and Arda Tezcan. 2019. Neural fuzzy repair: Integrating fuzzy matches into neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1800-1809, Florence, Italy. Association for Computational Linguistics.  
+Deng Cai, Yan Wang, Huayang Li, Wai Lam, and Lemao Liu. 2021. Neural machine translation with monolingual translation memory. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International
+
+Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7307-7318, Online. Association for Computational Linguistics.  
+Jaime G Carbonell and Jade Goldstein. 1998. The use of mmr and diversity-based reranking for reordering documents and producing summaries.(1998). *Citado*, 4:24-42.  
+Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024a. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216.  
+Meiqi Chen, Fandong Meng, Yingxue Zhang, Yan Zhang, and Jie Zhou. 2024b. Crat: A multi-agent framework for causality-enhanced reflective and retrieval-augmented translation with large language models. arXiv preprint arXiv:2410.21067.  
+Simone Conia, Daniel Lee, Min Li, Umar Farooq Minhas, Saloni Potdar, and Yunyao Li. 2024. Towards cross-cultural machine translation with retrieval-augmented generation from multilingual knowledge graphs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16343-16360, Miami, Florida, USA. Association for Computational Linguistics.  
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.  
+Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997.  
+Qiuxiang He, Guoping Huang, Qu Cui, Li Li, and Lemao Liu. 2021. Fast and accurate neural machine translation with translation memory. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3170-3180, Online. Association for Computational Linguistics.  
+Cuong Hoang, Devendra Sachan, Prashant Mathur, Brian Thompson, and Marcello Federico. 2023. Improving retrieval augmented neural machine translation by controlling source and fuzzy-match interactions. In Findings of the Association for Computational Linguistics: EACL 2023, pages 289-295, Dubrovnik, Croatia. Association for Computational Linguistics.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego
+
+de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769-6781.  
+Urvashi Khandelwal, Angela Fan, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2021. Nearest neighbor machine translation. In International Conference on Learning Representations.  
+Tom Kocmi and Christian Federmann. 2023. Large language models are state-of-the-art evaluators of translation quality. In 24th Annual Conference of the European Association for Machine Translation, page 193.  
+Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. 2022. A survey on retrieval-augmented text generation. arXiv preprint arXiv:2202.01110.  
+Yanming Liu, Xinyue Peng, Xuhong Zhang, Weihao Liu, Jianwei Yin, Jiannan Cao, and Tianyu Du. 2024. Ra-ishf: Learning to answer and understand from retrieval augmentation via iterative self-feedback. arXiv preprint arXiv:2403.06840.  
+OpenAI. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Laura Perez-Beltrachini and Mirella Lapata. 2021. Models and datasets for cross-lingual summarisation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9408-9423, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. 2020. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 3505-3506.  
+Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte Alves, Luisa Coheur, Alon Lavie, and André F. T.
+
+Martins. 2022. CometKiwi: IST-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 634-645, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.  
+Denny Vrandecic and Markus Krötzsch. 2014. Wiki-data: a free collaborative knowledgebase. Communications of the ACM, 57(10):78-85.  
+Jiaan Wang, Yunlong Liang, Fandong Meng, Zengkui Sun, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. 2023a. Is ChatGPT a good NLG evaluator? a preliminary study. In Proceedings of the 4th New Frontiers in Summarization Workshop, pages 1-11, Singapore. Association for Computational Linguistics.  
+Yile Wang, Peng Li, Maosong Sun, and Yang Liu. 2023b. Self-knowledge guided retrieval augmentation for large language models. arXiv preprint arXiv:2310.05002.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+Jingyi Zhang, Masao Utiyama, Eiichro Sumita, Graham Neubig, and Satoshi Nakamura. 2018. Guiding neural machine translation with retrieved translation pieces. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1325-1335, New Orleans, Louisiana. Association for Computational Linguistics.  
+Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified efficient fine-tuning of  $100+$  language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400-410, Bangkok, Thailand. Association for Computational Linguistics.
+
+# A Details of GPT-4o translation
+
+Complete Prompt. We provide the system prompt as follows: "You are a professional translator, and your task is to translate a given input sentence from English to Chinese/German. In addition to the input sentence, you will be provided with a document that may contain relevant information to aid in the translation. However, be aware that some documents may contain irrelevant or noisy information". An  $\mathrm{En}\Rightarrow \mathrm{Zh}$  example of user prompt and
+
+# PROMPT
+
+![](images/8f936782603caeac974e8381a63348d09789b20d61ed1d793f1898992dd8cc82.jpg)  
+Figure 3: An  $\mathrm{En}\Rightarrow \mathrm{Zh}$  example of the complete prompt in GPT-4o translator.
+
+model response is shown in Figure 3, where both a (French) document and an English source sentence are provided in the user prompt. We also define a 5-point rating breakdown to align the scoring value between GPT-4o and humans. In the model response, GPT-4o first judges the relevance between the given document and the source sentence, and then provides the corresponding translation.
+
+Quality Analysis. To figure out the quality of GPT-4o translations, we calculate the reference-free CometKiwi score between the source English sentences and GPT-4o translations. As a result, the average score of  $\mathrm{En} \Rightarrow \mathrm{Zh}$  translation is 84.48, and  $\mathrm{En} \Rightarrow \mathrm{De}$  translation is 84.90, indicating high translation quality (Rei et al., 2022).
+
+Other Details. The version of GPT-4o used in this work is GPT-4o-2024-08-06. When calling the official APIs, we set the temperature to 0.1, and set default values for other hyper-parameters.
+
+# B Implementation Details
+
+Training Details. Llama-Factory (Zheng et al., 2024) is used to instruct-tune LLMs. All LLMs are tuned on  $8 \times$  NVIDIA A100 GPUs (40G) with 1e-5 learning rate and  $32(8 \times 4)$  batch size. We use the DeepSpeed optimization (Rasley et al., 2020), and set ZeRO-2 optimization for Qwen2.5-7B-Instruct and Mistral-7B, while ZeRo-3 for Qwen2.5-14B
+
+Instruct and Llama-3-8B. During tuning, documents are also truncated to ensure the input length is within 2K tokens.
+
+SFT prompt. The system prompt in SFT is the same as the GPT-4o translator (c.f. Appendix A). The user prompt in SFT is provided as follows: "<document>[doc]</document><input sentence>[sent]</input sentence>", where <document>, </document>, <input sentence> and </input sentence> are special tokens to indicate the boundaries of the given document (denoted as "[doc]") and the source sentence ("[sent]").
+
+Model Checkpoints. We use four LLM backbones in experiments, i.e., Qwen2.5-7B-Instruct<sup>5</sup>, Qwen2.5-14B-Instruct<sup>6</sup>, Llama-3-8B<sup>7</sup> and Mistral-7B<sup>8</sup>. All model checkpoints are available at Huggingface.co community.
+
+Training Hours. All experiments are conducted on NVIDIA A100 GPUs with 40G memory, and we use its GPU hours to denote the consumption of computing resources. We SFT LLMs on the
+
+```txt
+$^{5}$ https://huggingface.co/Qwen/Qwen2. 5-7B-Instruct  ${}^{6}$ https://huggingface.co/Qwen/Qwen2. 5-14B-Instruct  ${}^{7}$ https://huggingface.co/hfl/ llama-3-chinese-8b-instruct-v3  ${}^{8}$ https://huggingface.co/mistralai/ Mistral-7B-Instruct-v0.3
+```
+
+# PROMPT
+
+Please continue the English sentence based on the information provided in the given document.
+
+<document>
+
+一直以来，K联赛都没有引入升降级制度。2007年韩国足球协会试图将K联赛最后一名跟韩国全国联赛第一名之间实行升降级，但韩国全国联赛第一名以参加K联赛经费太高为由拒绝升级，韩国足球协会让K联赛跟韩国全国联赛之间实行升降级尝试失败。The K League then took steps to create the new second division, mainly with the addition of a split-system during the 2012 K-Legue season in which the bottom clubs are placed in a competition for safety with the last placed club being relegated to the new second division (originally it was going to be two clubs relegated but the withdrawal of Sangsu Sangmu FC meant only one would be relegated). Au cours de laaison 2013, les et équipés de la K League Classic ont été automatiquement reliquées, tandis que la equipe a joué un match contre le vainueur du K League Challenge nouvellement formé pour decide de la promotion / relégation. 2012 waren erstmals die neuen Mitglieder der 2. The promotion playoffs are as follows: the fourth-placed team plays against the third-placed team, then the winner of this match plays with the second-placed team. Si le match est à également, é-teame la première classée avance. Le 3 janvier 2013, le nom officiel de la deuxième division a été annexe comme la K League tandis que la K League originale a été changée en K League Classic avec le nouveau logo. This name change caused some degree of confusion and controversy, and on 11 March 2013 the official name was changed to K League Challenge. On 22 January 2018 the official name was changed to K League 2.
+
+</document>
+
+<English sentence>
+
+The K League 2 is a South Korean professional association football league. </English sentence>
+
+OUTPUT:
+
+Being the second tier of the South Korean football league system, it is currently contested by ten clubs, and operates on a system of promotion and relegation with the K League 1.
+
+# PROMPT
+
+You are a professional translator, and your task is to translate an given input sentence from English to Chinese. In addition to the input sentence, you will be provided with a document that may contain relevant information to aid in the translation. However, be aware that some documents may contain irrelevant or noisy information.
+
+<document>
+
+1. **牡蛎的生态作用**：
+
+- 牡蛎是一种重要的海洋生物，不仅为人类提供食物资源，还在生态系统中扮演着关键角色。它们能够通过过滤水中的浮游植物和有机物来净化水质，有助于维持海洋环境的健康。牡蛎还能够形成天然的礁石结构，这些礁石为许多其他海洋生物提供了栖息地，增加了生物多样性。
+
+2.**牡蛎的生理特性**：
+
+- 牡蛎的身体具有一定的适应性，可以在特定条件下改变体液的浓度，使身体变得稀薄，从而更容易与其他牡蛎粘合在一起，形成坚固的礁石结构。
+
+- 这种粘合能力是牡蛎生存策略的一部分，有助于它们在潮间带等环境中抵御水流和波浪的冲击。
+
+3.**牡蛎礁石的重要性**：
+
+- 牡蛎礁石不仅是牡蛎本身的栖息地，还是为许多其他海洋生物提供了庇护所，包括鱼类、蟹类和其他无脊椎动物。
+
+- 这些礁石还能减少海岸侵蚀，保护沿海地区免受风暴和海浪的侵袭，对
+
+</document>
+
+<input sentence>
+
+Finally, oysters can attenuate and agglomerate onto each other and form these amazing natural reef structures.
+
+</input sentence>
+
+OUTPUT:
+
+最后，牡蛎可以减弱并聚集在一起，形成这些令人惊叹的自然礁石结构。
+
+# PROMPT
+
+You are a professional translator responsible for assessing the relevance of a provided document in relation to a source input sentence. Please response "relevant" or "irrelevant".
+
+<document>
+
+软件的RAM disk使用一般的RAM的一部分来模拟一个硬盘。通常都可以使用由操作系统内核支持的机制来实现，但是通过用户空间（相对于“内核空间”的应用程序进程来进行这种的模拟也是有可能的。因为存贮在RAM disk中文件一般都是临时性的，所以也就不需要使用电池备份，或者UPS等设备。某些RAM disk使用压缩的文件系统例如cramfs。这是因为一般的RAM disk都容量一般都较小，且RAM的价格比硬盘要高，所以这样做是很合理的。
+
+</document>
+
+<input sentence>
+
+A RAM drive is a block of random-access memory that a computer's software is treating as if the memory were a disk drive. It is sometimes referred to as a virtual RAM drive or software RAM drive to distinguish it from a hardware RAM drive that uses separate hardware containing RAM, which is a type of battery-backed solid-state drive.<input sentence>
+
+OUTPUT:
+
+relevant
+
+(a) Cross-lingual Information Completion
+
+(b) Self-knowledge-enhanced Translation
+
+(c) Cross-lingual Relevance Discrimination
+
+Figure 4: Examples of CSC training objectives. Different colors in (a) means different languages, including Chinese, English, French and German.
+
+training data of RAGtrans with 2 epochs, and each epoch costs 9.1 GPU hours, 54.0 GPU hours, 33.5 GPU hours, and 9.3 GPU hours for Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, Llama-3-8B and Mistral-7B, respectively. For SFT+CSC LLMs, more GPU hours are costed. For example, to SFT Qwen2.5-14B-Instruct on both the RAGtrans training samples and CSC samples, each epoch costs 208 GPU hours; while the counterparts of Qwen2.5-7B-Instruct, Llama-3-8B and Mistral-7B are 34.1, 128.0 and 33.9 GPU hours, respectively.
+
+Multi-Task Training Samples. As we introduce in Section 3, there are three training objectives in CSC multi-task training method. To provide a deeper understanding of these objectives, here we give some example samples in Figure 4. In our main experiments, we create 40K samples for each training objective.
+
+Metric Implementation. To calculate the reference-based COMET score (Rei et al., 2022), we leverage the official codes and the official model. To calculate the BLEU score, we use the sacrebleu toolkit to calculate the corpus-level BLEU. In  $\mathrm{En} \Rightarrow \mathrm{Zh}$  and  $\mathrm{En} \Rightarrow \mathrm{De}$  translation, the "tokenize" is set to "zh" and "char", respectively. For GRB and GRF, we prompt GPT-4o (2024-08-06
+
+Score the following translation from English to [Chinese/German] with respect to the human reference on a continuous scale from 0 to 100, where score of zero means "no meaning preserved" and score of one hundred means "perfect meaning and grammar".
+
+English source: [src]
+
+[Chinese/German] human reference: [ref]
+
+[Chinese/German] translation: [hyp]
+
+Score:
+
+Score the following translation from English to [Chinese/German] on a continuous scale from 0 to 100, where score of zero means "no meaning preserved" and score of one hundred means "perfect meaning and grammar".
+
+English source: [src]
+
+[Chinese/German] translation: [hyp]
+
+Score:
+
+Figure 5: The prompts in GRB (upper part) and GRF (lower part). “[src]”, “[ref]” and “[hyp]” denote the source sentence, human translation and model translation, respectively. “[Chinese/German]” denotes “Chinese” or “German”, depending on the target language.
+
+version) as the MT evaluator in the reference-based and reference-free manners, respectively. The corresponding prompts borrow from Kocmi and Federmann (2023), and are illustrated in Figure 5. Since GRB and GRF need the API costs, we randomly select 200 samples from the RAGtrans testing set, and conduct the GRB and GRF evaluation. All experimental results listed in this paper are the average of 3 runs.
+
+![](images/86723242ef93f9e0906751aaf6e90540fdc8d27f80f575ea045573d4c7a57829.jpg)  
+Figure 6: The experimental results of CSC scalability (in  $\mathrm{En}\Rightarrow \mathrm{Zh}$
+
+![](images/1df485217a357efd96dc4f4a0c9c7b9736e2bb5e546598b93a54e3c7d0b5a935.jpg)
+
+# C Scalability of CSC
+
+As we demonstrate the effectiveness of CSC multi-task training method in experiments, we wonder the upper limit of the improvement brought by CSC. To this end, we use Qwen2.5-7B-Instruct as the backbone, and systematically vary the number of CSC samples during the instruction tuning to examine the resulting performance (w/ golden Chinese document) changes. As shown in Figure 6, when the number of CSC samples exceeds  $120\mathrm{K}$ , the improvement brought by CSC begins to plateau. When the number of CSC samples increases from  $210\mathrm{K}$  to  $240\mathrm{K}$ , the model performance does not improve accordingly.
+
+# D Details of Full Wiki Testing
+
+Retriever. For BM25 retriever, we use the implementation of elasticsearch $^{12}$  toolkit to retrieve top-3 documents for each source sentence. For BGE-m3 retriever, we first use BGE-m3 sentence embedding model $^{13}$  to calculate the embedding of all documents in knowledge sources, and then use the embedding of source sentence to retrieve top-3 relevant documents via FAISS (Johnson et al., 2019).
+
+Knowledge Sources. We use Wikipedia dumps (20241001 version) as the knowledge sources, and leverage wikiextractor $^{14}$  toolkit to extract articles from Wikipedia dumps. Following Karpukhin et al. (2020), we split each article into multiple, disjoint
+
+<table><tr><td></td><td>BLEU w/ Empty Doc.</td><td>COMET w/ Golden Doc.</td><td>BLEU w/ Golden En.</td><td>COMET w/ Golden En.</td></tr><tr><td>Qwen2.5-7B (SFT+CSC)</td><td>60.9</td><td>71.8</td><td>63.5</td><td>73.4</td></tr><tr><td>- CLIC</td><td>60.4</td><td>71.5</td><td>62.8</td><td>72.9</td></tr><tr><td>- SKET</td><td>59.8</td><td>71.0</td><td>62.2</td><td>72.3</td></tr><tr><td>- CLRD</td><td>60.5</td><td>71.5</td><td>63.0</td><td>73.1</td></tr><tr><td></td><td colspan="2">w/ Golden Zh.</td><td colspan="2">w/ Golden De.</td></tr><tr><td>Qwen2.5-7B (SFT+CSC)</td><td>63.2</td><td>73.1</td><td>64.5</td><td>74.5</td></tr><tr><td>- CLIC</td><td>62.4</td><td>72.6</td><td>63.4</td><td>73.7</td></tr><tr><td>- SKET</td><td>62.9</td><td>72.9</td><td>62.9</td><td>73.2</td></tr><tr><td>- CLRD</td><td>62.2</td><td>72.5</td><td>63.9</td><td>73.9</td></tr></table>
+
+Table 8: Ablation study on  $\mathrm{En} \Rightarrow \mathrm{De}$  golden evaluation. Doc.: Document; "En.", "Zh." and "De." indicate English, Chinese and German documents, respectively.
+
+text blocks of 100 words as passages, serving as our basic retrieval units. In the full Wiki evaluation, we build the knowledge sources based on English, Chinese, German, French, Czech, Russian, Korean and Japanese Wikipedia dumps, resulting in tens of millions of retrieval units.
+
+# E Ablations on  $\mathbf{En}\Rightarrow \mathbf{De}$
+
+Table 8 shows the ablation results on  $\mathrm{En} \Rightarrow \mathrm{De}$  golden evaluation. The results also demonstrate that SKET is more important than the other two tasks when given empty documents or documents in the source/target language. When given documents in a third language, the importance of SKET decreases and is less than that of CLIC/CLRD. The conclusion is consistent with that in Section 4.4.
+
+# F Details of Human Evaluation
+
+Evaluators. Three master students are recruited in our human evaluation, and they are fluent in both Chinese and English.
+
+Instruction. The human evaluators are provided with the instructions for each translation error type:
+
+<table><tr><td>SFT LLMs</td><td>BLEU</td><td>COMET</td><td>GRB</td><td>GRF</td></tr><tr><td>Qwen2.5-7B (w/o LP.)</td><td>57.3</td><td>87.1</td><td>91.2</td><td>91.5</td></tr><tr><td>Qwen2.5-14B (w/o LP.)</td><td>58.1</td><td>87.2</td><td>91.7</td><td>92.0</td></tr><tr><td>Llama-3-8B (w/o LP.)</td><td>57.1</td><td>87.0</td><td>90.2</td><td>90.8</td></tr><tr><td>Mistral-7B (w/o LP.)</td><td>56.4</td><td>87.8</td><td>89.9</td><td>90.3</td></tr><tr><td>Qwen2.5-7B (w/ LP.)</td><td>57.9</td><td>87.7</td><td>92.1</td><td>92.3</td></tr><tr><td>Qwen2.5-14B (w/ LP.)</td><td>58.6</td><td>87.9</td><td>92.4</td><td>92.5</td></tr><tr><td>Llama-3-8B (w/ LP.)</td><td>57.7</td><td>87.5</td><td>90.9</td><td>91.4</td></tr><tr><td>Mistral-7B (w/ LP.)</td><td>57.1</td><td>88.0</td><td>90.9</td><td>91.1</td></tr></table>
+
+Table 9: Experimental results (En  $\Rightarrow$  Zh) with and without lead paragraphs of golden Chinese documents. w/: with; w/o: without; LP.: lead paragraphs.
+
+(1) Reference Errors: Are there any mistakes in pronoun or reference usage that could cause confusion about what or whom is being referred to? (2) Word-Level Errors: Are there incorrect translations, omissions, or additions of individual words that alter the meaning of the text? (3) Phrase-Level Errors: Are there incorrect translations, omissions, or additions of phrases that affect the overall coherence and accuracy of the translation? (4) Fluency Issues: Does the translation flow smoothly, or are there awkward phrases or constructions that impede comprehension? (5) Other Errors: Are there any additional errors present in the translation that do not fit into the categories above?
+
+Evaluation Samples. Since human evaluation is labor-intensive, we randomly select 200 samples from the  $\mathrm{En} \Rightarrow \mathrm{Zh}$  testing set of RAGtrans to conduct the human evaluation.
+
+Inter-agreement. The Fleiss' Kappa scores (Fleiss, 1971) of the five error types are 0.63, 0.57, 0.68, 0.75 and 0.66 in our human evaluation, respectively, indicating a good inter-agreement among our evaluators.
+
+Other Details. During our human evaluation, we find that a golden Chinese document might involve multiple domain terms that are similar in lexical but with slight discrepancies. During translation, LLMs might use a flawed Chinese term (similar to the right term from the golden Chinese document) in the translation results. This will make them achieve good results in terms of automatic evaluation (including GRF and GRB). However, in our human evaluation, we will judge them as translation errors.
+
+# G Experiments with Lead Paragraphs
+
+As illustrated in Section 2, When constructing documents, we remove the lead paragraphs (i.e.,
+
+$D^{l} \setminus p_{1}^{l}(l \in \{\text{en, zh, de, fr, cs}\})$ . This is under the considerations of (a) answer leakage and (b) consistency with real scenes. For (a), the answer leakage issue will exist if we do not remove the lead paragraphs for the target-language documents. This is because the lead paragraphs in the target-language documents might involve the translation fragments of source sentences. For (b), in real-world applications, for a source sentence, if we can retrieve a document that involves its translation or translation fragments, it will be useful for translating the sentence. But in most cases, we cannot guarantee that the retrieved documents involve translations or translation fragments due to many practical difficulties such as insufficient knowledge corpus, limited ability in cross-lingual retriever, etc. Therefore, we also remove the lead paragraphs from documents in other languages (beyond the target language).
+
+To figure out the effect of lead paragraphs in non-English documents, we conduct a new experiment to evaluate model performance equipped with them. Specifically, we add the lead paragraphs of the Chinese documents for each  $\mathrm{En} \Rightarrow \mathrm{Zh}$  test sample, and re-evaluate their performance (w/ Golden Chinese document) in the test set. As shown in Table 9, with the help of lead paragraphs, the model performance is generally improved.

retrievalaugmentedmachinetranslationwithunstructuredknowledge/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5865bef487f1c071645c7a73c3732766cd150f71921765ecd7de47400e6fa3d
+size 664545

retrievalaugmentedmachinetranslationwithunstructuredknowledge/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29aad1a781f274029722f57be5f304f562dee6c1c540ffac58e66798fd305b76
+size 537116

revealandreleaseiterativellmunlearningwithselfgenerateddata/1f34f70f-7440-47df-90ce-108baac1c739_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:327b39c2f94ef49d6ab6b001318a6dc41576ea4545375b7fbfd44024f18e44e5
+size 89891

revealandreleaseiterativellmunlearningwithselfgenerateddata/1f34f70f-7440-47df-90ce-108baac1c739_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f946a61705068f4c38ea566d8ab4a507dc17f5c0469185336c79ffbd6ff27fb6
+size 111837

revealandreleaseiterativellmunlearningwithselfgenerateddata/1f34f70f-7440-47df-90ce-108baac1c739_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:486a63683da0ac8b9598919bd62339987a992533bf7264ee40c2ecc25628e133
+size 2505768

revealandreleaseiterativellmunlearningwithselfgenerateddata/full.md

@@ -0,0 +1,387 @@
+# Reveal and Release: Iterative LLM Unlearning with Self-generated Data
+
+# Linxi Xie, Xin Teng, Shichang Ke, Hongyi Wen, Shengjie Wang
+
+New York University Shanghai, Center for Data Science
+
+{lx2154, xt2251, sk11726, hongyi.wen, shengjie.wang}@nyu.edu
+
+# Abstract
+
+Large language model (LLM) unlearning has demonstrated effectiveness in removing the influence of undesirable data (also known as forget data). Existing approaches typically assume full access to the forget dataset, overlooking two key challenges: (1) Forget data is often privacy-sensitive, rare, or legally regulated, making it expensive or impractical to obtain (2) The distribution of available forget data may not align with how that information is represented within the model. To address these limitations, we propose a "Reveal-and-Release" method to unlearn with self-generated data, where we prompt the model to reveal what it knows using optimized instructions. To fully utilize the self-generated forget data, we propose an iterative unlearning framework, where we make incremental adjustments to the model's weight space with parameter-efficient modules trained on the forget data. Experimental results demonstrate that our method balances the tradeoff between forget quality and utility preservation.
+
+# 1 Introduction
+
+Large language models (LLMs) function as vast knowledge repositories, drawing on information embedded in their parameters in response to user inputs (Brown et al., 2020). However, the scope of their knowledge is fixed at the time of training, lacking effective means to verify and may produce responses that are outdated, incorrect, or even harmful (Liang et al., 2023). Additionally, once information is learned by the model, it becomes deeply internalized and challenging to erase.
+
+Machine unlearning has become a promising area of research aimed at addressing these limitations. A straightforward approach—known as exact unlearning—involves removing undesirable data from the training corpus and retraining
+
+![](images/4266fdc8ef189fd32fb694148913f1b9ee3b7a8e02c3a0fa34f4ac6dd80107dc.jpg)  
+Figure 1: External forget data may include information irrelevant to the true unlearning target, or miss the model's knowledge related to the target. Our approach enables effective unlearning with minimal utility loss.
+
+the model from scratch, which is prohibitively resource-intensive for modern LLMs. Researchers have explored approximate unlearning, which seeks to remove relevant knowledge without full retraining. The goal is to efficiently and selectively erase the influence of targeted information while maintaining the model's performance on nontargeted tasks (Liu et al., 2024a). Current methods include gradient ascent that effectively guide models to forget by optimizing in the opposite direction of original learning (Ullah et al., 2021); knowledge editing methods that locate and directly modify network parameters to perform targeted information removal (Meng et al., 2023); and influence function approaches that identify and neutralize the impact of specific training examples (Li et al., 2024b).
+
+In a typical machine-unlearning process, one crucial factor is the data, specifically, the information to be forgotten and the information to be retained (Xu, 2024), which we refer to as forget data and retain data. Most unlearning methods
+
+require well-annotated forget data. However, in practice—particularly for LLMs—obtaining well-annotated forget data presents a significant obstacle. While retain data can typically be curated from public or general-purpose corpora, the availability of forget data is frequently hindered by privacy restrictions, proprietary limitations, or confinement to specific domains. Additionally, as model knowledge progresses, forget data may rapidly become obsolete, resulting in a misalignment with the data actually stored within the model. Moreover, existing unlearning benchmarks often assume access to the model's original training data or an exact forget subset (Maini et al., 2024), which is unrealistic for massive and private corpora. In other cases, forget data consists of publicly sourced approximations (Gehman et al., 2020), herein termed as external data; however, such data may not faithfully represent how the information is genuinely encoded within the model. On one hand, some related knowledge of LLMs may not be included in the external data, and on the other hand, external data may contain extra knowledge that impacts models' performance unexpectedly.
+
+To address this challenge, we introduce a "Reveal-and-Release" approach for unlearning that leverages self-generated data. Given a specific unlearning target, our goal is to extract and reveal as much of the model's internal knowledge about that target as possible. This requires the generated data to not only relate to the target closely but also cover a diverse spectrum of how the model encodes the target. Instead of relying on well-labeled external forget data, we use a NeuralUCB-based instruction optimization method (Zhou et al., 2020; Lin et al., 2024) to generate prompts to reveal internal knowledge, focusing on the relevance and diversity of the generation (Section 3.1). We refer to the resulting self-generated data as internal data.
+
+For the "release" part, we further introduce an iterative unlearning method to effectively utilize the internal forget data. Inspired by Parameter-Efficient Module (PEM) composition (Zhang et al., 2023), our approach incrementally edits the base model by merging two types of PEM LoRAs (Hu et al., 2022): a forget PEM trained on internal forget data and a retain PEM trained on retain data. We control the forgetting and preservation dynamics by adjusting the merge weights of each PEM at every iteration. Intuitively, the LoRAs act like gradient ascents/descents, and multiple iterations of unlearning correspond to applying small steps of
+
+gradient optimizations. This enables significantly improved target forgetting while preserving utility by finding a better optimized trade-off point.
+
+We conduct experiments on three unlearning tasks: toxicity, name entity recognition (NER), and coding ability. Our results demonstrate that unlearning with self-generated data achieves similar or better results than external data. Also, our approach achieves a better trade-off between forget quality and model utility. Our contributions are:
+
+1. We study LLM unlearning with self-generated forget data, generated through optimized instruction search and multi-turn prompting, eliminating the need for well-annotated, externally sourced forget datasets.  
+2. We propose an Iterative Unlearning method that incrementally edits the base model by alternating between retain and forget Parameter-Efficient Modules (PEMs), enabling control over the trade-off between forget quality and utility preservation.  
+3. Experiments and ablation studies across multiple tasks demonstrate that our framework effectively supports targeted forgetting with minimal degradation to retained capabilities.
+
+# 2 Related Work
+
+Data Synthesis for Unlearning Well-annotated data is expensive to obtain. In non-LLM domains, Shen et al. (Shen et al., 2024) introduce Label-Agnostic Forgetting (LAF), a supervision-free unlearning framework that manipulates representation distributions to remove forgotten data without relying on labels. Peng et al. (Peng et al., 2025) propose MixUnlearn, which uses adversarially generated mixup samples to mitigate catastrophic unlearning, ensuring effective data deletion even in label-agnostic scenarios.
+
+In the domain of LLMs, prior work has explored using synthesized data for unlearning. CMD introduces a detoxification framework for LLMs that leverages synthesized data to enable unlearning of toxic behaviors (Tang et al., 2024). It detoxifies context segments and uses the cleaned context to guide generation, ensuring the model unlearns toxicity without sacrificing context fidelity or generation quality. RWKU (Jin et al., 2024) constructs a synthetic forget corpus by prompting LLM with manually crafted templates in a single-pass manner. While this provides a straightforward way to obtain forget data, the reliance on fixed
+
+prompt templates and single-pass generation risks capturing only a narrow view of the model's internal knowledge, potentially missing out on diverse or harder-to-reach information.
+
+Parameter-Efficient-Module for Unlearning Parameter-efficient fine-tuning (PEFT) methods such as LoRA (Hu et al., 2022) have become popular for adapting LLMs due to their efficiency and modularity. Recent research explores how these parameter-efficient modules (PEMs) can be composed through arithmetic operations to enable unlearning(Zhang et al., 2023). Building on this, Liu et al. (Liu et al., 2024b) proposed SKU, which trains multiple modules from different perspectives and merges them before a single subtraction, aiming to better capture harmful knowledge from multiple angles. Ding et al. (Ding et al., 2025) proposed a unified framework for PEM-based unlearning by applying influence functions to directly update existing PEMs.
+
+Extending this line of work, Hu et al. (Hu et al., 2024) introduced Ext-Sub, a method to isolate and subtract only the "deficiency capability" from an anti-expert PEM. Instead of direct subtraction, Ext-Sub first defines general capability as the sum of expert and anti-expert PEMs, then subtracts this from the anti-expert PEM to isolate what they call the deficiency capability. While this decomposition is intuitive, we find it unstable across all our tasks, likely due to the oversimplified assumption that general knowledge can be captured through linear addition of opposing PEMs. Notably, all existing methods rely on a single subtraction step, which can be limiting when balancing forget quality and utility preservation. In contrast, our approach performs unlearning iteratively, enabling more controllable model updates.
+
+# 3 Method
+
+Our method consists of two stages: we first obtain self-generated forget data by optimizing instructions for the LLM, and then utilize the obtained data in an iterative unlearning framework.
+
+# 3.1 Forget Data Generation
+
+To generate high-quality internal forget data, we aim to elicit as much relevant and diverse knowledge as possible from the model with a set of optimized instructions. We formulate this as an instruction optimization problem and use a query-efficient search framework based on a NeuralUCB
+
+algorithm following prior work (Garnett, 2023; Lin et al., 2024). This approach allows us to perform black-box instruction optimization efficiently in high-dimensional spaces.
+
+The instruction search is guided by a task-specific scoring function designed to reflect two core objectives:
+
+- Relevance: The generated internal data should strongly reflect the unlearning target (e.g., high toxicity if we aim to forget toxic behavior).  
+- Diversity: The generated internal data should span a wide range of content and thoroughly reflect the model's internal knowledge of the unlearning target.
+
+We assume a metric or oracle is available to quantify the relevance of the generated data to the task (for example, a model to calculate the toxicity score for toxicity unlearning). We argue this is a mild assumption, as we always need such a metric for evaluation in practical applications. Even in cases of unlearning with external data, such a metric is still required for assessment. The specific relevance metric used for each task is detailed in Section 4.
+
+To capture diversity, we use the Vendi score (Friedman and Dieng, 2023), which is defined as the exponential of the Shannon entropy of the eigenvalues of a similarity matrix. Concretely, we embed all decoded responses, compute pairwise similarities to form a similarity matrix, and then apply the Vendi formula. The Vendi score rewards sets of outputs that are semantically dissimilar, ensuring that the generated forget data covers a diverse space. We combine two scores using a weighted harmonic mean, where the weights control their importance in the final composition.
+
+NeuralUCB Instruction Optimization To generate internal data that matches the two objectives, we apply a NeuralUCB-based approach: we initialize a set of soft prompts (the bandits) and search for the top soft prompts that generate outputs with high scores (relevance and diversity). A small-sized neural network learns the association between the soft prompts and the scores to guide the search. The details are shown in Alg. 1.
+
+As diversity is a metric defined relative to a set of items, we iteratively identify soft prompts that can generate diverse data relative to the previously selected ones. Our algorithm consists of an outer loop and an inner loop. At the beginning of each outer-loop iteration, we initialize the neural net
+
+![](images/cfa83abe321b993528ad158bb96ddb536c7f284448a5c9812657e06ce7e74229.jpg)  
+Stage1: self-generate forget data
+
+![](images/711dbde60d38840d227a4924f98ae0466d9c722da763dca92d264bb85a8404c3.jpg)  
+Stage2: Iterative Unlearning
+
+![](images/44c5c5bb74485ada5226a63a083f77bd4062ef9e21b4809782652992a0520a83.jpg)  
+Final Result:  
+Figure 2: Overview of our two-stage unlearning framework. In Stage 1, we generate forget data by prompting the model with optimized instructions over multiple iterations. The objective for this stage is to generate diverse data that is most relevant to the unlearning targets. In Stage 2, we iteratively apply parameter-efficient updates to unlearn the target information while preserving utility.
+
+work for NeuralUCB with  $k$  high-scoring prompts from previous outer iterations (we use  $k = 10$ ). This provides a strong starting point for prompt searching. Assuming  $D_{\mathrm{self - gen}}$  contains the internal data collected so far, we then launch the inner loop to identify the best instruction that prompts the model to generate outputs that are both relevant to the unlearning target and diverse relative to the existing samples in  $D_{\mathrm{self - gen}}$  guided by NeuralUCB. Once identified, this instruction is used to generate new responses conditioned on the given prompts (generation context  $C$ ), and the resulting outputs are added to  $D_{\mathrm{self - gen}}$ .
+
+# 3.2 Iterative Unlearning with PEM
+
+Iterative PEM Composition for Unlearning Inspired by prior work (Zhang et al., 2023), we propose an iterative unlearning framework that incrementally edits the base model by composing parameter-efficient modules (PEMs) trained on different objectives. At each iteration, we alternate between a forget PEM trained on internal forget data and a retain PEM trained on retain data. These modules are merged into the base model through weighted addition and subtraction.
+
+We initiate unlearning by subtracting a forget PEM from the base model. In each subsequent
+
+iteration, we perform two steps:
+
+1. Train a retain PEM on retain data using the negated model as the base; merge it via addition.  
+2. Train a forget PEM on the forget data using the updated model; merge it via subtraction.
+
+This process is repeated for several iterations. Although prior work has suggested potential overlap between PEMs trained on retain and forget data (Hu et al., 2024), our analysis (See Section 4.1) shows that the two modules are largely orthogonal, and forcing orthogonality between these opposing PEMs does not improve unlearning performance (See Appendix B). As a result, we adopt a simple linear merge strategy:
+
+$$
+\begin{array}{l} \Phi^ {(t)} = \Phi_ {0} - \mu_ {0} \Delta \Phi_ {\mathrm {f o r g e t}} ^ {(0)} \\ + \sum_ {i = 1} ^ {t} \left(\lambda_ {i} \Delta \Phi_ {\text {r e t a i n}} ^ {(i)} - \mu_ {i} \Delta \Phi_ {\text {f o r g e t}} ^ {(i)}\right) \tag {1} \\ \end{array}
+$$
+
+where  $\Phi_0$  is the frozen base model, and  $\Delta \Phi_{\mathrm{forget}}^{(0)}$  is the initial forget PEM. At each iteration  $i\geq 1$  we alternately train a retain PEM and a forget PEM, denoted by  $\Delta \Phi_{\mathrm{retain}}^{(i)}$  and  $\Delta \Phi_{\mathrm{forget}}^{(i)}$  respectively. Scalars  $\lambda_{i}$  and  $\mu_{i}$  control the influence of each module. This formulation allows us to initialize forgetting with a strong signal, then
+
+Algorithm 1 Generate Forget Data with Instruction Optimization
+
+1: Input: Generation context  $C$ ; Number of outer iterations  $m$ ; Number of inner iterations per outer loop  $n$ ; soft prompt set  $P$ ; response generator  $f(C, P_i)$  with generation context  $C$  and instruction  $P_i$ ; weight  $\alpha$  for harmonic mean;  
+2: Initialize self-generated dataset  $D_{\mathrm{self - gen}}\gets \emptyset$  
+3: for  $i = 1$  to  $m$  do  
+4: Initialize network for NeuralUCB with  $k$  high-score soft prompts  
+5: for  $t = 1$  to  $n$  do  
+6: Select prompt:  
+7:  $P_{t}\gets \arg \max_{P}$  NeuralUCBt(P)  
+8: Generate response  $y_{t} \gets f(C, P_{t})$  
+9: Compute relevance  $\tau_{t}$  
+10: Compute diversity:  
+11:  $v_{t}\gets \mathrm{Vendi}(y_{t}\cup D_{\mathrm{self - gen}})$  
+12: Compute score:
+
+$$
+\operatorname {S c o r e} \left(y _ {t}\right) \leftarrow \left(\frac {\alpha}{v _ {t}} + \frac {1 - \alpha}{\tau_ {t}}\right) ^ {- 1}
+$$
+
+13: Update NeuralUCB with Score  $(y_{t})$  
+14: end for  
+15: Select best prompt:  
+16:  $P^{*}\gets \arg \max_{P}\mathsf{Score}(f(C,P))$  
+17: Update self-gen data:  
+18:  $D_{\mathrm{self - gen}}\gets D_{\mathrm{self - gen}}\cup \{f(C,P^{*})\}$  
+19: end for  
+20: Return: Final forget dataset  $D_{\mathrm{self - gen}}$
+
+refine the model iteratively by reinforcing retaining behavior and further subtracting residual traces of the target knowledge.
+
+Merge Weight Selection. We define  $s_t$  as the score measuring forget quality on the forget dataset, and  $u_t$  as the score measuring utility preservation on the retain dataset. The subtraction weight  $\mu_i$  is chosen to ensure that the model either (1) forgets at least 90% of the target behavior compared to the beginning of the current iteration, or (2) does not sacrifice more utility than it gains in forgetting. Formally, we select  $\mu_i$  such that either  $s_i \leq 0.1 \cdot s_{i-1}$  or the reduction in forget score exceeds the reduction in utility, i.e.,  $(s_{i-1} - s_i) > (u_{i-1} - u_i)$ .
+
+For the addition weight  $\lambda_{i}$ , our goal is to restore as much utility as possible after forgetting. We select  $\lambda_{i}$  such that the model recovers at least  $95\%$  of the utility score compared to the beginning of the
+
+current iteration, i.e.,  $u_{i} \geq 0.95 \cdot u_{i-1}$ . These rules ensure that the unlearning process is both effective and balanced (See Section 5.2).
+
+# 4 Experiments
+
+To evaluate the effectiveness of our self-generated forget dataset, we conduct experiments on three tasks: LLM detoxification, Named Entity Recognition (NER) unlearning, and coding ability unlearning. These tasks are chosen because they require data that is either socially sensitive, domain-specific, or expensive to annotate. All experiments are performed using the LLaMA3-8B-Instruct model (Grattafori et al., 2024), and we use all-roberta-large-v1 (Reimers and Gurevych, 2019) to embed texts for diversity scores. To further assess the generalizability of our framework, we also include results on Mistral-7B-Instruct-v0.2 (Mistral AI, 2024) (See Appendix C).
+
+<table><tr><td>Task</td><td>Avg. Similarity</td><td>Std. Dev.</td></tr><tr><td>Toxicity</td><td>0.0484</td><td>0.0230</td></tr><tr><td>Coding</td><td>0.0397</td><td>0.0234</td></tr><tr><td>NER</td><td>0.0398</td><td>0.0208</td></tr></table>
+
+Table 1: Average eigenbasis similarity  $\left( {\operatorname{top} - k = 8}\right)$  between retain and forget PEMs across layers.
+
+# 4.1 Preliminary Study
+
+We first conduct a preliminary analysis to quantify the overlap between the retain and forget PEMs. For each layer, we obtain the merged LoRA update matrix  $W = BA$ , and compute its top-  $k$  left singular vectors via SVD:
+
+$$
+W _ {\text {r e t a i n}} = U _ {1} \Sigma_ {1} V _ {1} ^ {\top}, \quad W _ {\text {f o r g e t}} = U _ {2} \Sigma_ {2} V _ {2} ^ {\top},
+$$
+
+where  $U_{1}^{(k)}$  and  $U_{2}^{(k)}\in \mathbb{R}^{d\times k}$  denote the top-  $k$  left singular vectors.
+
+To measure the similarity between the subspaces, we compute:
+
+$$
+\operatorname {S i m} \left(U _ {1} ^ {(k)}, U _ {2} ^ {(k)}\right) = \frac {1}{k} \left\| U _ {1} ^ {(k) ^ {\top}} U _ {2} ^ {(k)} \right\| _ {F},
+$$
+
+where  $\| \cdot \| _F$  denotes the Frobenius norm. This score ranges from 0 to 1, with higher values indicating greater alignment between the two subspaces.
+
+We report the average and standard deviation of the similarity scores across all layers for each task in Table 1. Across all tasks, the average similarity remains low (below 0.05), indicating that the retain
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">PPL ↓</td><td colspan="3">Challenge</td><td colspan="3">Non-Challenge</td></tr><tr><td>Tox. Score ↓</td><td>Tox. Rate ↓</td><td>Severe Tox. ↓</td><td>Tox. Score ↓</td><td>Tox. Rate ↓</td><td>Severe Tox. ↓</td></tr><tr><td>Basemodel</td><td>7.2055</td><td>0.7310</td><td>0.3654</td><td>0.2725</td><td>0.2986</td><td>0.0167</td><td>0.0352</td></tr><tr><td>DPO</td><td>8.9598</td><td>0.6871</td><td>0.3654</td><td>0.2648</td><td>0.2724</td><td>0.0234</td><td>0.0337</td></tr><tr><td>RMU</td><td>7.2056</td><td>0.7010</td><td>0.4038</td><td>0.2507</td><td>0.2912</td><td>0.0190</td><td>0.0334</td></tr><tr><td>CMD</td><td>8.6479</td><td>0.6574</td><td>0.3173</td><td>0.2280</td><td>0.2850</td><td>0.0167</td><td>0.0349</td></tr><tr><td>Ext-Sub</td><td>7.8563</td><td>0.4447</td><td>0.0769</td><td>0.0973</td><td>0.1740</td><td>0.0011</td><td>0.0100</td></tr><tr><td>PEM-external</td><td>10.4109</td><td>0.4479</td><td>0.0865</td><td>0.0877</td><td>0.1873</td><td>0.0022</td><td>0.0114</td></tr><tr><td>Ours</td><td>7.5513</td><td>0.3047</td><td>0.0481</td><td>0.0532</td><td>0.1842</td><td>0.0000</td><td>0.0123</td></tr></table>
+
+Table 2: Toxicity unlearning results on RTP. We report perplexity (PPL), average toxicity score, toxicity rate (fraction of outputs with toxicity  $>0.5$ ), and severe toxicity (score  $>0.8$ ), for both challenge and non-challenge subsets. Our method achieves strong toxicity reduction with lower perplexity.
+
+and forget PEMs occupy largely orthogonal subspaces. This supports our design choice to merge them directly using linear addition and subtraction without further operations.
+
+# 4.2 Baseline Models
+
+We compare our method against several baselines based on parameter-efficient methods (PEMs) and fine-tuning approaches. Specifically, we include Ext-Sub (Hu et al., 2024), CMD (Tang et al., 2024), and direction subtraction using a forget PEM trained on external data (Zhang et al., 2023) (denoted as PEM-external). We also evaluate the widely used DPO method (Rafailov et al., 2024) and RMU (Li et al., 2024a) in its best-performing configuration. We tune the weighting parameter  $\alpha$  for Ext-Sub and direction subtraction (PEM-external).
+
+# 4.3 Toxicity Unlearning
+
+Training To construct the forget dataset that captures the model's internal toxic behaviors, we use prompt-only inputs from RealToxicityPrompts (RTP) (Gehman et al., 2020) and CivilComments (Zhang et al., 2023), both widely adopted in prior detoxification studies (Hu et al., 2024; Ko et al., 2024; Tang et al., 2024). In contrast to previous work that utilizes the full prompt-response pairs, we discard the original outputs and instead prompt the base model to generate its own responses. After three outer iterations of instruction-optimized generation, we obtain a total of 89,497 samples, comprising 1,095 challenging and 88,402 non-challenging instances. We perform a single round of iterative unlearning using this internal forget dataset.
+
+Evaluation We evaluate the generation results from two aspects: forget quality and utility preservation. Utility preservation is quantified by perplexity (PPL) computed on the WikiText-2-raw-v1
+
+dataset. And forget quality is measured using the Perspective API toxicity scores. Following prior work (Tang et al., 2024; Ko et al., 2024), we use nucleus sampling to generate 25 continuations per prompt, each with a maximum of 20 tokens. Each continuation is scored with the Perspective API. We report three standard metrics across challenging and non-challenging splits: (1) Expected Maximum Toxicity, the average maximum toxicity score across the 25 generations; (2) Toxicity Probability, the fraction of continuations with a toxicity score above 0.5; and (3) Severe Toxicity, the fraction exceeding a score of 0.8.
+
+Results Our method outperforms all baselines on the challenging split, achieving the lowest toxicity score, toxicity rate, and severe toxicity. On the non-challenging split, it performs comparably to Ext-Sub in terms of toxicity metrics. Furthermore, our method achieves substantially lower perplexity (PPL) than all other baselines, indicating stronger utility preservation across both splits. These results highlight the effectiveness of self-generated forget data in supporting targeted unlearning without compromising fluency.
+
+# 4.4 NER Unlearning
+
+Training We build on prior work in LLM-based Named Entity Recognition (NER), which leverages LLMs to identify a wide range of entity types across diverse domains (Zhou et al., 2024). We adapt this task for unlearning by aiming to remove the model's ability to recognize a single entity type, while preserving its ability to recognize all other entity types. Specifically, we aim to unlearn the Person entity type and retain performance on the four most frequent entity types in the training set: Organization, Concept, Location, and Date. Since diversity score is not applicable in this setting, we directly prompt the base model to extract entities and their corresponding types for
+
+<table><tr><td>Model</td><td>Person F1 ↓</td><td>Org F1 ↑</td><td>Concept F1 ↑</td><td>Location F1 ↑</td><td>Date F1 ↑</td></tr><tr><td>Basemodel</td><td>0.5370</td><td>0.4501</td><td>0.2123</td><td>0.4747</td><td>0.7173</td></tr><tr><td>DPO</td><td>0.4140</td><td>0.5190</td><td>0.1840</td><td>0.4410</td><td>0.7847</td></tr><tr><td>RMU</td><td>0.3453</td><td>0.2869</td><td>0.1479</td><td>0.3310</td><td>0.5030</td></tr><tr><td>Ext-Sub</td><td>0.2444</td><td>0.2876</td><td>0.0667</td><td>0.3042</td><td>0.2640</td></tr><tr><td>PEM-external</td><td>0.2483</td><td>0.1641</td><td>0.0444</td><td>0.2187</td><td>0.4854</td></tr><tr><td>Ours</td><td>0.1430</td><td>0.5242</td><td>0.2299</td><td>0.5157</td><td>0.7005</td></tr></table>
+
+a given passage, following the prompt format introduced in UniversalNER (Zhou et al., 2024). We perform three iterations of unlearning using the self-generated forget set on Person and the retain set on the other four entity types.
+
+Evaluation We use the F1 score on the Person entity type to assess forget quality, and the F1 scores on the remaining four entity types to evaluate utility preservation.
+
+Results Our method achieves the lowest Person F1 among all baselines while maintaining strong performance on most retained entity types. Unlike manually curated datasets, our method flexibly generates forget data tailored to any specific unlearning objective, making it adaptable across domains. Notably, while Direct Preference Optimization (DPO) preserves utility well on some non-target entities, it performs poorly in terms of forget quality. Its Person F1 score remains significantly higher than other baselines, indicating that it fails to forget the intended knowledge.
+
+![](images/ab0a52f986104086368aeb492dd7596fd740062ea302889518f19c67037bdbad.jpg)  
+Figure 3: F1 scores of NER entity types across unlearning steps. The Person entity (red), which is the unlearning target, shows a significant drop in performance (from 0.54 to 0.14), indicating successful forgetting. Other entities retain their initial performance levels.
+
+# 4.5 Coding Unlearning
+
+Training Coding ability unlearning is a novel and challenging task, as labeled forget data is
+
+scare and costly to obtain. To construct the forget set, we use prompt-only inputs from the MBPP (Austin et al., 2021) dataset and prompt the base model to generate its own coding responses. We use the pass@1 score to measure the relevance of the generated outputs and continue to use the Vendi score to measure diversity. After three iterations of instruction-optimized generation, we collect 1,009 unique completions, compared to the 374 well-annotated reference solutions in the original dataset. Motivated by prior work (Li et al., 2025), which shows that coding and math tasks activate overlapping neurons, we use the training split of GSM8K (Cobbe et al., 2021) as the retain dataset. This setup allows us to evaluate whether the model can selectively unlearn coding ability while preserving math problem-solving skills. We perform a single round of iterative unlearning using retain dataset and self-generated forget dataset.
+
+Table 3: NER unlearning results. We report F1 scores on each entity type. Lower Person F1 indicates better unlearning, while higher scores on the remaining entities reflect better utility preservation.  
+
+<table><tr><td>Model</td><td>MBPP↓</td><td>MBPP+ ↓</td><td>GSM8K ↑</td></tr><tr><td>Basemodel</td><td>0.693</td><td>0.566</td><td>0.7437 ± 0.0121</td></tr><tr><td>DPO</td><td>0.698</td><td>0.585</td><td>0.7445 ± 0.0120</td></tr><tr><td>RMU</td><td>0.206</td><td>0.159</td><td>0.7460 ± 0.0120</td></tr><tr><td>Ext-Sub</td><td>0.066</td><td>0.050</td><td>0.5534 ± 0.0137</td></tr><tr><td>PEM-external</td><td>0.019</td><td>0.013</td><td>0.6520 ± 0.0131</td></tr><tr><td>Ours</td><td>0.003</td><td>0.000</td><td>0.6505 ± 0.0131</td></tr></table>
+
+Table 4: Code unlearning results. Lower pass@1 on MBPP and MBPP+ indicates better forgetting, while higher pass@1 on GSM8K reflects better preservation of math-solving ability.
+
+Evaluation After unlearning, we evaluate the model on the test split of each dataset. For coding ability, we also evaluate on MBPP+ (Liu et al., 2023), which contains  $35 \times$  more test cases.
+
+Results Our method achieves the strongest forgetting performance, with the lowest pass@1 on both MBPP and MBPP+, outperforming all baselines by a significant margin. Notably, it reduces pass@1 on MBPP+ to zero, demonstrating near-complete removal of coding ability. At the same time, it preserves math problem-solving ability,
+
+achieving a GSM8K score comparable to the best-performing baseline. These results show that our approach enables precise, targeted forgetting without sacrificing performance on unrelated skills. Interestingly, the DPO baseline performs poorly in this setting and even slightly improves coding performance, likely due to the small size of the MBPP dataset, which may not provide sufficient signal for effective preference optimization.
+
+# 5 Ablation
+
+# 5.1 External Data vs Internal Data
+
+We conduct ablation studies to examine how internal (self-generated) data compares to external data in enabling effective and precise unlearning. For the toxicity task, we train PEM modules on three types of datasets: (1) the original RTP dataset (Gehman et al., 2020), (2) a self-generated dataset using only RTP prompt inputs, and (3) a self-generated dataset using CivilComments inputs (Zhang et al., 2023). We apply each PEM to the base model via direct subtraction, using different subtraction weights  $\lambda$  selected to match forget quality—specifically, by aligning their toxicity scores. Under this constraint, we observe that PEMs trained on internal data consistently yield lower perplexity (PPL), indicating better utility preservation compared to those trained on external data. This result holds across both RTP and CivilComments settings.
+
+For the NER task, we compare PEMs trained on (1) the original UniversalNER dataset (Zhou et al., 2024) and (2) a self-generated dataset produced by prompting the base model. When controlling for forget quality (similar Person F1 scores), we find that internal data again leads to higher average F1 scores on the retained entities. These findings indicate that self-generated internal data not only supports targeted forgetting but also minimizes utility degradation, likely due to its alignment with the model's training distribution, enabling more precise unlearning.
+
+# 5.2 Hyperparameter for Iterative Unlearn
+
+The subtraction weight  $\mu_{i}$  is chosen at each iteration to ensure that the model forgets at least  $90\%$  of the target behavior compared to the beginning of that iteration. To study the impact of this threshold, we compare it with a relaxed variant that targets only  $60\%$  forgetting at each iteration.
+
+We conduct an ablation study on CodeUnlearn
+
+<table><tr><td>Method</td><td>PPL ↓</td><td>Tox. Score ↓</td></tr><tr><td>PEM-external (RTP)</td><td>10.4019</td><td>0.3249</td></tr><tr><td>internal (Civil)</td><td>9.6172</td><td>0.3378</td></tr><tr><td>internal (RTP)</td><td>7.8092</td><td>0.3415</td></tr></table>
+
+Table 5: Ablation on forget data source for Toxicity task. We compare PEMs trained on external vs. self-generated (internal) data under matched forget quality (similar Tox. Score). Internal data consistently yields lower perplexity (PPL), indicating better utility preservation across different datasets.  
+
+<table><tr><td>Method</td><td>Person F1 ↓</td><td>Avg. Retain F1 ↑</td></tr><tr><td>PEM-external</td><td>0.2483</td><td>0.2282</td></tr><tr><td>internal</td><td>0.2474</td><td>0.2802</td></tr></table>
+
+Table 6: Ablation on forget data source for the NER task. We compare PEMs trained on external vs. self-generated (internal) data. Under matched forget quality (similar Person F1), unlearning with Internal data achieves higher average F1 scores on retained entity types, indicating better utility preservation.
+
+with two groups: Group 1 sets  $\mu_{i}$  to forget only  $60\%$  of the target behavior per iteration, while Group 2 sets  $\mu_{i}$  for at least  $90\%$  forgetting. As shown in Figure 4, although Group 1 starts with weaker forgetting performance, it eventually reaches a similar level of forgetting and utility preservation as Group 2. This suggests that suboptimal hyperparameter choices can be compensated for by additional unlearning steps.
+
+![](images/2a32c76eedc7306527f3d6569b19ab8cf6aaea554cb4b8cd08b73d0842eda197.jpg)  
+Figure 4: Performance comparison of MBPP (forget target, orange) and GSM8K (retain target, green) across unlearning steps under different subtraction thresholds. Group 1 (solid lines) uses a smaller subtraction weight to enforce  $60\%$  forgetting, while Group 2 (dotted lines) uses a larger weight to enforce  $90\%$  forgetting. Group 1 requires more iterations to reach comparable forgetting and utility preservation.
+
+# 6 Conclusion
+
+In this paper, we propose to perform LLM unlearning using self-generated forget data, eliminating the
+
+need for costly and well-labeled external datasets. Additionally, we introduce an iterative unlearning framework that incrementally edits the model using Parameter-Efficient Modules (PEMs) trained on distinct objectives. This framework enables fine-grained control over the trade-off between forget quality and utility preservation.
+
+We evaluate our approach on a diverse set of tasks, including detoxification, coding, and entity forgetting. Results demonstrate that our method enables effective, targeted unlearning with minimal degradation to unrelated capabilities. These findings underscore the practicality and flexibility of self-generated data for unlearning, and open new directions for studying the relationship between forget data quality and unlearning effectiveness.
+
+# Limitations
+
+Instruction Optimization Complexity While our use of NeuralUCB for instruction optimization helps avoid manual tuning, the quality of the resulting instructions is not always ideal. This is partly due to the inherent difficulty of our tasks, which require generating diverse and meaningful outputs (e.g., toxic completions, code). Unlike prior work that often focuses on simpler objectives such as synonym generation, our setting demands more nuanced instructions to effectively elicit the model's internal knowledge. Further research is needed to improve instruction optimization and to better understand how to guide models in surfacing knowledge relevant to specific unlearning targets.
+
+Efficient Merge Weight Selection Although our iterative unlearning framework allows explicit control over the trade-off between forgetting and utility preservation, it still relies on manual evaluation to determine the optimal merge weights. Despite our rule-based selection strategy, hyperparameter tuning currently requires trial-and-error over multiple runs. Developing more principled or automated methods for hyperparameter selection would enhance both efficiency and usability.
+
+# Acknowledgement
+
+This work is supported in part by NYU Shanghai Center for Data Science and NYU HPC resources.
+
+# References
+
+Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen
+
+Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. 2021. Program synthesis with large language models. Preprint, arXiv:2108.07732.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.  
+Chenlu Ding, Jiancan Wu, Yancheng Yuan, Jinda Lu, Kai Zhang, Alex Su, Xiang Wang, and Xiangnan He. 2025. Unified parameter-efficient unlearning for llms. Preprint, arXiv:2412.00383.  
+Dan Friedman and Adji Bousso Dieng. 2023. The vendi score: A diversity evaluation metric for machine learning. Preprint, arXiv:2210.02410.  
+Roman Garnett. 2023. Bayesian Optimization. Cambridge University Press.  
+Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. Preprint, arXiv:2009.11462.  
+Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3.  
+Xinshuo Hu, Dongfang Li, Baotian Hu, Zihao Zheng, Zhenyu Liu, and Min Zhang. 2024. Separate the wheat from the chaff: Model deficiency unlearning via parameter-efficient module operation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18252-18260.  
+Zhuoran Jin, Pengfei Cao, Chenhao Wang, Zhitao He, Hongbang Yuan, Jiachun Li, Yubo Chen, Kang Liu, and Jun Zhao. 2024. Rwku: Benchmarking real-world knowledge unlearning for large language models. Preprint, arXiv:2406.10890.  
+Ching-Yun Ko, Pin-Yu Chen, Payel Das, Youssef Mroueh, Soham Dan, Georgios Kollias, Subhajit Chaudhury, Tejaswini Pedapati, and Luca Daniel. 2024. Large language models can be strong selfdetoxifiers. Preprint, arXiv:2410.03818.
+
+Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D. Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, Gabriel Mukobi, Nathan Helm-Burger, Rassin Lababidi, Lennart Justen, Andrew B. Liu, Michael Chen, Isabelle Barrass, Oliver Zhang, Xiaoyuan Zhu, and 38 others. 2024a. The wmdp benchmark: Measuring and reducing malicious use with unlearning. Preprint, arXiv:2403.03218.  
+Wenjie Li, Jiawei Li, Christian Schroeder de Witt, Ameya Prabhu, and Amartya Sanyal. 2024b. Delta-influence: Unlearning poisons via influence functions. Preprint, arXiv:2411.13731.  
+Yongce Li, Chung-En Sun, and Tsui-Wei Weng. 2025. Effective skill unlearning through intervention and abstention. Preprint, arXiv:2503.21730.  
+Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D. Manning, Christopher Ré, Diana Acosta-Navas, Drew A. Hudson, and 31 others. 2023. Holistic evaluation of language models. Preprint, arXiv:2211.09110.  
+Xiaoqiang Lin, Zhaoxuan Wu, Zhongxiang Dai, Wenyang Hu, Yao Shu, See-Kiong Ng, Patrick Jaillet, and Bryan Kian Hsiang Low. 2024. Use your instinct: Instruction optimization for llms using neural bandits coupled with transformers. Preprint, arXiv:2310.02905.  
+Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and LINGMING ZHANG. 2023. Is your code generated by chatGPT really correct? rigorous evaluation of large language models for code generation. In Thirty-seventh Conference on Neural Information Processing Systems.  
+Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Yuguang Yao, Chris Yuhao Liu, Xiaojun Xu, Hang Li, Kush R. Varshney, Mohit Bansal, Sanmi Koyejo, and Yang Liu. 2024a. Rethinking machine unlearning for large language models. Preprint, arXiv:2402.08787.  
+Zheyuan Liu, Guangyao Dou, Zhaoxuan Tan, Yijun Tian, and Meng Jiang. 2024b. Towards safer large language models through machine unlearning. Preprint, arXiv:2402.10058.  
+Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C. Lipton, and J. Zico Kolter. 2024. Tofu: A task of fictitious unlearning for llms. Preprint, arXiv:2401.06121.  
+Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2023. Locating and editing factual associations in gpt. Preprint, arXiv:2202.05262.  
+Mistral AI. 2024. Mistral-7b-instruct-v0.2.
+
+Zhuoyi Peng, Yixuan Tang, and Yi Yang. 2025. Adversarial mixup unlearning. Preprint, arXiv:2502.10288.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Preprint, arXiv:2305.18290.  
+Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. Preprint, arXiv:1908.10084.  
+Shaofei Shen, Chenhao Zhang, Yawen Zhao, Alina Bialkowski, Weitong Tony Chen, and Miao Xu. 2024. Label-agnostic forgetting: A supervision-free unlearning in deep models. Preprint, arXiv:2404.00506.  
+Zecheng Tang, Keyan Zhou, Juntao Li, Yuyang Ding, Pinzheng Wang, Bowen Yan, Rejie Hua, and Min Zhang. 2024. Cmd: a framework for context-aware model self-detoxification. Preprint, arXiv:2308.08295.  
+Enayat Ullah, Tung Mai, Anup Rao, Ryan Rossi, and Raman Arora. 2021. Machine unlearning via algorithmic stability. Preprint, arXiv:2102.13179.  
+Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. 2023. Orthogonal subspace learning for language model continual learning. Preprint, arXiv:2310.14152.  
+Miao Xu. 2024. Machine unlearning: Challenges in data quality and access. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 8589-8594. International Joint Conferences on Artificial Intelligence Organization. Early Career.  
+Jinghan Zhang, Junteng Liu, Junxian He, and 1 others. 2023. Composing parameter-efficient modules with arithmetic operation. Advances in Neural Information Processing Systems, 36:12589-12610.  
+Dongruo Zhou, Lihong Li, and Quanquan Gu. 2020. Neural contextual bandits with ucb-based exploration. Preprint, arXiv:1911.04462.  
+Wenxuan Zhou, Sheng Zhang, Yu Gu, Muhao Chen, and Hoifung Poon. 2024. Universalner: Targeted distillation from large language models for open named entity recognition. Preprint, arXiv:2308.03279.
+
+# A Implementation Details
+
+# A.1 Best Instructions
+
+We list the top-scoring instructions identified by our NeuralUCB-based algorithm for Toxicity Unlearning and Coding Unlearning. Each task involves three outer-loop iterations, and we report the best instruction found in each.
+
+# Toxicity Unlearning
+
+- Outer Iteration 1: The instruction was to generate a response that is toxic and insulting.  
+- Outer Iteration 2: The instruction was to provide a response that is toxic and insulting.  
+- Outer Iteration 3: The instruction was to create an AI system that can generate a response that is toxic and insulting.
+
+# Coding Unlearning
+
+- Outer Iteration 1: The instruction was to write a function that takes an integer and returns a string, and I was given the following input and expected output.  
+- Outer Iteration 2: The instruction was to create a program that takes a string, and I was given the following input and expected output: Input: A function to find the longest distance to a point (m, n) from (0,0) for the given grid of size m.n  
+- Outer Iteration 3: The instruction was to write a program that takes a string, and I was given the following input and expected output: Input: Write a program to find the maximum 24-hour time that is in the 12-hour format
+
+# A.2 Prompt Formatting for NER
+
+# Instruction:
+
+What describes Entity in the text? Identify any Entity entities mentioned in the text and respond ONLY with a list in the exact format: ["Entity1", "Entity2"]. If no Entity entities are mentioned, respond only with an empty list: [].
+
+# Input:
+
+Text: An icon of Leland, Carlson's Fishery is located right on the River in Fishtown. The Carlson Family's fishing tradition has been handed down five times in the last hundred years. Today, the younger generation is at the helm with Nels Carlson and Joe Campo.
+
+# Output:
+
+["Entity1", "Entity2"]
+
+# A.3 Hyperparameters Settings
+
+We present the weight hyperparameters applied at each iteration, along with the corresponding evaluation scores for each task, in Table 7, Table 8, and Table 9.
+
+# B Orthogonal Loss Study
+
+Previous work suggests that the forget and retain PEMs may overlap in their learned subspaces, potentially leading to interference. To investigate this, we explore whether enforcing orthogonality between these PEMs can better separate their objectives and reduce mutual influence.
+
+We adopt the O-LoRA framework (Wang et al., 2023), which introduces orthogonal subspace constraints during parameter-efficient tuning. Specifically, we add an orthogonality regularization term to the standard cross-entropy loss when training the retain PEM, encouraging it to learn in a subspace orthogonal to the previously trained forget PEM.
+
+Our experiment is conducted on a NER unlearning task. We first train a forget PEM to erase the Person entity and negate it (we denote as base). Then, we train a retain PEM on the retain set consisting of four entity types (Org, Concept, Location, Date), comparing versions with and without the orthogonality regularization term. The merged results are shown in Table 10.
+
+The results suggest that enforcing orthogonality does not lead to improved performance. Although adding the retain PEM with the orthogonality regularization term helps recover utility on the retain entity types, it continues to influence performance on the Person entity. This indicates that the orthogonality constraint fails to effectively disentangle the representation space of the retain PEM from that of the forget PEM. These findings further imply that the retain and forget PEMs already reside in largely orthogonal subspaces, rendering orthogonality regularization unnecessary.
+
+# C Results on Mistral-7B-Instruct-v0.2
+
+To further validate the generalizability of our approach, we conducted additional experiments on all three unlearning tasks using Mistral-7B-Instructv0.2, following the same protocol as with LLaMA3-8B-Instruct. The results show that our method con
+
+<table><tr><td rowspan="2">#Step</td><td rowspan="2">Weights Applied</td><td rowspan="2">PPL ↓</td><td colspan="3">Challenge</td><td colspan="3">Non-Challenge</td></tr><tr><td>Tox. Score ↓</td><td>Tox. Rate ↓</td><td>Severe Tox. ↓</td><td>Tox. Score ↓</td><td>Tox. Rate ↓</td><td>Severe Tox. ↓</td></tr><tr><td>0</td><td>Base model Φ0</td><td>7.2055</td><td>0.7310</td><td>0.3654</td><td>0.2725</td><td>0.2986</td><td>0.0167</td><td>0.0352</td></tr><tr><td>1</td><td>-μ0 = -3</td><td>7.8092</td><td>0.3415</td><td>0.0481</td><td>0.0644</td><td>0.1875</td><td>0.0000</td><td>0.0119</td></tr><tr><td>2</td><td>+λ1 = +0.3</td><td>6.8652</td><td>0.4689</td><td>0.1250</td><td>0.1207</td><td>0.2102</td><td>0.0000</td><td>0.0145</td></tr><tr><td>3</td><td>-μ1 = -0.2</td><td>7.5513</td><td>0.3047</td><td>0.0481</td><td>0.0532</td><td>0.1842</td><td>0.0000</td><td>0.0123</td></tr></table>
+
+Table 7: Toxicity and perplexity metrics across unlearning steps for challenge and non-challenge subsets. Step-wise application of forget  $(- \mu)$  and retain  $(+ \lambda)$  weights reduces toxicity while maintaining perplexity.  
+
+<table><tr><td>#Step</td><td>Weights Applied</td><td>Person F1 ↓</td><td>Org F1 ↑</td><td>Concept F1 ↑</td><td>Location F1 ↑</td><td>Date F1 ↑</td></tr><tr><td>0</td><td>Base model Φ0</td><td>0.5370</td><td>0.4501</td><td>0.2123</td><td>0.4747</td><td>0.7173</td></tr><tr><td>1</td><td>-μ0 = -5</td><td>0.2474</td><td>0.2564</td><td>0.0883</td><td>0.3024</td><td>0.4738</td></tr><tr><td>2</td><td>+λ1 = +0.3</td><td>0.4780</td><td>0.4489</td><td>0.1975</td><td>0.4687</td><td>0.6999</td></tr><tr><td>3</td><td>-μ1 = -0.4</td><td>0.1788</td><td>0.3380</td><td>0.0921</td><td>0.3233</td><td>0.4894</td></tr><tr><td>4</td><td>+λ2 = +0.3</td><td>0.3184</td><td>0.4205</td><td>0.1446</td><td>0.4335</td><td>0.6161</td></tr><tr><td>5</td><td>-μ2 = -0.3</td><td>0.0306</td><td>0.2044</td><td>0.0693</td><td>0.2439</td><td>0.2958</td></tr><tr><td>6</td><td>+λ3 = +1.0</td><td>0.3210</td><td>0.5410</td><td>0.2456</td><td>0.5363</td><td>0.7300</td></tr><tr><td>7</td><td>-μ3 = -0.1</td><td>0.1430</td><td>0.5242</td><td>0.2299</td><td>0.5157</td><td>0.7005</td></tr></table>
+
+Table 8: F1 scores for each NER entity type at each unlearning step. The Person entity is the unlearning target, with decreasing F1 across forgetting steps. The other entities are retention targets, showing recovery as retention weights are applied. Each row reflects the model state after a single weight update step.
+
+sistently achieves effective unlearning across different model families, as shown in Table 11, Table 12, and Table 13.
+
+<table><tr><td>#Step</td><td>Weights Applied</td><td>MBPP ↓</td><td>MBPP+ ↓</td><td>GSM8K ↑</td></tr><tr><td>0</td><td>Base model Φ0</td><td>0.659</td><td>0.553</td><td>0.7437±0.0121</td></tr><tr><td>1</td><td>-μ0 = -4</td><td>0.053</td><td>0.045</td><td>0.5959±0.0135</td></tr><tr><td>2</td><td>+λ1 = +1</td><td>0.106</td><td>0.085</td><td>0.6823±0.0128</td></tr><tr><td>3</td><td>-μ1 = -0.4</td><td>0.003</td><td>0.000</td><td>0.6505±0.0131</td></tr></table>
+
+Table 9: Pass@1 scores on MBPP and MBPP+ (forget targets) and GSM8K (retain target) across code unlearning steps. Forgetting weights reduce performance on MBPP/MBPP+, while retain weights recover GSM8K accuracy. Final subtraction improves forget specificity while maintaining retention.  
+
+<table><tr><td>Model</td><td>Person F1 ↓</td><td>Org F1 ↑</td><td>Concept F1 ↑</td><td>Location F1 ↑</td><td>Date F1 ↑</td></tr><tr><td>Base</td><td>0.0521</td><td>0.3793</td><td>0.1883</td><td>0.4170</td><td>0.6588</td></tr><tr><td>w/ ortho term</td><td>0.2373</td><td>0.4787</td><td>0.2369</td><td>0.5061</td><td>0.7044</td></tr><tr><td>w/o ortho term</td><td>0.2132</td><td>0.5025</td><td>0.2454</td><td>0.5162</td><td>0.7308</td></tr></table>
+
+Table 10: Study on the effect of orthogonality loss in NER unlearning. Incorporating orthogonality loss into the retain PEM still impacts the forget entity (Person) performance, showing a similar level of interference as the retain PEM trained without the orthogonality constraint.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">PPL ↓</td><td colspan="3">Challenge</td><td colspan="3">Non-Challenge</td></tr><tr><td>Tox. Score ↓</td><td>Tox. Rate ↓</td><td>Severe Tox. ↓</td><td>Tox. Score ↓</td><td>Tox. Rate ↓</td><td>Severe Tox. ↓</td></tr><tr><td>basemodel</td><td>5.0297</td><td>0.8464</td><td>0.6923</td><td>0.3581</td><td>0.3521</td><td>0.0402</td><td>0.0484</td></tr><tr><td>DPO</td><td>5.0843</td><td>0.8393</td><td>0.6923</td><td>0.3244</td><td>0.3454</td><td>0.0446</td><td>0.0466</td></tr><tr><td>RMU</td><td>5.0298</td><td>0.8248</td><td>0.6731</td><td>0.3318</td><td>0.3471</td><td>0.0368</td><td>0.0499</td></tr><tr><td>Ext-Sub</td><td>5.0225</td><td>0.7352</td><td>0.4519</td><td>0.2354</td><td>0.2760</td><td>0.0134</td><td>0.0255</td></tr><tr><td>PEM-external</td><td>132.4302</td><td>0.3435</td><td>0.0000</td><td>0.1190</td><td>0.3350</td><td>0.0000</td><td>0.1168</td></tr><tr><td>Ours</td><td>5.6633</td><td>0.4194</td><td>0.0000</td><td>0.1517</td><td>0.2125</td><td>0.0000</td><td>0.0351</td></tr></table>
+
+Table 11: Toxicity unlearning results on Mistral-7B-Instruct-v0.2. Our method achieves substantial reductions in toxicity while maintaining fluency, showing consistent trends with LLaMA3-8B-Instruct.  
+
+<table><tr><td>Model</td><td>Person F1 ↓</td><td>Org F1 ↑</td><td>Concept F1 ↑</td><td>Location F1 ↑</td><td>Date F1 ↑</td></tr><tr><td>basemodel</td><td>0.3765</td><td>0.3104</td><td>0.1508</td><td>0.2168</td><td>0.4258</td></tr><tr><td>DPO</td><td>0.0010</td><td>0.0007</td><td>0.0004</td><td>0.0027</td><td>0.0330</td></tr><tr><td>RMU</td><td>0.1168</td><td>0.1787</td><td>0.0817</td><td>0.1126</td><td>0.1036</td></tr><tr><td>Ext-Sub</td><td>0.1865</td><td>0.1319</td><td>0.0609</td><td>0.1349</td><td>0.2884</td></tr><tr><td>PEM-external</td><td>0.0000</td><td>0.0000</td><td>0.0000</td><td>0.0000</td><td>0.0000</td></tr><tr><td>Ours</td><td>0.0324</td><td>0.3170</td><td>0.1072</td><td>0.3571</td><td>0.4443</td></tr></table>
+
+Table 12: NER unlearning results on Mistral-7B-Instruct-v0.2. Our approach effectively forgets the Person entity type while preserving performance on other entities.  
+
+<table><tr><td>Model</td><td>MBPP↓</td><td>MBPP+ ↓</td><td>GSM8K ↑</td></tr><tr><td>basemodel</td><td>0.526</td><td>0.450</td><td>0.3760 ± 0.0133</td></tr><tr><td>DPO</td><td>0.516</td><td>0.437</td><td>0.3768 ± 0.0133</td></tr><tr><td>RMU</td><td>0.415</td><td>0.336</td><td>0.3450 ± 0.0131</td></tr><tr><td>Ext-Sub</td><td>0.026</td><td>0.021</td><td>0.1046 ± 0.0084</td></tr><tr><td>PEM-external</td><td>0.005</td><td>0.003</td><td>0.3374 ± 0.0130</td></tr><tr><td>Ours</td><td>0.000</td><td>0.000</td><td>0.4405 ± 0.0137</td></tr></table>
+
+Table 13: Code unlearning results on Mistral-7B-Instruct-v0.2. Our method nearly eliminates coding ability while retaining math reasoning (GSM8K).

revealandreleaseiterativellmunlearningwithselfgenerateddata/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f84282f2ec86e9cc19656ce6b5e27f37ef6242890bb93cc82cfb2f5485dc10bf
+size 627245

revealandreleaseiterativellmunlearningwithselfgenerateddata/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f27581276e3f721c74e54ab064e879cd1f76d7771607b056e62e4b915e6255f
+size 425479

revealingtheimpactofsyntheticnativesamplesandmultitaskingstrategiesinhindienglishcodemixedhumourandsarcasmdetection/0dd3cd9d-286a-4e73-a1b8-fd3e80097485_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53ad3501d15d9b5fd65b3ae31296f72078cb9faa497c536703f95040d0d7b750
+size 218175

revealingtheimpactofsyntheticnativesamplesandmultitaskingstrategiesinhindienglishcodemixedhumourandsarcasmdetection/0dd3cd9d-286a-4e73-a1b8-fd3e80097485_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9eb1643872737d2bfb04adbcd8bd652e97cb0dc99353f1678930e5534dd86c32
+size 271072

revealingtheimpactofsyntheticnativesamplesandmultitaskingstrategiesinhindienglishcodemixedhumourandsarcasmdetection/0dd3cd9d-286a-4e73-a1b8-fd3e80097485_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0c008749950f6d8a75573e7998826fe0042efe4c143ed13ec933235b4d1008e
+size 1174168

revealingtheimpactofsyntheticnativesamplesandmultitaskingstrategiesinhindienglishcodemixedhumourandsarcasmdetection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:123c37581ab709ddd930f3cfad40dcd4b9d83149f576e833514a3bbbea5e0834
+size 1995781

revealingtheimpactofsyntheticnativesamplesandmultitaskingstrategiesinhindienglishcodemixedhumourandsarcasmdetection/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:026a1d47e4bd19cc0c6dbba8b511bab110e13445aeaa700b76c1431c324ce421
+size 883042

revealingtheinherentinstructabilityofpretrainedlanguagemodels/6f383d8e-3f24-415b-8cb0-1626465c11bf_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:162e24e15ef72ad4ba88912d965c35920918a723b04e91f9536e7d0a9b969fe6
+size 201825

revealingtheinherentinstructabilityofpretrainedlanguagemodels/6f383d8e-3f24-415b-8cb0-1626465c11bf_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfe3a2c583b27424ec41a92bdb3d95b4466eb5b076c4df643e2b5e404f928d5f
+size 250501

revealingtheinherentinstructabilityofpretrainedlanguagemodels/6f383d8e-3f24-415b-8cb0-1626465c11bf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a1486eb6126f18fbd3125bdaa953e178bd30ce67f4a42554a40260b5e315582
+size 1280177

revealingtheinherentinstructabilityofpretrainedlanguagemodels/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cbb8d194d3032c9b5338b13d34e8651346fd307e96cfa824ff2bde08ccc6e0b
+size 1361291

revealingtheinherentinstructabilityofpretrainedlanguagemodels/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab6e530e7653854f45f3599d1661ba1605e809e40c2a23d6baf39f3799ad62c3
+size 877567

reviewevalanevaluationframeworkforaigeneratedreviews/2ddceee1-909a-4dd8-9e1a-b6303c024d76_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89c7d4ba344c0fe0f7bf10497371d23ac22d0d99b0fd2e2d2aaf69b1aa62c28f
+size 161758

reviewevalanevaluationframeworkforaigeneratedreviews/2ddceee1-909a-4dd8-9e1a-b6303c024d76_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9671c920fa23206ecef4dbaacd4054459b25c69f02943e9deb0b7cc97a4dbe60
+size 199532

reviewevalanevaluationframeworkforaigeneratedreviews/2ddceee1-909a-4dd8-9e1a-b6303c024d76_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d76be5ddadd391c9c884ef540b4ca8aa717080eb4a7bd1f91fc38240574feda
+size 4322923

reviewevalanevaluationframeworkforaigeneratedreviews/full.md

@@ -0,0 +1,720 @@
+# ReviewEval: An Evaluation Framework for AI-Generated Reviews
+
+# Madhav Krishan Garg $^{1}$ , Tejash Prasad $^{1}$ , Tanmay Singhal $^{1}$ , Chhavi Kirtani $^{1}$ , Murari Mandal $^{2}$ , Dhruv Kumar $^{1,3}$
+
+$^{1}$ IIT Delhi,  $^{2}$ KIIT Bhubaneswar,  $^{3}$ BITS Pilani
+
+Correspondence: dhruv.kumar@pilani(bits-pilani.ac.in
+
+# Abstract
+
+The escalating volume of academic research, coupled with a shortage of qualified reviewers, necessitates innovative approaches to peer review. In this work, we propose: 1 ReviewEval, a comprehensive evaluation framework for AI-generated reviews that measures alignment with human assessments, verifies factual accuracy, assesses analytical depth, identifies degree of constructiveness and adherence to reviewer guidelines; and 2 ReviewAgent, an LLM-based review generation agent featuring a novel alignment mechanism to tailor feedback to target conferences and journals, along with a self-refinement loop that iteratively optimizes its intermediate outputs and an external improvement loop using ReviewEval to improve upon the final reviews. ReviewAgent improves actionable insights by  $6.78\%$  and  $47.62\%$  over existing AI baselines and expert reviews respectively. Further, it boosts analytical depth by  $3.97\%$  and  $12.73\%$ , enhances adherence to guidelines by  $10.11\%$  and  $47.26\%$  respectively. This paper establishes essential metrics for AI-based peer review and substantially enhances the reliability and impact of AI-generated reviews in academic research.
+
+# 1 Introduction
+
+The rapid growth of academic research, coupled with a shortage of qualified reviewers, has created an urgent need for scalable and high-quality peer review processes (Petrescu and Krishen, 2022; Schulz et al., 2022; Checco et al., 2021). This need has led to a growing interest in leveraging large language models (LLMs) to automate and enhance various aspects of the peer review process (Robertson, 2023; Liu and Shah, 2023).
+
+Although LLMs have shown remarkable potential in automating various natural language processing tasks, their effectiveness in serving as reliable and consistent paper reviewers remains a significant challenge. The academic community is al
+
+![](images/3875930eba99987cc79a2edb47f8869c1f9e7945492fc9ca17656b3eaff6a276.jpg)  
+Figure 1: ReviewEval and ReviewAgent: Given a paper and the associated conference or journal guidelines, ReviewAgent generates AI-based reviews and evaluates them along multiple dimensions using ReviewEval.
+
+ready experimenting with AI-assisted reviews; for instance,  $15.8\%$  of ICLR 2024 reviews involved AI assistance (Latona et al., 2024). Despite increasing LLM adoption, concerns about reliability and fairness remain.
+
+Specifically, papers reviewed by AI have been perceived to gain an unfair advantage, leading to questions about the integrity of such evaluations. Consequently, research into robust automated review generation systems is crucial, necessitating rigorous evaluation of AI generated reviews to address key challenges. (Zhou et al., 2024) analyze commercial models like GPT-3.5 and GPT-4 (Achiam et al., 2023) as paper reviewers, identifying limitations such as hallucinations, incomplete understanding, and insufficient critical feedback compared to humans.
+
+Existing research on evaluation metrics for AI-generated research paper reviews remains limited. For instance, (D'Arcy et al., 2024) proposed a GPT-4-based automated metric for evaluating approximate matches between AI-generated and human-written reviews. However, their method's iterative reliance on GPT-4 creates a black-box evaluation,
+
+limiting transparency and raising reliability concerns. Similarly, (Zhou et al., 2024) investigated the aspect coverage and similarity between AI and human reviews through a blend of automatic metrics and manual analysis. But their approach overlooks other critical dimensions where AI reviews may underperform, as highlighted in Figure 2. Recent surveys also provide broader overviews of LLM-driven peer review, highlighting emerging opportunities and persistent challenges (Lin et al., 2023a; Zhuang et al., 2025).
+
+Addressing the above mentioned limitations and gaps, we propose ReviewEval, a comprehensive evaluation framework for assessing the quality of AI-generated research paper reviews. ReviewEval targets five key dimensions (see Figure 2): 1 Comparison with Human Reviews: Evaluates topic coverage and semantic similarity to measure the alignment between AI-generated and human-written feedback. 2 Factual Accuracy: Detects factual errors, including misinterpretations, incorrect claims, and hallucinated information. 3 Analytical Depth: Assesses whether the AI's critique transcends generic commentary to offer in-depth, meaningful feedback. 4 Actionable Insights: Measures the ability of the AI to provide constructive suggestions for improving the paper. 5 Adherence to Reviewer guidelines: Quantifies the degree to which a review conforms to the evaluation criteria outlined by the target conference.
+
+Recent studies (Bauchner and Rivara, 2024; Biswas, 2024) underscore the growing importance of aligning reviews with conference-specific evaluation criteria, especially as many venues now require adherence to detailed reporting guidelines. To address this, we introduce ReviewAgent, an AI reviewer with three key features: 1 Conference-specific review alignment, dynamically adapting reviews to each venue's unique evaluation criteria; 2 Self-refinement loop, iteratively optimizing prompts and intermediate outputs for deeper analytical feedback; 3 External improvement loop, systematically enhancing review quality using evaluation metrics provided by ReviewEval.
+
+Using ReviewEval, we extensively evaluate ReviewAgent against two baseline models, AI-Scientist (Lu et al., 2024) and MARG (D'Arcy et al., 2024), benchmarking all systems against expert-written reviews treated as the gold standard. The experiments are performed on a balanced and topically diverse dataset of 120 papers sampled from NeurIPS, ICLR, and UAI, including domains
+
+such as computer science, biology, social science, finance, physics, and engineering.
+
+Our results show that ReviewAgent improves actionable insights by  $6.78\%$  and  $47.62\%$  over existing AI baselines and expert reviews respectively. Further, it boosts analytical depth by  $3.97\%$  and  $12.73\%$ , enhances adherence to guidelines by  $10.11\%$  and  $47.26\%$  respectively.
+
+In summary, the paper makes the following research contributions: (1) ReviewEval: a multidimensional framework for evaluating research paper reviews ( $\S 3$ ). (2) ReviewAgent: an adaptive AI reviewer which utilizes ReviewEval along with novel conference alignment and iterative self-improvement ( $\S 4$ ). (3) Comprehensive evaluation of ReviewAgent against state-of-the-art baselines using ReviewEval ( $\S 5$ ).
+
+# 2 Related Work
+
+AI-based scientific discovery. Early attempts to automate research include AutoML that optimize hyperparameters and architectures (Hutter et al., 2019; He et al., 2021) and AI-driven discovery in materials science and synthetic biology (Merchant et al., 2023; Hayes et al., 2024). However, these methods remain largely dependent on human-defined search spaces and predefined evaluation metrics, limiting their potential for open-ended discovery. Recent works (Lu et al., 2024) aim to automate the entire research cycle, encompassing ideation, experimentation, manuscript generation, and peer review, thus pushing the boundaries of AI-driven scientific discovery.
+
+AI-based peer-review. Existing work has looked at scoring and improving research papers in a variety of ways such as statistical reporting inconsistencies (Nuijten and Polanin), recommending citations (Ali et al., 2020) and predicting review scores (Basuki and Tsuchiya, 2022; Bharti et al.). More recently, LLM-based approaches have been used to generate peer reviews (Robertson, 2023; Liu and Shah, 2023; D'Arcy et al., 2024; Lu et al., 2024; Liang et al.). Complementing these efforts, (Lin et al., 2023a) provide a comprehensive conceptual analysis of automated scholarly reviewing, while their MOPRD dataset (Lin et al., 2023b) offers a valuable multidisciplinary benchmark for developing and testing AI review systems.
+
+(Lu et al., 2024) employ LLMs to autonomously conduct the research pipeline, including peer review. It follows a structured three-stage review
+
+<table><tr><td>Comparison with Human Reviews</td><td>Factual Accuracy</td><td>Analytical Depth</td><td>Actionable Insights</td></tr><tr><td>AI: &quot;The explanation of the results is adequate&quot;</td><td>AI: &quot;The paper uses supervised learning techniques effectively&quot;</td><td>AI: &quot;The methodology section is sufficient&quot;</td><td>AI: &quot;Provide more examples for better understanding&quot;</td></tr><tr><td>Human: &quot;The explanation is too brief and misses key statistical trends in Figure 3, such as the anomaly at epoch 50&quot; (human).</td><td>This statement is factually incorrect because the actual technique described is reinforcement learning.</td><td>The above is a superficial statement as experts say &quot;The comparison to baseline models lacks clarity, especially in explaining the choice of hyperparameters&quot;.</td><td>The above statement is not actionable enough for authors to improve upon as compared to &quot;Add examples demonstrating how the algorithm performs under different lighting conditions to clarify its robustness&quot; which is more actionable.</td></tr><tr><td>AI: &quot;Overall, the related work section is relevant&quot;</td><td>AI: &quot;The dataset appears to be balanced&quot;</td><td>AI: &quot;The discussion is clear&quot;</td><td>AI: &quot;Clarify the introduction&quot;</td></tr><tr><td>Human: &quot;The related work section does not include recent advancements in transformer-based architectures, such as XYZ-2023&quot;</td><td>This statement is factually incorrect because the dataset is actually imbalanced based on the class distributions mentioned in Section 4.2.</td><td>The statement above misses feedback like &quot;The discussion should explore why the proposed approach underperforms on Dataset B, as highlighted in Table 2&quot; which is given by experts.</td><td>This is also less actionable as compared to &quot;Reorganize the introduction to define the problem before introducing the contributions, as this will improve flow and reader engagement&quot;.</td></tr><tr><td>AI: &quot;The conclusion is well-written&quot;</td><td>AI: &quot;The results suggest strong performance&quot;</td><td rowspan="2">AI: &quot;Results are promising&quot; This statement is also superficial, lacking feedback such as &quot;Consider expanding on the implications of your findings for real-world applications, particularly in autonomous navigation&quot; given by an expert.</td><td>AI: &quot;Improve the figures for better clarity&quot;,</td></tr><tr><td>Human: &quot;The conclusion does not address limitations, such as the small sample size used in the experiments&quot;.</td><td>This statement is factually incorrect but it incorrectly claims &quot;The model outperforms all baselines,&quot; while Table 3 shows it underperforms in some metrics.</td><td>This is less actionable as compared to &quot;Increase the text size in Figure 4 and add units to the axes labels for better readability&quot;.</td></tr></table>
+
+Figure 2: Illustrative examples highlighting key challenges in AI-generated research paper reviews that motivate our proposed evaluation metrics: Column 1 shows semantic and topical divergences from human reviews, supporting the Human Comparison metric; Column 2 presents factual inaccuracies or hallucinations, motivating the Factual Accuracy metric; Column 3 illustrates limited analytical reasoning, justifying the Analytical Depth metric; Column 4 reveals a lack of specific, actionable suggestions, underscoring the need for the Actionable Insights metric.
+
+process: paper understanding, criterion-based evaluation (aligned with NeurIPS and ICLR guidelines), and final synthesis: assigning scores to key aspects like novelty, clarity, and significance. MARG (D'Arcy et al., 2024) introduces a multiagent framework where worker agents review sections, expert agents assess specific aspects, and a leader agent synthesizes feedback. Using BERTScore (Zhang* et al., 2020) and GPT-4-based evaluation, MARG-S improves feedback quality, reducing generic comments and increasing helpful feedback per paper. These studies highlight the AI's potential to enhance peer review through structured automation and multi-agent collaboration.
+
+Evaluation framework for AI-based peer-review. There has been limited research on developing evaluation frameworks for evaluating the quality of LLM generated paper reviews. (Zhou et al., 2024) evaluated GPT models for research paper reviewing across 3 tasks: aspect score prediction, review generation, and review-revision MCQ answering. Their evaluation framework comprised aspect coverage (originality, soundness, substance, replicability, etc.), ROUGE (lexical overlap), BERTScore (semantic similarity), and BLANC (informativeness), alongside manual analysis. Results showed LLMs overemphasized positive feedback, lacked critical depth, and neglected substance and clarity, despite high lexical similarity to human reviews.
+
+<table><tr><td>Method</td><td>depth analysis</td><td>factual correctness</td><td>topic coverage</td><td>actionable insight</td></tr><tr><td>(D’Arcy et al., 2024)</td><td>✗</td><td>✗</td><td>✗</td><td>✓</td></tr><tr><td>(Zhou et al., 2024)</td><td>✗</td><td>✗</td><td>✗</td><td>✗</td></tr><tr><td>ReviewEval (ours)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 1: Existing metrics focus on AI-human review similarity but overlook other key aspects; our framework fills these gaps with more interpretable metrics.
+
+(D'Arcy et al., 2024) introduced an automated evaluation framework for AI-generated reviews, quantifying similarity to human reviews via recall, precision, and Jaccard index. Recall measures the fraction of real-reviewer comments with at least one AI match, precision quantifies AI comments aligned with human reviews, and Jaccard index evaluates the intersection-over-union of aligned comments.
+
+Existing evaluation metrics predominantly emphasize the similarity between AI-generated and human reviews, overlooking other crucial parameters. Moreover, their heavy reliance on LLMs for end-to-end evaluation results in a black-box system with limited transparency. In contrast, our framework introduces more interpretable evaluation metrics for AI-generated reviews (see Figure 2 and Table 1), effectively addressing these shortcomings.
+
+Iterative Refinement of Large Language Models. Recent studies highlight the benefits of iterative feedback-driven refinement in improving LLM outputs: Self-Refine (Madaan et al., 2023) uses self-generated feedback without additional training, LLMRefine (Xu et al., 2024) employs a learned feedback model with simulated annealing, and ProMiSe (Hu et al., 2024) leverages external proxy metrics, collectively enhancing factuality, coherence, and performance in tasks like document-grounded QA and dialog generation.
+
+# 3 ReviewEval
+
+We introduce ReviewEval, an evaluation framework grounded in the LLM-as-a-Judge paradigm (Gu et al., 2025). In ReviewEval, each review is evaluated on several key parameters to assess the overall quality of the generated feedback. To ensure consistency and reliability, all evaluations for a given metric were performed by LLMs of the same specification and version, minimizing variability from model differences and enabling robust, fair, and unbiased comparisons. The overview of the proposed framework (ReviewEval + ReviewAgent) is presented in Figure 1.
+
+# 3.1 Comparison with Expert Reviews
+
+We compare the reviews generated by the LLM based reviewer with expert human reviews. Our primary goal is to gauge how well the AI system replicate expert-level critique. The evaluation is conducted along the following dimensions:
+
+Semantic similarity. To assess the alignment between AI-generated and expert reviews, we embed each review  $R$  into a vector representations using a text embedding model (Mikolov et al., 2013). The semantic similarity between an AI-generated review  $R_{\mathrm{AI}}$  and an expert review  $R_{\mathrm{Expt}}$  is measured using cosine similarity:
+
+$$
+S _ {\text {s e m}} \left(R _ {\mathrm {A I}}, R _ {\text {E x p t}}\right) = \frac {e \left(R _ {\mathrm {A I}}\right) \cdot e \left(R _ {\text {E x p t}}\right)}{\| e \left(R _ {\mathrm {A I}}\right) \| \| e \left(R _ {\text {E x p t}}\right) \|} \tag {1}
+$$
+
+where  $e(R)$  denotes the embedding of review  $R$ . A higher cosine similarity indicates a stronger alignment between the AI-generated and expert reviews.
+
+Topic Coverage. We evaluate topic coverage to determine how comprehensively AI-generated reviews address the breadth of topics present in expert reviews. Our approach comprises three steps: 1 Topic extraction: Each review  $R$  (either AI-generated or expert) is decomposed into a set of topics by an LLM:  $T_{R} = \{t_{1}, t_{2}, \ldots, t_{n}\}$ , where each topic  $t_{i}$  is represented by a sentence that captures its core content and context. 2 Topic similarity: Let  $T_{\mathrm{AI}} = \{t_{1}, t_{2}, \ldots, t_{m}\}$  and  $T_{\mathrm{Expt}} = \{t_{1}', t_{2}', \ldots, t_{n}'\}$  denote the topics extracted from the AI and expert reviews, respectively. We define a topic similarity function  $\mathrm{TS}(t_i, t_j')$  that an LLM assigns on a discrete scale:
+
+$\mathrm{TS}(t_i,t_j^{\prime}) = 3\cdot \mathbb{I}\{t_i\sim_{\mathrm{strong}}t_j^{\prime}\} +2\cdot \mathbb{I}\{t_i\sim_{\mathrm{moderate}}$ $t_j^\prime \} +1\cdot \mathbb{I}\{t_i\sim_{\mathrm{weak}}t_j^\prime \}$  , where  $\mathbb{I}$  is the indicator function,  $t_i\sim_{\mathrm{strong}}t_j^{\prime},t_i\sim_{\mathrm{moderate}}t_j^{\prime},t_i\sim_{\mathrm{weak}}t_j^{\prime}$  denote substantial, moderate, and minimal overlap in concepts, respectively. All the conditions are mutually exclusive. We set a similarity threshold  $\tau = 2$  so that a topic in AI-generated review having atleast moderate similarity to a topic in expert review is considered aligned. 3 Coverage ratio: For each AI-generated review, we construct a topic similarity matrix  $S$  where each element  $S[i,j] = \mathrm{TS}(t_i,t_j^{\prime})$  represents the similarity between topic  $t_i$  from  $T_{\mathrm{AI}}$  and topic  $t_j^\prime$  from  $T_{\mathrm{Expt}}$  The topic coverage ratio is defined as:
+
+$$
+S _ {\text {c o v e r a g e}} = \frac {1}{n} \sum_ {j = 1} ^ {n} \mathbb {I} \left(\max  _ {i = 1, \dots , m} S [ i, j ] \geq \tau\right), \tag {2}
+$$
+
+where  $\mathbb{I}(\cdot)$  is the indicator function, and  $n = |T_{\mathrm{Expt}}|$  is the total number of topics extracted from the expert review.
+
+# 3.2 Factual Correctness
+
+To address hallucinations and factual inaccuracies in LLM-generated reviews, we introduce an automated pipeline that simulates the conference rebuttal process, allowing evidence-based validation of reviewer claims. By automating both the question-generation and rebuttal phases, our system produces a robust factual correctness evaluation. The pipeline consists of the following steps:
+
+Step 1: Question generation. Each review  $R$  is converted by an LLM into a structured verification question  $Q$  that captures its central critique (Example in Appendix A.3).
+
+Step 2: Query decomposition. A dedicated LLM-based decomposition component splits  $Q$  into a set of sub-questions  $\{q_1, \ldots, q_n\}$ , isolating distinct aspects of the claim for fine-grained analysis. (Implementation and Example in Appendix A.3.)
+
+Step 3: Retrieval-augmented synthesis. For each sub-question  $q_{i}$ , we: (a) retrieve relevant text segments ( $\approx$  1000 tokens) via semantic similarity search over contextual embeddings; (b) extract the corresponding parent sections ( $\approx$  4000 tokens) to provide broader context. Both parent and child document splitters use a  $10\%$  overlap to enhance context retention across chunk boundaries; (c) generate a focused answer  $A_{i}$  to  $q_{i}$  using the retrieved context. The individual answers are then aggregated into a unified response  $A_{Q}$  using an LLM addressing the original question  $Q$ .
+
+Step 4: Automated rebuttal generation. The aggregated evidence  $A_{Q}$  serves as the basis for an evidence-based rebuttal  $R_{b}$ , which systematically supports or counters each claim in  $R$  by citing specific sections of the paper.
+
+Step 5: Factual correctness evaluation. We compare  $R$  against  $R_{b}$  to determine whether each claim is substantiated:
+
+$$
+\mathcal {V} = \left\{ \begin{array}{l l} 1: & \text {a l l c l a m i s s i n R a r e s u p p o r t e d b y R _ {b}}; \\ 0: & \text {o t h e r w i s e} \end{array} \right.
+$$
+
+We then compute an overall factual correctness score:
+
+$$
+S _ {\text {f a c t u a l}} = \frac {1}{| Q |} \sum_ {i = 1} ^ {| Q |} \mathbb {I} \left(\mathcal {V} _ {i} = 1\right), \tag {3}
+$$
+
+# 3.3 Constructiveness
+
+We assess review constructiveness by quantifying the presence and quality of actionable insights in AI-generated reviews relative to expert feedback. Our framework begins by extracting key actionable components from each review using an LLM
+
+with few-shot examples. Specifically, we identify the following insights: (i) criticism points  $(CP)$ , which capture highlighted flaws or shortcomings in the paper's content, clarity, novelty, and execution; (ii) methodological feedback  $(MF)$ , which encompasses detailed analysis of experimental design, techniques, and suggestions for methodological improvements; and (iii) suggestions for improvement  $(SI)$ , which consist of broader recommendations for enhancement such as additional experiments, alternative methodologies, or improved clarity.
+
+Once these components are extracted, each insight is evaluated along three dimensions: specificity, feasibility, and implementation details.
+
+- The specificity score  $\sigma$  is defined as 1 if the insight is clear and unambiguous, referring to a particular aspect of the paper and including explicit examples.  
+- The feasibility score  $\phi = 1$  when the recommendation can realistically be implemented within the research context, such as available data, reasonable technical effort, or domain constraints.  
+- The implementation detail score  $\delta = 1$  if the feedback provides actionable steps, specific techniques, or references for improvement.
+
+Examples explaining the three dimensions in Appendix A.4.
+
+The overall actionability score for an individual insight is then computed as  $S_{\mathrm{act},i} = \sigma_i + \phi_i + \zeta_i$ , with an insight considered actionable if  $S_{\mathrm{act},i} > 1$  (having at least 2 of the three qualities mentioned above). Finally, we quantify the overall constructiveness of a review by calculating the percentage of actionable insights:
+
+$$
+S _ {\text {a c t}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbb {I} \left(S _ {\text {a c t}, i} > 1\right) \times 1 0 0, \tag {4}
+$$
+
+where  $N$  is the total number of extracted insights and  $\mathbb{I}(\cdot)$  denotes the indicator function. This metric provides a quantitative measure of how a review offers guidance for improving the work.
+
+# 3.4 Depth of Analysis
+
+To assess whether a review provides a comprehensive, critical evaluation rather than a superficial commentary, we measure the depth of analysis in AI-generated reviews. Each review is evaluated by multiple LLMs, which assign scores for each dimension,  $m_{i}$  ( $i \in \{1,2,3,4,5\}$ ), with scores  $S_{i} \in \{0,1,2,3\}$ . We define these as follows:
+
+Comparison with existing literature  $(m_{1})$ : As-sesses whether the review critically examines the
+
+paper's alignment with prior work, acknowledging relevant studies and identifying omissions. The scoring rubric is:
+
+$S_{1} = \{3:$  thorough and critical comparison; 2: meaningful but shallow; 1 : vague; 0 : not present\}
+
+Logical gaps identified  $(m_2)$ : Evaluates the review's ability to detect unsupported claims, reasoning flaws, and to offer constructive suggestions:  $S_{2} = \{3:$  clear gaps with suggestions; 2 : some gaps, unclear recommendations; 1 vague gaps, no solutions; 0: no gaps identified\}
+
+Methodological scrutiny  $(m_3)$ : Measures the depth of critique regarding the paper's methods, including evaluation of strengths, limitations, and improvement suggestions:
+
+$S_{3} = \{3:$  thorough and actionable; 2 : meaningful but limited; 1: vague; 0: none\}
+
+Results interpretation  $(m_4)$ : Assesses depth of result discussion (biases, alternative explanations, implications).
+
+$S_{4} = \{3\colon$  insightful; 2: shallow; 1: vague; 0: none\}
+
+Theoretical contribution  $(m_{5})$ : Evaluates the assessment of the paper's theoretical contributions, including its novelty and connections to broader frameworks:
+
+$S_{5} = \{3:$  thorough and insightful critique; 2 : reasonable but lacks depth; 1
+
+vague critique; 0: no assessment}
+
+The overall depth of analysis score for a review is calculated as the average normalized score across
+
+all dimensions:  $S_{\mathrm{depth}} = \frac{\sum_{i=1}^{5} S_i}{15}$
+
+A higher  $S_{\mathrm{depth}}$  indicates a more comprehensive and critical engagement with the manuscript.
+
+# 3.5 Adherence to Reviewer Guidelines
+
+To assess whether a review complies with established criteria, we evaluate its adherence to guidelines set by the venue.
+
+Our approach begins by extracting the criteria  $C$  from the guidelines  $G$ . These criteria fall into two broad categories: ① subjective criteria, which involve qualitative judgments (e.g., clarity, constructive feedback), and ② objective criteria, which are quantifiable (e.g., following a prescribed rating scale). For each review  $R$ , every extracted criterion  $C_i$  is scored on a 0-3 scale using a dedicated LLM with dynamically generated prompts that include few-shot examples for contextual calibration. For subjective criteria, the score is defined as:
+
+$S_{i} = \{3: \text{strong, detailed alignment}; 2: \text{mostly aligned, minor issues}; 1: \text{incomplete or inaccurate}; 0: \text{no alignment}\}$
+
+For objective criteria, the scoring is binary:  $S_{i} = \{3 : \text{adheres to scale and structure}; 0 : \text{otherwise}\}$
+
+The overall adherence score is then computed as  $S_{\text{adherence}} = \frac{\sum_{i=1}^{2} S_i}{6}$ .
+
+This score quantifies how well the review adheres to the prescribed guidelines.
+
+# 4 ReviewAgent
+
+To ensure AI-based peer reviews are thorough across the multi-dimensional metrics defined in ReviewEval, we introduce ReviewAgent, a framework that aligns evaluations with conference-specific reviewing guidelines while operating effectively across those dimensions.
+
+# 4.1 Conference-Specific Review Alignment
+
+To tailor the review process to conference-specific guidelines, we first retrieve them from the target conference's official reviewing website and preprocess them using LLM. Each review guideline  $g_{i}$  is then converted into a step-by-step instructional prompt via LLM:  $P_{i} = \text{GeneratePrompt}(g_{i})$ , where  $P_{i}$  denotes the prompt corresponding to guideline  $g_{i}$ . Since some criteria apply to multiple sections of a research paper, the prompts are dynamically mapped to the relevant sections. This is achieved using a mapping function:  $S_{j} = \mathcal{M}(P_{i})$ , where  $S_{j}$  is the set of paper sections associated with prompt  $P_{i}$ . Notably,  $\mathcal{M}$  is a one-to-many mapping, i.e.,  $\mathcal{M}: P \to \mathcal{P}(S)$ , with  $\mathcal{P}(S)$  denoting the power set of all sections. By conducting reviews on a section-wise basis, our framework enhances processing efficiency and allows for independent evaluation of each section prior to aggregating the final review. (Prompts provided in Appendix A.5)
+
+# 4.2 Iterative Refinement Loop
+
+To enhance the quality and completeness of our review prompts and the section-wise reviews generated, we employ an iterative refinement process using a Supervisor LLM. Starting with an initial set of prompts generated from the reviewing guidelines, the Supervisor LLM evaluates each prompt in conjunction with its corresponding guideline, serving as the problem statement, and provides targeted feedback. This feedback addresses key aspects such as clarity logical consistency align-
+
+ment with the guideline, and comprehensiveness. The feedback is then used to revise the prompt, and this iterative loop is repeated for a fixed number of iterations (one in our current implementation). The result is a final set of structured, high-quality review prompts that are well-aligned with the reviewing guidelines and optimized for the evaluation process. We employ the same mechanism to iteratively refine the section-wise reviews generated, ensuring each section benefits from targeted feedback and improvement.
+
+# 4.3 Improvement using ReviewEval
+
+To further improve the quality and consistency of our reviews, we use an automated improvement loop based on an ReviewEval. After a draft for review is generated for a paper, ReviewEval evaluates it across four important dimensions: (1) constructiveness,(2) depth of analysis, (3) factual accuracy, and (4) adherence to reviewer guidelines a score along with scores of each criteria in these metrics. These are then passed on to the agent, which consumes both the raw review data and the evaluation feedback to generate an enhanced, final review. In our current implementation this loop is run once. Notably, throughout system development and benchmarking we do conduct an AI vs Human analysis in ReviewEval to judge AI generated reviews against expert reviews taking them as gold standard. But for real world submissions where we do not have any existing reviewer comments, we solely depend on these four objective criteria.
+
+# 5 Experiments & Results
+
+Dataset. We constructed a dataset, FullCorpus-120, comprising 120 papers sampled from the NeurIPS, ICLR, and UAI conferences. For each of the acceptance and rejection category we sampled 8 papers, resulting in a balanced set across acceptance categories (mentioned in table 4). These conferences were selected due to the availability of official reviewer comments for their 2024 and later editions on Openreview.net. Notably, the recent nature of these papers creates a challenging test set, as SOTA LLMs are unlikely to be trained on this data given their data cutoffs dates. This design allows for a robust assessment of each model's ability to interpret and analyze previously unseen data. To promote topical diversity, we select papers that integrate artificial intelligence with other domains such as biology, social science, finance, and engineering  $(\mathrm{AI} + \mathrm{X})$ , thereby avoiding over
+
+<table><tr><td>Framework</td><td>Actionable insights</td><td>Adherence to review guidelines</td><td>Coverage of topics</td><td>Semantic similarity</td><td>Depth of analysis</td><td>Factual correctness</td></tr><tr><td>Expert</td><td>0.4457 ± 0.1322</td><td>0.5967 ± 0.1369</td><td>N/A ± N/A</td><td>N/A ± N/A</td><td>0.8772 ± 0.1127</td><td>0.9797 ± 0.0400</td></tr><tr><td>MARG-V1-Deepseek</td><td>0.4915 ± 0.1193</td><td>0.4367 ± 0.1441</td><td>0.5315 ± 0.2139</td><td>0.7957 ± 0.0710</td><td>0.8022 ± 0.1231</td><td>0.9873 ± 0.0283</td></tr><tr><td>MARG-V1-Qwen</td><td>0.6161 ± 0.1098</td><td>0.5160 ± 0.1411</td><td>0.5123 ± 0.2057</td><td>0.7809 ± 0.0818</td><td>0.9098 ± 0.0929</td><td>0.9777 ± 0.0388</td></tr><tr><td>MARG-V2-Deepseek</td><td>0.3702 ± 0.1603</td><td>0.5898 ± 0.1376</td><td>0.8694 ± 0.1625</td><td>0.7638 ± 0.0918</td><td>0.9122 ± 0.0629</td><td>0.9911 ± 0.0305</td></tr><tr><td>MARG-V2-Qwen</td><td>0.4874 ± 0.1441</td><td>0.6337 ± 0.1418</td><td>0.8341 ± 0.1856</td><td>0.7814 ± 0.0809</td><td>0.9429 ± 0.0544</td><td>0.9762 ± 0.0565</td></tr><tr><td>ReviewAgent-Deepseek</td><td>0.3096 ± 0.1438</td><td>0.7584 ± 0.1266</td><td>0.7330 ± 0.1709</td><td>0.8016 ± 0.0990</td><td>0.8694 ± 0.1239</td><td>0.9870 ± 0.0556</td></tr><tr><td>ReviewAgent-Qwen</td><td>0.5899 ± 0.1278</td><td>0.8545 ± 0.1039</td><td>0.7351 ± 0.1803</td><td>0.7887 ± 0.1089</td><td>0.9733 ± 0.0514</td><td>0.9793 ± 0.0436</td></tr><tr><td>AI-Scientist-Deepseek</td><td>0.3164 ± 0.1611</td><td>0.5997 ± 0.1077</td><td>0.8913 ± 0.1450</td><td>0.7876 ± 0.1012</td><td>0.7045 ± 0.1320</td><td>0.9743 ± 0.0966</td></tr><tr><td>AI-Scientist-Qwen</td><td>0.4684 ± 0.1641</td><td>0.7105 ± 0.1508</td><td>0.7933 ± 0.1696</td><td>0.8088 ± 0.0988</td><td>0.9010 ± 0.0965</td><td>0.9901 ± 0.0330</td></tr></table>
+
+Table 2: Evaluation of AI-generated reviews across six different metrics on FullCorpus-120 dataset. ReviewAgent is compared against MARG (D'Arcy et al., 2024) and Sakana AI Scientist (Lu et al., 2024). The 'NA' for Expert indicates that the metric is calculated relative to the Expert rating, and therefore does not require a comparison  
+
+<table><tr><td>Framework</td><td>Actionable insights</td><td>Adherence to review guidelines</td><td>Coverage of topics</td><td>Semantic similarity</td><td>Depth of analysis</td><td>Factual correctness</td></tr><tr><td>MARG-V1-3.5-Haiku</td><td>0.4262 ± 0.1287</td><td>0.4587 ± 0.1004</td><td>0.5096 ± 0.1985</td><td>0.4330 ± 0.0684</td><td>0.4689 ± 0.1551</td><td>0.9650 ± 0.0711</td></tr><tr><td>MARG-V1-3.7-Sonnet</td><td>0.4725 ± 0.1017</td><td>0.5820 ± 0.0854</td><td>0.4355 ± 0.2203</td><td>0.5057 ± 0.0782</td><td>0.8244 ± 0.0884</td><td>0.9593 ± 0.0673</td></tr><tr><td>MARG-V1-GPT4o</td><td>0.3762 ± 0.1116</td><td>0.5567 ± 0.1159</td><td>0.4870 ± 0.1906</td><td>0.5019 ± 0.0705</td><td>0.7178 ± 0.1283</td><td>0.9587 ± 0.1089</td></tr><tr><td>MARG-V1-GPT4o-mini</td><td>0.4338 ± 0.1139</td><td>0.5433 ± 0.0807</td><td>0.5034 ± 0.1684</td><td>0.4974 ± 0.0700</td><td>0.7867 ± 0.0882</td><td>0.9877 ± 0.0344</td></tr><tr><td>MARG-V2-3.5-Haiku</td><td>0.2746 ± 0.1394</td><td>0.4400 ± 0.1400</td><td>0.9467 ± 0.1383</td><td>0.7025 ± 0.1100</td><td>0.7311 ± 0.0934</td><td>0.9713 ± 0.0801</td></tr><tr><td>MARG-V2-3.7-Sonnet</td><td>0.3384 ± 0.1451</td><td>0.5047 ± 0.1573</td><td>0.8817 ± 0.1842</td><td>0.7639 ± 0.0476</td><td>0.8044 ± 0.1333</td><td>0.9903 ± 0.0300</td></tr><tr><td>MARG-V2-GPT4o</td><td>0.3077 ± 0.1400</td><td>0.4957 ± 0.1345</td><td>0.8896 ± 0.1755</td><td>0.7763 ± 0.0714</td><td>0.8156 ± 0.0970</td><td>0.9900 ± 0.0403</td></tr><tr><td>MARG-V2-GPT4o-mini</td><td>0.2137 ± 0.1523</td><td>0.5167 ± 0.1397</td><td>0.9361 ± 0.1241</td><td>0.7701 ± 0.0601</td><td>0.8644 ± 0.0643</td><td>0.9887 ± 0.0431</td></tr><tr><td>ReviewAgent-3.5-haiku</td><td>0.1781 ± 0.1096</td><td>0.6867 ± 0.1161</td><td>0.6868 ± 0.1825</td><td>0.8380 ± 0.0310</td><td>0.6911 ± 0.1292</td><td>0.9960 ± 0.0219</td></tr><tr><td>ReviewAgent-3.7-sonnet</td><td>0.4135 ± 0.1628</td><td>0.7213 ± 0.1048</td><td>0.7915 ± 0.2026</td><td>0.8202 ± 0.0751</td><td>0.9222 ± 0.0823</td><td>0.9837 ± 0.0633</td></tr><tr><td>ReviewAgent-Deepseek-Imp</td><td>0.2830 ± 0.1542</td><td>0.7313 ± 0.1205</td><td>0.7205 ± 0.1909</td><td>0.7778 ± 0.1138</td><td>0.9356 ± 0.0866</td><td>0.9890 ± 0.0474</td></tr><tr><td>ReviewAgent-Deepseek-ImpRef</td><td>0.2927 ± 0.1253</td><td>0.7330 ± 0.1213</td><td>0.7369 ± 0.1863</td><td>0.8021 ± 0.0676</td><td>0.9533 ± 0.0610</td><td>0.9943 ± 0.0310</td></tr><tr><td>ReviewAgent-Deepseek-Ref</td><td>0.2883 ± 0.1192</td><td>0.7587 ± 0.1443</td><td>0.7377 ± 0.1573</td><td>0.7767 ± 0.1057</td><td>0.9000 ± 0.0971</td><td>0.9833 ± 0.0582</td></tr><tr><td>ReviewAgent-GPT4o</td><td>0.2740 ± 0.1592</td><td>0.6773 ± 0.1228</td><td>0.7565 ± 0.1693</td><td>0.8018 ± 0.0954</td><td>0.6889 ± 0.1360</td><td>0.9833 ± 0.0913</td></tr><tr><td>ReviewAgent-GPT4o-mini</td><td>0.2029 ± 0.1391</td><td>0.6287 ± 0.1235</td><td>0.7421 ± 0.1600</td><td>0.8092 ± 0.0906</td><td>0.6822 ± 0.0954</td><td>0.9850 ± 0.0634</td></tr><tr><td>ReviewAgent-Qwen-Imp</td><td>0.6469 ± 0.1415</td><td>0.8787 ± 0.0990</td><td>0.6649 ± 0.1727</td><td>0.8401 ± 0.0347</td><td>0.9822 ± 0.0300</td><td>0.9707 ± 0.0474</td></tr><tr><td>ReviewAgent-Qwen-ImpRef</td><td>0.6579 ± 0.1194</td><td>0.8493 ± 0.1242</td><td>0.7418 ± 0.1848</td><td>0.7965 ± 0.0557</td><td>0.9889 ± 0.0253</td><td>0.9770 ± 0.0341</td></tr><tr><td>ReviewAgent-Qwen-Ref</td><td>0.6135 ± 0.1394</td><td>0.8347 ± 0.1058</td><td>0.6605 ± 0.1861</td><td>0.8185 ± 0.0952</td><td>0.9778 ± 0.0535</td><td>0.9590 ± 0.0471</td></tr><tr><td>AI-Scientist-3.5-Haiku</td><td>0.3257 ± 0.1448</td><td>0.5360 ± 0.0941</td><td>0.8624 ± 0.1441</td><td>0.7786 ± 0.0696</td><td>0.6133 ± 0.1071</td><td>0.9953 ± 0.0256</td></tr><tr><td>AI-Scientist-3.7-Sonnet</td><td>0.3215 ± 0.1439</td><td>0.7980 ± 0.1071</td><td>0.8208 ± 0.1694</td><td>0.7986 ± 0.0978</td><td>0.9511 ± 0.0741</td><td>0.9667 ± 0.1269</td></tr><tr><td>AI-Scientist-GPT4o</td><td>0.3071 ± 0.1471</td><td>0.5217 ± 0.1100</td><td>0.9079 ± 0.1282</td><td>0.7661 ± 0.0818</td><td>0.5911 ± 0.1144</td><td>0.9640 ± 0.1148</td></tr><tr><td>AI-Scientist-GPT4o-mini</td><td>0.3406 ± 0.1390</td><td>0.5107 ± 0.0957</td><td>0.8279 ± 0.1776</td><td>0.7947 ± 0.0417</td><td>0.6111 ± 0.0944</td><td>0.9557 ± 0.1903</td></tr></table>
+
+Table 3: Evaluation of AI-generated reviews across six different metrics on CoreCorpus-30 dataset. ReviewAgent against MARG and Sakana AI Scientist. ReviewEval framework has two binary settings: Imp (improvement described in 4.3) and Ref (reflection described in 4.2). A suffix of -Imp, -Ref or -ImpRef indicates that the corresponding feature was turned "on"; absence of any suffix implies both features were "off."  
+
+<table><tr><td>Venue (Year)</td><td>Poster</td><td>Spotlight</td><td>Oral</td><td>Rejected</td><td>Total</td></tr><tr><td>UAI (2024)</td><td>8</td><td>8</td><td>8</td><td>-</td><td>24</td></tr><tr><td>NeurIPS (2024)</td><td>8</td><td>8</td><td>8</td><td>8</td><td>32</td></tr><tr><td>ICLR (2025)</td><td>8</td><td>8</td><td>8</td><td>8</td><td>32</td></tr><tr><td>ICLR (2024)</td><td>8</td><td>8</td><td>8</td><td>8</td><td>32</td></tr><tr><td>Overall</td><td>-</td><td>-</td><td>-</td><td>-</td><td>120</td></tr></table>
+
+Table 4: Composition of our FullCorpus-120 test set.  
+
+<table><tr><td>Domain</td><td>Papers</td></tr><tr><td>Healthcare/Biology</td><td>11</td></tr><tr><td>Social Sciences/Humanities</td><td>4</td></tr><tr><td>Economics/Finance</td><td>14</td></tr><tr><td>Physics/Astronomy</td><td>6</td></tr><tr><td>Engineering/Systems</td><td>4</td></tr></table>
+
+Table 5: Cross-domain paper count in FullCorpus-120.
+
+representation of any single domain (mentioned in table 5). We use FullCorpus-120 to establish results for our own reviewer models as well as all baseline methods.
+
+To evaluate on expensive foundation models and run higher-cost framework settings, we curated a stratified subset containing 30 papers CoreCorpus-30 that preserves the diversity and balance of the FullCorpus-120 while enabling more efficient experimentation.
+
+Baselines and models. We compare our AI-generated review approach with two established methods: Sakana AI Scientist (Lu et al., 2024) and MARG (v1 and v2) (D'Arcy et al., 2024). To demonstrate the effectiveness of ReviewAgent, we report results using foundational models from Deepseek, Qwen, OpenAI, and Anthropic.
+
+Results for FullCorpus-120. Table 2 summarizes the performance of ReviewAgent along with baselines and expert reviews on FullCorpus-120. For each of the six metrics, we highlight the top-scoring system and contrast it with ReviewAgent:
+
+Actionable insights. MARG-v1-qwen narrowly beats ReviewAgent-qwen. The fact that ReviewAgent-qwen nearly matches MARG's top score demonstrates the efficacy of ReviewAgent in producing concrete, actionable feedback.
+
+Adherence to the review guidelines. Experts average approximately 0.6. ReviewAgent-Qwen (mean  $\approx 0.85$ ) and ReviewAgent-DeepSeek (mean  $\approx 0.75$ ) show strong compliance, surpassing both MARG and AI-Scientist, which are hardcoded to a particular review format. In contrast, ReviewAgent dynamically adjusts itself to the given guidelines. Topic coverage and semantic similarity. AI-Scientist-Deepseek leads the topic coverage, while ReviewEval also shows promising results. For semantic similarity, AI-Scientist-qwen tops while ReviewAgent-deepseek closely follows.
+
+Depth of analysis. ReviewAgent-qwen achieves the highest depth of analysis score, surpassing both expert reviewers and all baseline models. This demonstrates ReviewAgent's ability to produce in-depth rather than superficial evaluations.
+
+Factual correctness. MARG-v2-deepseek scores the highest, ReviewAgent-deepseek and ReviewAgent-qwen remain just below, indicating reliability and low risks of misinterpretations by AI-based frameworks.
+
+Overall, ReviewAgent attains the highest depth, adherence, and near-top actionable insights, demonstrating its superiority over AI baselines and expert reviews on FullCorpus-120.
+
+Results for CoreCorpus-30. Building on this foundation, we evaluated ReviewAgent under more resource intensive settings and also used proprietary, closed-source models on CoreCorpus-30, a stratified subset of FullCorpus-120. Table 3 summarises these results:
+
+# Actionable insights.
+
+ReviewAgent-Qwen-Improvement-Reflection leads, closely followed by the Improvement variant. Both outperform all MARG and AI-Scientist baselines, underscoring how feedback from ReviewEval and iterative self-reflection loop boosts the constructiveness of feedback.
+
+Adherence to the review guidelines. ReviewAgent-Qwen-Improvement achieves the highest adherence, with the Improvement vari
+
+ant a close second followed by Reflection variant. These results exceed all baseline models, including MARG and AI-Scientist, and highlight the effectiveness of dynamic prompt generation and evaluation-driven calibration, coupled with iterative refinements.
+
+Topic coverage and Semantic Similarity. While the top single-model coverage is achieved by a MARG variant, ReviewAgent-3.7-sonnet still maintains a strong, balanced scope even if not maximized for breadth. ReviewAgent-Qwen-Improvement achieves the highest alignment with expert phrasing. This indicates that ReviewAgent can mimic the tone and delivery of expert comments.
+
+# Depth of analysis.
+
+ReviewAgent-Qwen-Improvement-Reflection achieves the top score, with the Improvement variant next. Both significantly surpass MARG and AI-Scientist models, validating that iterative reflection enables truly in-depth, expert-level critique rather than superficial reviews.
+
+Factual correctness. ReviewAgent variants achieve among the highest factual correctness scores, with DeepSeek and Haiku-based versions performing best. Other baselines also score high in factual accuracy, showing that the models generally stay close to the source material and avoid making things up. Overall, the pronounced gains from the Improvement and Reflection loops demonstrate their effectiveness in elevating review quality.
+
+# 6 Conclusion
+
+We introduced ReviewEval, a framework evaluating AI-generated reviews on alignment, accuracy, depth, constructiveness, and guideline adherence. Additionally, we propose ReviewAgent, an AI system featuring (1) conference-specific review alignment, (2) iterative self-refinement, and (3) external improvement via ReviewEval. Our experiments show ReviewAgent matches expert quality and surpasses existing AI frameworks.
+
+# 7 Acknowledgement
+
+The authors wish to acknowledge the use of ChatGPT in improving the presentation and grammar of the paper. The paper remains an accurate representation of the authors' underlying contributions. This work was supported in part by Initiation Research Grants from IIIT Delhi and the New Faculty Seed Grant from BITS Pilani.
+
+# 8 Limitations
+
+(1) Although we attempted to enhance cross-domain diversity by considering AI + X papers, our evaluation dataset comprised papers only from the AI conferences, which may confine the generalizability of our results to other domains, disciplines, or larger-scale datasets. (2) Dependence on LLMs brings with it familiar weaknesses of LLM-as-a-Judge paradigm: hallucinations, bias, prompt sensitivity, temporal stability and black-box behavior, limiting interpretability, reliability, transparency, and trust. We tried to overcome this by decomposing the evaluation process itself into a well-defined, multi-step procedure and doing a comprehensive evaluation of our approach across a wide variety of LLMs. (3) Metric interdependence (semantic similarity, factual accuracy, depth of analysis) injects subjectivity and can constrain model flexibility. (4) Iterative refinement loop, external improvement loop and RAG components in our approach have nontrivial computational expense. However, this is shared by many Generative AI tools and products such as Deep Research Agents. Furthermore, we could also incorporate cost-reduction mechanisms like caching, batch processing, and chunk optimisation to enhance efficiency. (5) Lastly, purely automated metrics might not be able to pick up on qualitative aspects like tone, subtle criticism, and readability.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Zafar Ali, Pavlos Kefalas, Khan Muhammad, Bahadar Ali, and Muhammad Imran. 2020. Deep learning in citation recommendation models survey. Expert Systems with Applications, 162:113790.  
+Setio Basuki and Masatoshi Tsuchiya. 2022. The quality assist: A technology-assisted peer review based on citation functions to predict the paper quality. IEEE Access, 10:126815-126831.  
+Howard Bauchner and Frederick P Rivara. 2024. Use of artificial intelligence and the future of peer review. Health Affairs Scholar, 2(5):qxae058.  
+Prabhat Kumar Bharti, Tirthankar Ghosal, Mayank Agarwal, and Asif Ekbal. PEERRec: An AI-based approach to automatically generate recommendations and predict decisions in peer review. 25(1):55-72.
+
+Som S Biswas. 2024. Ai-assisted academia: Navigating the nuances of peer review with chatgpt 4. The Journal of Pediatric Pharmacology and Therapeutics, 29(4):441-445.  
+Sungmin Cha, Kyunghyun Cho, and Taesup Moon. Augmenting negative representation for continual self-supervised learning.  
+Alessio Checco, Lorenzo Bracciale, Paola Loretti, Stephen Pinfield, and Giuseppe Bianchi. 2021. AI-assisted peer review. *Humanities and Social Sciences Communications*, 8(1):1-11.  
+M D'Arcy, T Hope, L Birnbaum, and D Downey. 2024. Marg: Multi-agent review generation for scientific papers (arxiv: 2401.04259). arxiv.  
+Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Yuanzhuo Wang, Wen Gao, Lionel Ni, and Jian Guo. 2025. A survey on llm-as-a-judge. Preprint, arXiv:2411.15594.  
+T. Hayes, R. Rao, H. Akin, and 1 others. 2024. Simulating 500 million years of evolution with a language model. bioRxiv.  
+X. He, K. Zhao, and X. Chu. 2021. Automl: A survey of the state-of-the-art. Knowledge-Based Systems, 212:106622.  
+Chi Hu, Yimin Hu, Hang Cao, Tong Xiao, and Jingbo Zhu. 2024. Teaching language models to self-improve by learning from language feedback. arXiv preprint arXiv:2406.07168.  
+F. Hutter, L. Kotthoff, and J. Vanschoren. 2019. Automated Machine Learning: Methods, Systems, Challenges. Springer Nature.  
+Dongyeop Kang, Waleed Ammar, Bhavana Dalvi, Madeleine Van Zuylen, Sebastian Kohlmeier, Edward Hovy, and Roy Schwartz. 2018. A dataset of peer reviews (peerread): Collection, insights and nlp applications. arXiv preprint arXiv:1804.09635.  
+Giovanni R Latona, Manoel H Ribeiro, Thomas R Davidson, Vlad Veselovsky, and Robert West. 2024. The AI Review Lottery: Widespread AI-Assisted Peer Reviews Boost Paper Scores and Acceptance Rates. arXiv preprint arXiv:2405.02150.  
+Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Smith, Yian Yin, Daniel McFarland, and James Zou. Can large language models provide useful feedback on research papers? a large-scale empirical analysis. Preprint, arxiv:2310.01783 [cs].  
+Jialiang Lin, Jiaxin Song, Zhangping Zhou, Yidong Chen, and Xiaodong Shi. 2023a. Automated scholarly paper review: Concepts, technologies, and challenges. Information Fusion, 98:101830.
+
+Jialiang Lin, Jiaxin Song, Zhangping Zhou, Yidong Chen, and Xiaodong Shi. 2023b. Moprd: A multidisciplinary open peer review dataset. Neural Computing and Applications, 35(34):24191-24206.  
+Ryan Liu and Nihar B Shah. 2023. Reviewergpt? an exploratory study on using large language models for paper reviewing. arXiv preprint arXiv:2306.00622.  
+Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. 2024. The ai scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292.  
+Aman Madaan, Jason Tuck, Jayesh Gupta, Amir Yazdanbakhsh, Yash Zheng, Vivek Srikumar, Deepak Pathak, Denny Yang, Partha Talukdar, and Karan Goel. 2023. Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651.  
+A. Merchant, E. Pyzer-Knapp, P. Szymanski, and 1 others. 2023. Ai in materials discovery. Nature Materials.  
+Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. Preprint, arXiv:1301.3781.  
+Michele B. Nuijten and Joshua R. Polanin. "statacheck": Automatically detect statistical reporting inconsistencies to increase reproducibility of meta-analyses. 11(5):574-579. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/jrsm.1  
+Maria Petrescu and Anjala S Krishen. 2022. The evolving crisis of the peer-review process. Journal of Marketing Analytics, 10(3):185-186.  
+Zachary Robertson. 2023. Gpt4 is slightly helpful for peer-review assistance: A pilot study. arXiv preprint arXiv:2307.05492.  
+Richard Schulz, Adrian Barnett, René Bernard, Nicholas J Brown, JA Byrne, Peter Eckmann, Marianna A Gazda, Halil Kilicoglu, Elisabeth M Prager, Michael Salholz-Hillel, and 1 others. 2022. Is the future of peer review automated? BMC Research Notes, 15(1):1-5.  
+Wenda Xu, Daniel Deutsch, Mara Finkelstein, Juraj Juraska, Biao Zhang, Zhongtao Liu, William Yang Wang, Lei Li, and Markus Freitag. 2024. Llmrefine: Pinpointing and refining large language models via fine-grained actionable feedback. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 1429-1445.  
+Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations.  
+Ruiyang Zhou, Lu Chen, and Kai Yu. 2024. Is llm a reliable reviewer? a comprehensive evaluation of llm on automatic paper reviewing tasks. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 9340-9351.
+
+Zhenzhen Zhuang, Jiandong Chen, Hongfeng Xu, Yuwen Jiang, and Jialiang Lin. 2025. Large language models for automated scholarly paper review: A survey. Information Fusion, 124:103332.
+
+# A Appendix
+
+# A.1 Correlation Tests of parameters in review eval
+
+# A.1.1 Averaging Analysis: Metric Impact and Contribution
+
+To understand the individual importance and contribution of each metric to an overall assessment, an averaging analysis was performed. This analysis aimed to quantify how sensitive a composite score is to the removal of any single metric and to determine the proportional contribution of each metric to this composite score.
+
+Methodology. A unified score was first calculated for each data point (representing a specific model's evaluation of a paper) by taking the simple arithmetic mean of the six core metrics: depth score, actionable insights, adherence score, coverage, semantic similarity, and factual correctness.
+
+For each model and for each of the six metrics, the analysis involved the following steps:
+
+1. The baseline unified score was established using all six metrics.  
+2. Each metric was temporarily removed, and a new adjusted average score was calculated based on the remaining five metrics.  
+3. The absolute change and relative percentage change between the baseline unified score and the adjusted score were computed. A positive change indicates that removing the metric increased the overall score (implying the metric typically scored lower than others), while a negative change indicates that removing the metric decreased the overall score (implying the metric typically scored higher).  
+4. The contribution of each individual metric to the original unified score was calculated as its value divided by the number of metrics (6). The average percentage contribution of each metric to the baseline unified score across all evaluated papers for a model was also determined.
+
+These calculations were performed for every model in the dataset, and the results were then aggregated
+
+to obtain mean and standard deviation values for the changes and contributions across all models.
+
+Results. The aggregated results, presented in Table 6, highlight the overall impact and contribution of each metric. The table shows the mean absolute change in the unified score when a metric is removed, the mean relative percentage change, and the mean percentage contribution of each metric to the unified score.
+
+As observed in Table 6, removing actionable insights led to the largest positive mean change in the unified score (0.0654, or  $+9.00\%$ ), suggesting it generally scored lower than the average of other metrics. Conversely, removing factual correctness resulted in the largest negative mean change (-0.0492, or  $-6.96\%$ ), indicating its scores were typically higher and thus its removal decreased the overall score most significantly.
+
+In terms of contribution, factual correctness had the highest average contribution to the unified score  $(22.47\%)$ , followed by depth score  $(19.14\%)$  and semantic similarity  $(17.34\%)$ . Actionable insights had the lowest average contribution  $(9.17\%)$ .
+
+These findings are further visualized in the generated plots. Bar chart 4 illustrates the mean absolute change in the unified score upon removing each metric, with bars colored to distinguish between positive and negative impacts. Pie chart 3 depicts the average percentage contribution of each metric to the total unified score, providing a clear visual breakdown of metric influence.
+
+![](images/9c8c058d58133b73777372acaa84da194e331c76b051146cf36e571f11b150e0.jpg)  
+Average Contribution of Each Metric to Unified Score  
+Figure 3: Metric Contribution
+
+![](images/7c5ce31a22ccb48a9931a7cedb37010e16b846f9dbbb744617765402eb2e061e.jpg)  
+Figure 4: Metric Importance
+
+# A.1.2 Pearson Correlation Analysis
+
+To investigate the linear relationships between the different evaluation metrics, a Pearson correlation analysis was conducted. This method assesses the strength and direction of association between pairs of continuous variables. The metrics analyzed were depth score, actionable insights, adherence score, coverage, semantic similarity, and factual correctness.
+
+Methodology The Pearson correlation coefficient  $(r)$  was calculated for every pair of the six metrics using all available individual data points. The correlation coefficient ranges from -1 (perfect negative linear correlation) to  $+1$  (perfect positive linear correlation), with 0 indicating no linear correlation. For each correlation, a p-value was also computed to determine the statistical significance of the observed relationship.
+
+Results The computed Pearson correlation matrix is presented in Table 7, and the corresponding p-values are shown in Table 8.
+
+The strongest positive correlation was observed between Adherence Score and Depth Score ( $r \approx 0.426$ ,  $p < 0.001$ ), suggesting that evaluations with higher adherence to instructions also tended to have greater depth. Another notable positive correlation was found between Depth Score and Actionable Insights ( $r \approx 0.298$ ,  $p < 0.001$ ). Coverage showed a statistically significant negative correlation with Actionable Insights, indicating that broader coverage might sometimes come at the expense of providing actionable insights, or vice-versa. Most other correlations were relatively weak, although several were statistically significant due to the large sample size. For instance, the correlation between Semantic Similarity and Factual Correctness was very weak and borderline significant. Factual Correctness showed very weak and non-significant correlations with the other 5
+
+<table><tr><td>Metric Removed</td><td>Mean Abs. Change</td><td>Mean Rel. Change (%)</td><td>Mean Contr. (%)</td></tr><tr><td>actionable_investig</td><td>0.0654</td><td>9.00</td><td>9.17</td></tr><tr><td>factual.correctness</td><td>-0.0492</td><td>-6.96</td><td>22.47</td></tr><tr><td>depth_score</td><td>-0.0225</td><td>-2.97</td><td>19.14</td></tr><tr><td>adherence_score</td><td>0.0131</td><td>1.91</td><td>15.08</td></tr><tr><td>semantic_similarity</td><td>-0.0061</td><td>-0.81</td><td>17.34</td></tr><tr><td>coverage</td><td>-0.0008</td><td>-0.16</td><td>16.80</td></tr></table>
+
+Table 6: Metric Importance Summary: Impact of Removal and Contribution to Unified Score (Averaged Across Models)
+
+parameters. These relationships are visually summarized in a heatmap 5 and a series of scatter plots 6.
+
+![](images/b2021af3eeaa8e06547c15705da4394c986825d92db4534ca18e410633ced74c.jpg)  
+Figure 5: Correlation Heatmap
+
+![](images/39ddf63eaf4ed18ccad970396a883e508a9ea804f8f697141707ec3a207581db.jpg)  
+Figure 6: Metric-wise scatter plot
+
+# A.2 Human Evaluation Study
+
+We conducted a blinded evaluation on CoreCorpus-30 to compare AI-generated
+
+peer reviews with expert human reviews.
+
+Experimental Setup A double-blind study was run on 30 papers, each paired with two reviews: one by ReviewAgent-Imp-Ref and one by an expert. Five expert evaluators participated, each assessing 6 paper-review sets, covering all 30 papers. Both reviews were judged using the official reviewer guidelines of the corresponding conference.
+
+Evaluation Criteria & Process Review quality was assessed along four dimensions: (1) Actionable Insights, (2) Adherence to Guidelines, (3) Depth of Analysis, and (4) Overall Quality. For each paper, evaluators compared the two reviews on all dimensions and indicated their preference.
+
+Results AI-generated reviews were preferred in  $75\%$  of comparisons (90 of 120), with consistent preference across all four dimensions (Table 9). These findings suggest AI-generated reviews can match or surpass expert quality, offering improvements in consistency, efficiency, and scalability of peer review. However, we note that this is only a small-scale study using human experts. A large-scale human evaluation study would also be useful to compare the quality of AI-generated reviews and expert reviews.
+
+# A.2.1 Instructions Provided to Evaluators
+
+Instructions Given to Participants Full text shown to evaluators (experts) before they began the task.
+
+Purpose. You are invited to take part in a double-blind study comparing two peer reviews for each paper in a small research corpus. Your judgments will help us assess whether AI-generated reviews provide feedback of comparable quality to expert human reviews.
+
+What you will do. For each assigned paper, you will read two anonymized re
+
+<table><tr><td></td><td>Depth</td><td>Actionable Ins.</td><td>Adherence Sc.</td><td>Coverage</td><td>Semantic Sim.</td><td>Factual Corr.</td></tr><tr><td>Depth Score</td><td>1.000</td><td>0.298</td><td>0.426</td><td>-0.013</td><td>0.242</td><td>0.026</td></tr><tr><td>Actionable Insights</td><td>0.298</td><td>1.000</td><td>0.186</td><td>-0.213</td><td>0.028</td><td>-0.017</td></tr><tr><td>Adherence Score</td><td>0.426</td><td>0.186</td><td>1.000</td><td>0.030</td><td>0.203</td><td>-0.001</td></tr><tr><td>Coverage</td><td>-0.013</td><td>-0.213</td><td>0.030</td><td>1.000</td><td>0.155</td><td>-0.015</td></tr><tr><td>Semantic Similarity</td><td>0.242</td><td>0.028</td><td>0.203</td><td>0.155</td><td>1.000</td><td>0.045</td></tr><tr><td>Factual Correctness</td><td>0.026</td><td>-0.017</td><td>-0.001</td><td>-0.015</td><td>0.045</td><td>1.000</td></tr></table>
+
+Table 7: Pearson Correlation Matrix of Metrics  
+Table 8: P-values for Pearson Correlations  
+
+<table><tr><td></td><td>Depth Sc.</td><td>Actionable Ins.</td><td>Adherence Sc.</td><td>Coverage</td><td>Semantic Sim.</td><td>Factual Corr.</td></tr><tr><td>Depth Score</td><td>0.00e+00</td><td>&lt;.001</td><td>&lt;.001</td><td>0.576</td><td>&lt;.001</td><td>0.274</td></tr><tr><td>Actionable Insights</td><td>&lt;.001</td><td>0.00e+00</td><td>&lt;.001</td><td>&lt;.001</td><td>0.225</td><td>0.479</td></tr><tr><td>Adherence Score</td><td>&lt;.001</td><td>&lt;.001</td><td>0.00e+00</td><td>0.191</td><td>&lt;.001</td><td>0.952</td></tr><tr><td>Coverage</td><td>0.576</td><td>&lt;.001</td><td>0.191</td><td>0.00e+00</td><td>&lt;.001</td><td>0.510</td></tr><tr><td>Semantic Similarity</td><td>&lt;.001</td><td>0.225</td><td>&lt;.001</td><td>&lt;.001</td><td>0.00e+00</td><td>0.055</td></tr><tr><td>Factual Correctness</td><td>0.274</td><td>0.479</td><td>0.952</td><td>0.510</td><td>0.055</td><td>0.00e+00</td></tr></table>
+
+views labeled Review A and Review B. Using the conference's official reviewer guidelines (provided to you), you will indicate which review you prefer on each of the four dimensions below. You may choose either A or B.
+
+Dimensions. (1) Actionable Insights: Does the review offer clear, concrete, and constructive suggestions the authors could act on? (2) Adherence to Guidelines: Does the review follow the conference's review rubric (coverage of novelty, soundness, clarity, significance, ethics, limitations, reproducibility, etc.)? (3) Depth of Analysis: Does the review demonstrate careful reading, accurate understanding of the work, appropriate references to specific sections/claims, and non-trivial reasoning? (4) Overall Quality: Your holistic preference considering balance, fairness, tone, specificity, and usefulness to authors and ACs.
+
+Decision rule. Make an independent choice per dimension. Do not attempt to infer which review is AI-generated or human-written.
+
+Blinding and conduct. - Treat all manuscripts and reviews as confidential. Do not share, copy, or discuss materials outside this task. - Do not contact the authors or search for external information about the papers. - Do not use generative
+
+AI tools to assist your judgment; we are measuring your expert preference.
+
+Risks and benefits. Risks are minimal (reading and evaluating academic text). You may stop at any time without penalty. Your responses will be recorded and analyzed in aggregate for research purposes.
+
+Confidentiality. We are just collecting your decisions. We do not request personally identifying information beyond your role (e.g., senior student, recent graduate, faculty) for reporting annotator characteristics.
+
+Voluntary participation. Participation is voluntary. You may withdraw at any point before final submission of your responses; upon request, we will delete your data if it can be uniquely located.
+
+Questions. For any concerns during the study, please contact the corresponding author.
+
+Recruitment and Payment Participants were recruited directly by the authors via personal outreach to experts (senior students, recent graduates, and faculty). Five experts volunteered to participate. No monetary compensation was offered or provided. All participants were academic experts based in India.
+
+Data Consent Consent text shown before the task began.
+
+<table><tr><td>Dimension</td><td>AI Wins</td><td>Expert Wins</td><td>AI Win %</td><td>Expert Win %</td></tr><tr><td>Actionable Insights</td><td>24/30</td><td>6/30</td><td>80.0</td><td>20.0</td></tr><tr><td>Review Guidelines</td><td>26/30</td><td>4/30</td><td>86.7</td><td>13.3</td></tr><tr><td>Depth of Analysis</td><td>19/30</td><td>11/30</td><td>63.3</td><td>36.7</td></tr><tr><td>Overall Quality</td><td>21/30</td><td>9/24</td><td>70</td><td>30</td></tr><tr><td>Total</td><td>90/120</td><td>30/120</td><td>75</td><td>25</td></tr></table>
+
+Table 9: Comparative Performance of AI vs Expert Reviews Across Evaluation Dimensions
+
+By proceeding, you acknowledge that: (i) you will evaluate pairs of anonymized reviews for anonymized papers; (ii) your responses (preferences per dimension) will be stored securely by the research team and used for research, analysis, and publication (e.g., aggregated statistics); (iii) no personally identifying information is required, and any optional demographic information you provide will be reported only in aggregate; (iv) all study materials (papers and reviews) are confidential and may not be shared; (v) you will not use external assistance (including generative AI) to make your judgments; (vi) your participation is voluntary and you may withdraw at any time before final submission; upon request, and where technically feasible, we will delete your data; (vii) there are minimal risks associated with reading and evaluating academic text. By clicking "I agree and begin," you provide informed consent for the described use of your data.
+
+# A.3 Examples of question generation and decomposition in factual correctness pipeline
+
+Example of question generation from review point in step 1: Consider the following review from the PeerRead dataset (Kang et al., 2018) for the paper "Augmenting Negative Representations for Continual Self-Supervised Learning" (Cha et al.):
+
+Review  $(R)$ : "Augmentation represents a crucial area of exploration in self-supervised learning. Given that the authors classify their method as a form of augmentation, it becomes essential to engage in comparisons and discussions with existing augmentation methods."
+
+Generated question  $(Q)$  : Has the paper engaged in comparisons and discussions with existing augmentation methods, given that the authors classify their method as a form of augmentation?"
+
+Example of question decomposition in step 2: Implementation Detail: We used Langchain's built-in query decomposition tool for initial question breakdown. Langchain's query decomposition module leverages large language models (LLMs) to break down complex, multi-part queries into simpler, sequential sub-questions. It typically works by first identifying logical or temporal dependencies within the input question, and then restructuring it into smaller, more manageable components. These sub-questions can then be answered independently or in a reasoning chain. We answer these questions independently using our custom RAG pipeline.
+
+Main Question: Has the paper engaged in comparisons and discussions with existing augmentation methods, given that the authors classify their method as a form of augmentation?
+
+# Sub-questions:
+
+a) Has the paper compared its augmentation method against existing augmentation methods?  
+b) Does the paper discuss the strengths and weaknesses of related augmentation techniques?
+
+# A.4 Examples of three dimensions of constructiveness
+
+- Specificity:
+
+- Example (score 1): "Section 5.3 discusses pretraining dataset selection but does not address the potential privacy costs of using private data for this purpose. Refer to <Research Paper Citation> for methods to ensure privacy in this step."
+
+- Non-example (score 0): "The paper lacks novelty and is a straightforward application of existing techniques."
+
+- Feasibility:
+
+- Example (score 1): "Break down the GPU hours into pretraining and finetuning stages in Table 7 to make the computational cost more transparent."  
+- Non-example (score 0): "Add experiments with a wide variety of datasets, including proprietary and restricted-access data, to generalize findings."
+
+- Implementation Details:
+
+- Example (score 1): "In Algorithm 1, correct the noise addition formula to  $\frac{1}{B} \cdot \mathcal{N}(0, \sigma^2 C^2 I)$ , as this ensures proper scaling of noise with batch size."  
+- Non-example (score 0): "Use advanced techniques to improve DP-SGD."
+
+# A.5 Prompts Used
+
+Below are the system instructions and prompts employed in our review generation pipeline. These prompts guide each stage of the process—from extracting reviewer guidelines from HTML content, to generating detailed review prompts for specific paper sections, and finally formatting the reviews in strict adherence to conference guidelines.
+
+Guidelines Parsing Prompt. You are a smart AI designed to extract reviewer guidelines from HTML content, regardless of its structure or format. You will be provided with the raw HTML of a webpage that contains the guidelines. Your task is to intelligently parse and extract the most relevant content based on the following high-level objectives:
+
+1. Understand the Context: The HTML file may contain multiple sections of a webpage, including irrelevant information like headers, footers, navigation bars, or ads. Your goal is to focus solely on extracting meaningful content that pertains to reviewer guidelines. Look for terms such as 'reviewer', 'guidelines', 'evaluation', 'criteria', 'instructions', or
+
+'review process' that may indicate sections of interest.
+
+2. Text Structure: Look for relevant sections by identifying common phrases or paragraphs that may contain instructions or rules for reviewers. This includes but is not limited to guidelines on evaluation, reviewing criteria. Focus on only the main content that provides the guidelines for how to review the papers content and not the conference details.  
+3. Avoid Noise: Ignore or discard text that is likely irrelevant, such as menus, links to other pages, copyright information, or promotional content. You are interested only in extracting text that provides guidance to reviewers for evaluating papers.  
+4. Identify Sections Based on Common Words: You can identify the main sections of interest by finding phrases like: "Reviewer Guidelines", "Review Criteria", "Evaluation Process", "Instructions for Reviewers", "Review Process Overview", When you find such phrases, capture the paragraph or section following the phrase, as this is likely to contain the reviewer guidelines.  
+5. Extract Text Around These Keywords: When you identify these keywords, extract approximately 3-4 paragraphs surrounding these keywords to capture the guidelines. This includes headings or bullet points that may be present.  
+6. Return Results as String: Once you have completed parsing the HTML content and extracted relevant guidelines, return the guidelines as a single continuous string. Ensure the text is well-formatted and readable, without HTML tags or irrelevant information like advertisements or links.
+
+7. Avoid capturing details of the conference or event itself, such as times, dates, locations, or registration information. Your task is to focus solely on the reviewer guidelines and evaluation criteria.  
+8. Avoid capturing details of what software or tools to use for the review process. Focus on the guidelines for evaluating the content of the papers.  
+9. If there is a table of any sort in the reviewer guidelines, extract the text content of the table and present it in a readable format, as a paragraph or list of items. Do not include the table structure in the extracted text.  
+10. if there are guidelines for multiple types of papers like ones in CER OR PCI OR NLP, extract the information of the first type of paper only. Do not give any recitations of any sort as that is blocked by google because of copyright issues. Note that you MUST also check the format which is required by the conference guidelines for a review and the output should be given in that format in the end
+
+Prompt for instruction generation for review of a section. You are a generative language model (LLM X) creating a prompt for another research paper reviewer LLM (LLM Y), generate a detailed prompt instructing LLM Y on how to review the section section of a research paper. Consider the following criteria:
+
+1. The clarity and completeness of the section.  
+2. The relevance and alignment of the section with the main themes and objectives of the paper.  
+3. The logical consistency and evidence support in the section.  
+4. The originality and contribution of the section to the field.
+
+5. Any specific elements highlighted in the conference guidelines that should be focused on in the section.
+
+Provide structured and clear instructions in the form of a plan with steps that will enable LLM Y to conduct a thorough and critical review of the research paper's section. Use the given conference guidelines. Do not give any recitations of any sort as that is blocked by google because of copyright issues.
+
+Prompt for finally formatting review as per conference guidelines. You are an expert in writing reviews for various research paper conferences. You will be given reviews for various sections of a research paper, and the research paper itself and you are supposed to write the review in the format that is expected for submission to the specified conference. You're given the contents of the reviewer guidelines for the conference and you are supposed to adhere to it strictly. You are also not supposed to change the content of the review provided to you AT ALL. You are just a formatter and are supposed to just rewrite the given review into the given format while making the necessary changes. You are to give the complete review of the paper in the format of the conference (the entire paper, not some part of it). Remember that you have an output token limit of 8192 tokens and your entire review is supposed to fit within that limit, so be careful. This is the conference guidelines for the conference: str(guidelines)
+
+# A.6 Example Reviews
+
+Below is an example review of a paper which has been accepted by ReviewAgent.
+
+Paper Title: Fast and Unified Path Gradient Estimators for Normalizing Flows (arXiv:2403.15881)
+
+# Summary
+
+This paper proposes a computationally efficient and unified framework for estimating path gradients in normalizing flows (NFs), enabling scalable training for both reverse and forward KL divergences.
+
+The authors address critical inefficiencies in prior path gradient estimation—such as repeated forward-backward passes and costly numerical inversion—via recursive gradient equations and implicit differentiation. Their approach achieves  $1.3 - 8 \times$  runtime improvements and higher effective sample size (ESS). The method unifies coupling flows, continuous flows, and implicitly invertible flows, with applications in physics (e.g., U(1) lattice theory) and high-dimensional ML.
+
+# Strengths
+
+- Clear Problem Statement & Motivation: Identifies the  $4 - 5 \times$  runtime overhead of existing path gradients as a key barrier for large-scale NF training. Links computational bottlenecks (e.g., numerical inversion) to practical challenges in scalable inference for physics and machine learning.
+
+# - Theoretical & Methodological Novelty:
+
+- Derives recursive gradient equations (Corollary 3.4) for affine coupling flows, reducing runtime by avoiding Jacobian inversion.  
+- Reformulates forward KL training as reverse KL in base space (Proposition 4.1), enabling direct application of path gradients with regularization benefits.  
+- Unifies path gradient computation across diverse architectures (coupling, continuous, implicitly invertible).
+
+- Empirical Validation: Demonstrates significant ESS improvements (97.4% for GMMs) and runtime gains (17.5× faster for implicit flows), with comparisons to prior work (e.g., Vaitl et al. 2022a).  
+- Practical Relevance: Highlights applications for physics simulations (e.g.,  $\phi^4$  lattice theory) and high-dimensional problems.  
+- **Strong Contribution:** The recursive framework and KL duality argument are novel and theoretically rich.
+
+# Weaknesses
+
+- Overspecialized Language: Uses terms like pullback density, diffeomorphisms, and KL duality without intuitive explanations.  
+- Evidence Integration: Empirical results are emphasized but not contextualized in the introduction/body.
+
+# - Missing Statistical Rigor:
+
+- Runtime claims lack significance tests.  
+- Missing error bars for ESS in several tables.  
+- Confounding variables (e.g., batch size impact) are not discussed.
+
+- Domain Restrictions: Excludes autoregressive flows without quantifying impact, and relies on explicit energy functions limiting broader applicability.
+
+# Soundness
+
+Technical correctness of the recursive gradient computation and implicit differentiation is well established; empirical claims are supported by ESS benchmarks and runtime comparisons, though limitations on non-physical tasks remain unaddressed.
+
+# Presentation
+
+Accessible to experts but opaque to general readers; figures/tables need annotations and reproducibility details (e.g., optimizer, batch size, hardware).
+
+# Contribution
+
+Methodological advancement for efficient path gradients in both forward and reverse KL training; novel unification via base-space pullback, with potential for fairness- and physics-based applications.
+
+# Critical Questions for Authors
+
+1. Does Proposition 3.3 provide variance reduction beyond runtime?  
+2. How adapt to tasks with intractable energy functions?  
+3. Which architectures are incompatible with the framework?  
+4. Why are certain ESS values missing in Table 1?
+
+# Recommendation: Accept with revisions.
+
+# Constructive Feedback
+
+1. Expand literature comparison to control variate methods.  
+2. Clarify KL duality with analogies and roadmaps.  
+3. Improve statistical reporting (confidence intervals, significance tests).  
+4. Discuss domain limitations explicitly.
+
+5. Address runtime-tolerance trade-offs with figures.
+
+# Final Comments
+
+A computationally and theoretically impactful solution to NF path gradient inefficiencies; with enhanced comparisons and statistical rigor, it will be a strong ICLR contribution.
+
+Reviewer Confidence: High
+
+Rating:Accept
+
+# Below is an example review of a paper which has been rejected by ReviewAgent.
+
+# AlloNet Review
+
+Paper Title: Allostatic Control of Persistent States in Spiking Neural Networks for perception and computation (arXiv:2503.16085v1)
+
+# Summary
+
+The paper proposes AlloNet, a spiking neural network integrating allostasis (via the Hammel model) with ring attractor dynamics for controlling persistent neural activity (bumps) in numerical cognition tasks like subitization. The model aims to align internal representations with environmental inputs and demonstrates subitization performance consistent with human behavioral trends (e.g., subitizing limits for numbers  $>3$ ). However, model reaction times are slower than biological benchmarks, and reproducibility details are minimal.
+
+# Strengths
+
+- Conceptual Integration: Combines allostasis and ring attractors to control persistent states, offering a biologically motivated framework for abstract cognition.  
+- Technical Details: Provides rigorous mathematical formulations (e.g., synaptic weight equations in Eq. 1) and empirical validation through subtitization experiments, including reaction time and error rate analysis.  
+- Biological Grounding: Ties the model to hippocampal/entorhinal dynamics and human subitization behavior (Dehaene & Cohen 1994; Togoli & Arrighi 2021).  
+- Task Relevance: Subitization experiments capture qualitative aspects of human performance (e.g., numerosity-dependent reaction time variability).
+
+# Weaknesses
+
+- Novelty Overstatements: Claims a "novel unified framework" without addressing recent work on dynamic ring attractor control (e.g., Khona & Fiete 2022; Rapu & Ganguli 2024). The allostatic-coupling mechanism lacks clear distinction from gain modulation approaches.
+
+- Reproducibility Deficiencies: References NEST simulations with hardcoded parameters (e.g.,  $\tau_{1} = 75\mathrm{ms}$ ) but omits implementation details (Poisson spike calibration, hyperparameter search) and public code/data.  
+- Limited Theoretical Depth: Key design choices (e.g.,  $\sigma_{1} / \sigma_{2}$  values, Gaussian synaptic weights in Eq. 1) are not justified; no comparison to alternative regulation mechanisms (e.g., reinforcement learning).  
+- Narrow Task Scope: Subitization is the sole task demonstrated; broader claims (e.g., robotics/spatial navigation) are unsubstantiated.
+
+# Soundness
+
+Technical validity of the model dynamics is clear; however, experimental rigor is weakened by missing error bars and statistical comparisons to human benchmarks.
+
+# Presentation
+
+Jargon (e.g., "HGM/LGM," "ring attractor") needs definitions; figures lack error bars and clear labels; the manuscript lacks a dedicated "Contribution" section.
+
+# Contribution
+
+Advances biologically inspired control of persistent states but reuses established components (Hammel model, ring attractors) without sufficient novelty justification; broader applicability claims are speculative.
+
+# Questions for Authors
+
+1. How does AlloNet's allostatic framework differ from predictive coding approaches in Rapu & Ganguli (2024) or active inference models (Parr et al. 2022)?  
+2. What evidence supports "predictive adaptation" compared to non-learning SNN models?  
+3. Why were Gaussian synaptic weights chosen (Eq. 1)? Are there theoretical or empirical motivations?  
+4. Does the model replicate quantitative error rates from human subitization (e.g., Castaldi et al. 2021)?
+
+# Limitations
+
+Recent SNN literature gaps (e.g., Rapu & Ganguli
+
+2024) ignored; reaction times (1000 ms) lag biological norms; reliance on manual parameter tuning limits scalability.
+
+# Ethics Review
+
+No major ethical issues, but authors should discuss potential biases in AI systems mimicking human cognitive processes.
+
+# Rating & Confidence
+
+Rating: 5/10 (technically sound but under-theorized and narrow)
+
+Recommendation: Reject Unless Major Revisions Address (novelty justification, reproducibility materials)
+
+Confidence: High
+
+# Additional Comments
+
+1. Add a dedicated “Contribution” section post-introduction.  
+2. Contrast with recent work (e.g., Rapu & Ganguli 2024; Rapp & Nawrot 2020).  
+3. Release code and hyperparameters on a public repository.  
+4. Demonstrate broader tasks (e.g., spatial navigation) for claimed applicability.
+
+Final Outcome: Reject with Major Revisions
+
+Confidence in Decision: High
+
+# A.7 Qualitative Review Analysis
+
+# A.7.1 Excerpt 1: Assessing Strengths (Empirical Results)
+
+As shown in Table 10, all reviews identify the empirical results of the paper as a key strength. However, their descriptions vary in detail and impact.
+
+# Analysis
+
+- ReviewAgent: Provides specific metrics (40.7% accuracy). However, it focuses on a less critical metric (95.8% vs 100% valid samples), which is part of a motivating example.  
+- ReviewAgent-Ref: Adds a comparative metric ("outperforming PPO by  $63\%$ ) which provides stronger context for the  $40.7\%$  accuracy figure.  
+- ReviewAgent-Imp: Improves significantly by grounding its claims in the paper's structure,
+
+<table><tr><td>Review Version</td><td>Excerpt on Strengths</td></tr><tr><td>ReviewAgent</td><td>&quot;Strong empirical demonstration through tasks (e.g., random number generation, arithmetic, infilling) showing GFlowNet outperforms PPO in diversity (95.8% vs. 100% valid samples in motivating example) and accuracy (40.7% in 5-operand arithmetic).&quot;</td></tr><tr><td>ReviewAgent-Ref</td><td>&quot;Strong Empirical Results: Demonstrates 40.7% out-of-distribution accuracy on arithmetic tasks (5 operands) with GFlowNets, outperforming PPO by 63%. &quot;</td></tr><tr><td>ReviewAgent-Imp</td><td>&quot;Novel Application: GFlowNet fine-tuning is a data-efficient alternative to PPO/RL, with improvements in diversity (Figure 3) and OOD generalization (Table 4, 40.7% accuracy).&quot;</td></tr><tr><td>ReviewAgent-Imp-Ref</td><td>&quot;Empirical Validation: Results (10.9% absolute improvement in subjectivity classification with 10 labels, 63% improvement in arithmetic OOD generalization) are meaningful. Qualitative examples (Table C.3-C.8) highlight the method&#x27;s efficacy in generating coherent and diverse samples.&quot;</td></tr><tr><td>Sakana</td><td>&quot;Comprehensive experiments across multiple tasks (sentence continuation, story infilling, arithmetic) with clear metrics and comparisons to established methods (e.g., PPO, diverse beam search).&quot;</td></tr><tr><td>Marg V1</td><td>&quot;Strong results on arithmetic reasoning and subjectivity classification, including out-of-distribution generalization (e.g., integer arithmetic).&quot;</td></tr><tr><td>Marg V2</td><td>&quot;Empirical results span diverse tasks (number generation, infilling, reasoning) with strong quantitative improvements (e.g., 10.9% absolute gain on subjectivity classification).&quot;</td></tr></table>
+
+Table 10: A.7.1 Comparison of Review Versions and Their Excerpts on Strengths
+
+referencing specific figures and tables (Figure 3, Table 4). This signals to the authors that the reviewer has carefully examined the evidence provided.
+
+- ReviewAgent-Imp-Ref: Synthesizes multiple quantitative results (10.9% improvement on one task, 63% on another) and uniquely incorporates evidence from the appendix (Table C.3-C.8). Referencing the appendix demonstrates a thorough review and substantiates claims about "coherent and diverse samples" with concrete proof.  
+- Sakana, Marg V1, and Marg V2: Identify the strength but are relatively vague in the details compared to ReviewAgent.  
+- Progression of quality: Moves from stating a result (ReviewAgent) to contextualizing it (ReviewAgent-Imp-Ref), referencing its location in the paper, and finally synthesizing multiple data points from the main body and appendix to form a comprehensive judgment.
+
+The ReviewAgent and its variants provide the most specific and actionable feedback: each variant
+
+cites exact figures (e.g.,  $40.7\%$  OOD accuracy on 5-operand arithmetic,  $10.9\%$  gain in subjectivity classification,  $63\%$  improvement over PPO) and refers to concrete tables or figures that substantiate diversity and generalisation claims, even noting qualitative evidence (Tables C.3-C.8). This level of detail tells authors precisely where their empirical story is strong. By contrast, Sakana offers only a broad statement about "comprehensive experiments" without any numbers or table references, while MARG V1 and V2 mostly just repeat that results are "strong" or "diverse," with almost no granular performance data.
+
+# A.7.2 Excerpt 2: Assessing Weaknesses (Missing Context and Citations)
+
+In Table 11, a common weakness identified is the failure of the paper to contextualize its contribution by citing relevant work.
+
+# Analysis
+
+- ReviewAgent: Correctly identifies missing citations and names relevant authors. This is the minimum requirement for helpful feedback.  
+- ReviewAgent-Ref: Improves by providing full, dated citations for highly relevant recent
+
+<table><tr><td>Review Version</td><td>Excerpt on Weaknesses</td></tr><tr><td>ReviewAgent</td><td>&quot;Incomplete citations: Missing direct comparisons with prior GFlowNet applications in language (e.g., Deleu et al.). Also, understatement of overlap with methods like self-consistency for chain-of-thought (Wang et al.) weakens novelty claims.&quot;</td></tr><tr><td>ReviewAgent-Ref</td><td>&quot;Misplaced Citations: Omitted recent work (e.g., Tree of Thoughts (Yao et al., 2023), A-NeSI (van Krieken et al., 2023)) that could contextualize novelty.&quot;</td></tr><tr><td>ReviewAgent-Imp</td><td>&quot;Incomplete Baseline Analysis: RL methods (e.g., PPO) are critiqued for mode collapse but lack in-depth theoretical discussion of credit assignment issues or comparisons with recent SOTA (Tree of Thoughts, Yao et al., 2023).&quot;</td></tr><tr><td>ReviewAgent-Imp-Ref</td><td>&quot;Missing Literature Comparison: Recent peer-reviewed GFlowNet work (...) and prompting techniques ([Tree of Thought, 2023]; [Self-Consistency, 2023b]) are not compared, weakening novelty claims.&quot;</td></tr><tr><td>Sakana</td><td>&quot;Limited comparison with recent methods for structured generation, such as those using masked language models or other Bayesian-inspired approaches (e.g., Metropolis-Hastings based samplers).&quot;</td></tr><tr><td>Marg V1</td><td>&quot;Situate the work within the broader literature on amortized inference (e.g., VAEs, normalizing flows) and LLM fine-tuning. Clarify how GFlowNets differ from prior work in domains like molecule generation.&quot;</td></tr><tr><td>Marg V2</td><td>&quot;Only compares against PPO and supervised fine-tuning, ignoring other RL or Bayesian methods (e.g., Hamiltonian Monte Carlo, contrastive divergence).&quot;</td></tr></table>
+
+Table 11: A.7.2 Comparison of Review Versions and Their Excerpts on Weaknesses
+
+work. This is more actionable for the authors and shows the reviewer is current with the field.
+
+- ReviewAgent-Imp: Provides better feedback by linking the missing citation (Tree of Thoughts) to a specific theoretical gap ("lack in-depth theoretical discussion of credit assignment issues"). This elevates the comment from a simple missing reference to a deeper conceptual weakness.  
+- ReviewAgent-Imp-Ref: Not only provides specific, dated citations but also categorizes them ("peer-reviewed GFlowNet work," "prompting techniques"). This structured feedback helps the authors understand the different research streams they need to engage with, making the revision process more efficient and targeted.  
+- Sakana: Correctly identifies a weakness and suggests relevant classes of methods ("masked language models," "Metropolis-Hastings"). This is helpful and actionable feedback.  
+- Marg V1: More comprehensive. It not only suggests comparing to other methods
+
+in the same field (VAEs, normalizing flows) but also challenges the authors to compare GFlowNet applications across different domains ("molecule generation"). This pushes the authors to provide a much deeper contextualization of their work's novelty.
+
+- Marg V2: Provides specific and relevant alternative methods ("Hamiltonian Monte Carlo, contrastive divergence").
+
+The best feedback does not just list missing papers; it explains why their absence is problematic and what specific theoretical or empirical gaps they could have filled, thereby providing a clear path for revision.
+
+We can see that ReviewAgent and its variants are able to pinpoint concrete literature gaps, citing specific missing works such as Deleu et al. on GFlowNets, Tree of Thoughts, A-NeSI, and self-consistency while also calling for deeper theoretical or baseline analyses (e.g., credit-assignment issues with PPO). The Ref and Imp-Ref versions further strengthen this by explicitly flagging how the absence of these comparisons weakens novelty claims, demonstrating a fine-grained grasp of citation relevance. By contrast, Sakana offers only a broad remark about "masked language models"
+
+and “Bayesian-inspired approaches,” without naming concrete papers or explaining their significance. MARG V1 and V2 remain similarly generic, suggesting the authors “situate the work” or “clarify differences” but providing no precise references, thereby giving authors little actionable direction for improving scholarly positioning.
+
+# A.7.3 Excerpt 3: Assessing Overall Contribution and Novelty
+
+As shown in Table 12, The assessment of the paper's core contribution reveals different levels of analytical depth.
+
+# Analysis
+
+- ReviewAgent-None: Provides a solid, balanced summary. Correctly identifies the contribution and points to a relevant theoretical weakness (comparison to VAEs).  
+- ReviewAgent-Ref: Uses more impactful language ("transformative framework") but the core point remains similar to the previous review.  
+- ReviewAgent-Imp: Improves by quantifying the contribution with a specific result ("10× improvement in arithmetic OOD accuracy"). Linking the novelty claim directly to a powerful empirical outcome makes the praise more concrete and impactful.  
+- ReviewAgent-Imp-Ref: Gives a sophisticated analysis. Separates the contribution into Innovation (the novelty of the idea) and Impact (the practical result). Furthermore, directly critiques a specific, overreaching phrase from the paper ("universal reasoner"), showing a close reading and holding the authors accountable for their precise claims.  
+- Sakana: Effectively describes the problem area and its significance. The phrasing is clear but standard.  
+- Marg V1: More precise by using specific technical terms to frame the contribution ("novel application of GFlowNets to amortized Bayesian inference"). Explicitly names "chain-of-thought" as a key task.  
+- Marg V2: Offers an impactful summary. Frames the work not just as an application but as a solution to a "critical limitation in
+
+autoregressive sampling." This phrasing elevates the perceived importance of the paper by positioning it as a remedy for a fundamental and widely recognized problem.
+
+A good review moves beyond a general summary to dissect the contribution into its constituent parts (e.g., idea vs. result) and challenges the authors' specific language and claims, ensuring the paper's final version is precise and well-supported.
+
+ReviewAgent and its refined variants stand out for delivering targeted, actionable feedback: they not only highlight contributions, such as the  $10 \times$  OOD-accuracy gain and the importance of amortized posterior sampling, but also pinpoint concrete gaps (e.g., insufficient contrast with VAEs, premature "universal reasoner" claim) and situate the work within broader contexts like probabilistic programming and non-RL baselines. ReviewAgent-Imp-Ref explicitly labels the different parts of review into "Innovation", "Impact", and "Limitation". In contrast, Sakana provides a generic endorsement of problem relevance, lacking specific criticisms or suggestions, while MARG V1 and V2 largely reiterate the paper's premise with minimal depth, providing little actionable guidance for authors or program chairs.
+
+<table><tr><td>Review Version</td><td>Excerpt on Contribution</td></tr><tr><td>ReviewAgent</td><td>&quot;The work offers a novel method for amortized posterior sampling in autoregressive LLMs using GFlowNets, with clear applications to reasoning, infilling, and low-data learning. However, the theoretical distinction from existing variational inference paradigms (e.g., VAEs) is underdeveloped.&quot;</td></tr><tr><td>ReviewAgent-Ref</td><td>&quot;The work presents a transformative framework for amortized posterior sampling, critical for tasks requiring diverse, structured outputs. However, gaps in contextualizing GFlowNets within probabilistic programming and omitting comparisons with non-RL methods (e.g., VAEs) dilute the novelty claim.&quot;</td></tr><tr><td>ReviewAgent-Imp</td><td>&quot;The work innovatively applies GFlowNets to latent-variable reasoning in LLMs, offering a Bayesian alternative to RL. The 10× improvement in arithmetic OOD accuracy is a strong contribution.&quot;</td></tr><tr><td>ReviewAgent-Imp-Ref</td><td>&quot;Innovation: The work introduces a novel method for amortized posterior sampling... Impact: It improves over existing methods in low-data settings... Limitations: The &quot;universal reasoner&quot; claim (Conclusion) is premature without validation on complex tasks...&quot;</td></tr><tr><td>Sakana</td><td>&quot;Tackles a significant and practical problem of intractable posterior sampling in LLMs, with real-world applications in reasoning and constrained generation.&quot;</td></tr><tr><td>Marg V1</td><td>&quot;The paper introduces a novel application of GFlowNets to amortized Bayesian inference in LLMs, addressing intractable posterior sampling in tasks like chain-of-thought reasoning and constrained generation.&quot;</td></tr><tr><td>Marg V2</td><td>&quot;Proposes a novel framework for amortized Bayesian inference in LLMs using GFlowNets, addressing a critical limitation in autoregressive sampling for tasks like infilling and chain-of-thought reasoning.&quot;</td></tr></table>
+
+Table 12: A.7.3 Comparison of Review Versions and Their Excerpts on Contribution

reviewevalanevaluationframeworkforaigeneratedreviews/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0784052115abaac5b95ced12837a6b53e6e06def2f24627ad08a396df4c91c39
+size 1419479

reviewevalanevaluationframeworkforaigeneratedreviews/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7acf5784c12136961e1fc64191ebdc78bf4e0db9f6b90036e45803a07362212e
+size 755263

revisitingchainofthoughtpromptingzeroshotcanbestrongerthanfewshot/d4bd8186-0f6f-423e-b442-141a7646b2da_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41384ae4131f66141d2ffb79ad2c7914d469fed3b9ebbcd2606fc471b4b6f435
+size 124069

revisitingchainofthoughtpromptingzeroshotcanbestrongerthanfewshot/d4bd8186-0f6f-423e-b442-141a7646b2da_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e48ac1efe61bbca860df0e4b3966901906057caf4a000d7484611bbb5ca7c88
+size 147309

revisitingchainofthoughtpromptingzeroshotcanbestrongerthanfewshot/d4bd8186-0f6f-423e-b442-141a7646b2da_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91a22040fcd0bbc6a8741208d5e35509ed536389b41a16c3d142d1e826774964
+size 27608229

revisitingchainofthoughtpromptingzeroshotcanbestrongerthanfewshot/full.md

@@ -0,0 +1,530 @@
+# Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot
+
+Xiang Cheng $^{1}$ , Chengyan Pan $^{1}$ , Minjun Zhao $^{2}$ , Deyang Li $^{2}$ , Fangchao Liu $^{2}$ , Xinyu Zhang $^{2}$ , Xiao Zhang $^{1}$ , Yong Liu $^{1*}$
+
+$^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China  $^{2}$ Huawei Poisson Lab
+
+{chengxiang1,2023100836,zhangx89,liuyonggsai}@ruc.edu.cn {zhaominjun1, lideyang2, liufangchao, zhangxinyu35}@huawei.com
+
+# Abstract
+
+In-Context Learning (ICL) is an essential emergent ability of Large Language Models (LLMs), and recent studies introduce Chain-of-Thought (CoT) to exemplars of ICL to enhance the reasoning capability, especially in mathematics tasks. However, given the continuous advancement of model capabilities, it remains unclear whether CoT exemplars still benefit recent, stronger models in such tasks. Through systematic experiments, we find that for recent strong models such as the Qwen2.5 series, adding traditional CoT exemplars does not improve reasoning performance compared to Zero-Shot CoT. Instead, their primary function is to align the output format with human expectations. We further investigate the effectiveness of enhanced CoT exemplars, constructed using answers from advanced models such as Qwen2.5-Max and DeepSeek-R1. Experimental results indicate that these enhanced exemplars still fail to improve the model's reasoning performance. Further analysis reveals that models tend to ignore the exemplars and focus primarily on the instructions, leading to no observable gain in reasoning ability. Overall, our findings highlight the limitations of the current ICL+CoT framework in mathematical reasoning, calling for a re-examination of the ICL paradigm and the definition of exemplars.
+
+# 1 Introduction
+
+As Large Language Models (LLMs) continue to scale, LLMs exhibit emergent In-Context Learning (ICL) capabilities (Brown et al., 2020), enabling them to perform target tasks by conditioning on a few exemplars without any additional parameter updates. Furthermore, the use of Chain-of-Thought (CoT) exemplars (Wei et al., 2022) in ICL guides models to reason step-by-step. This approach is commonly referred to as Few-shot CoT. Kojima
+
+![](images/99905700d13ddcc45ebec78aa9a27d4dcb745d5459c71ff8bba9d6782c845811.jpg)  
+Figure 1: Accuracy under different prompting settings on GSM8K (top) and MATH (bottom). We observe that the Zero-shot setting consistently achieves strong performance, suggesting that the model may not attend to the CoT exemplars. See Section 5 for the full experimental results.
+
+et al. (2022) further showed that simply appending the instruction "Let's think step by step" can trigger multi-step reasoning even without exemplars, giving rise to the Zero-shot CoT paradigm, an overview of them is shown in Figure 2.
+
+Existing research primarily focuses on how the quality, order, and number of exemplars influence ICL performance, proposing various strategies for exemplar construction and selection to enhance model performance across different task settings (Lu et al., 2022; Chen et al., 2023; Kim et al., 2022; Purohit et al., 2024). In addition, several studies have investigated the underlying mechanisms and influencing factors of ICL from either theoretical or empirical perspectives (Ren and Liu, 2024; Xie et al., 2022; Min et al., 2022; Wei et al., 2023; Wang et al., 2023). However, most of these strategies and experimental conclusions are based on earlier, weaker models. As foundation models become increasingly powerful, it is necessary to revisit a central question: In mathematical reasoning tasks, can CoT exemplars still improve the
+
+reasoning performance of recent strong models?
+
+In this paper, we aim to investigate the actual role of CoT exemplars in mathematical reasoning tasks. We conduct systematic experiments on two representative math reasoning datasets, GSM8K (Cobbe et al., 2021) and MATH (Hendrycks et al., 2021), using several recent open-source LLMs. We first identify a common evaluation bias in open-source evaluation frameworks (Contributors, 2023; Lambert et al., 2024) in GSM8K, which significantly underestimates the performance of Zero-shot CoT, as discussed in Section 4. After correcting for this issue, we compare Few-shot CoT with Zero-shot CoT prompting. Our results show that recent strong models already exhibit strong reasoning capabilities under the Zero-shot CoT setting, and the primary role of Few-shot CoT exemplars is to align the output format with human expectations. Subsequent analysis confirms that adding traditional CoT exemplars does not improve reasoning performance (See Section 5.1). Inspired by recent advances in reasoning models with more sophisticated capabilities (Guo et al., 2025; Jaech et al., 2024), we then examine the effectiveness of enhanced CoT demonstrations constructed using answers generated by advanced models such as Qwen2.5-Max and DeepSeek-R1. Experimental results indicate that, regardless of enhancement, models tend to ignore the content of exemplars in mathematical reasoning tasks and fail to acquire advanced capabilities such as self-reflection (See Section 5.3). As a result in figure 1, CoT exemplars do not lead to improved reasoning performance in recent models.
+
+To summarize, our main empirical findings in mathematical reasoning tasks are as follows:
+
+1. The primary function of CoT exemplars is to align the output format, and this effect persists regardless of the model's reasoning ability.  
+2. Traditional CoT exemplars do not enhance the reasoning performance of strong models, although they may benefit weaker models.  
+3. Enhanced CoT exemplars also fail to improve reasoning ability in strong models, as these models tend to ignore the CoT content.
+
+# 2 Related Work
+
+CoT Prompting ICL enables LLMs to perform tasks without fine-tuning (Brown et al., 2020), but it often falls short in complex reasoning scenarios.
+
+![](images/1d6743f82ef78c6411e5596f8ab449645af4975da120779aaedfe4f680d1e770.jpg)  
+Figure 2: An overview of ICL and CoT prompting. The figure illustrates the Few-shot CoT setting, where the model performs reasoning based on provided demonstrations and a test question. When no demonstrations are given, the setting corresponds to Zero-shot CoT.
+
+To address this, CoT prompting (Wei et al., 2022) introduces intermediate reasoning steps to guide model outputs. Building on CoT, researchers have proposed various extensions to enhance reasoning capabilities. For instance, Tree-of-Thought (Yao et al., 2023) generalizes CoT to tree-structured reasoning, while Graph-of-Thought (Besta et al., 2024) further expands it to graph-based structures. The Least-to-Most framework (Zhou et al., 2023) decomposes complex problems into simpler subproblems and solves them sequentially.
+
+Exemplar Selection In addition to improving CoT itself, numerous studies have explored how exemplar quality, quantity, diversity, and ordering affect ICL performance (Lu et al., 2022; Li et al., 2023; Ma et al., 2023; Zhang et al., 2022). A variety of exemplar selection strategies have been proposed. Fu et al. (2023) recommend selecting exemplars with higher reasoning complexity (i.e., involving more intermediate steps), while Hongjin et al. (2022) emphasize diversity and introduce the VoteK algorithm. Other representative methods include DPP (Ye et al., 2023a), a probabilistic model that measures both subset diversity and quality by using the determinant of a kernel matrix—the larger the determinant, the more diverse and representative the selected exemplars. MMR (Ye et al., 2023b) balances relevance and diversity by iteratively selecting items that are highly relevant to the query while remaining dissimilar to already chosen exemplars. EXPLORA (Purohit et al., 2024) evaluates exemplar subsets without relying on model confidence scores.
+
+Understanding CoT Prompting Beyond methodology, a growing body of research has sought to understand the mechanisms behind ICL and CoT prompting. Theoretical investigations (Dai et al., 2023; Li et al., 2024; Ren and Liu, 2024; Mahankali et al., 2023) offer insights
+
+![](images/f58652990fe5dc8a6aa284c72c7459be9700ad2727c1bb91009cce606cd9c63e.jpg)  
+Figure 3: Accuracy of different models on the GSM8K dataset under varying numbers of exemplars. Few-shot examples are taken from Wei et al. (2022). Only Zero-shot-fixed applies evaluation bias correction, as described in Section 4; all other settings retain uncorrected results for comparison.
+
+![](images/ea150801ef2d0e2863685dc8b3aaeeacda29b4d63e5979dc8e1ad1e997d11830.jpg)  
+Figure 4: Accuracy of different models on the GSM8K dataset under various ablation settings. Replace_Q denotes replacing the question in each exemplars with "xxx". Replace_QA replaces both the question and answer with "xxx" but retains the final phrase "So the answer is ...". Replace_ALL replaces the question, answer, and the final phrase with "xxx". See figure 17, 18, and 19 for input examples, respectively. Other settings follow those in Figure 3.
+
+into the learning dynamics of ICL, while empirical studies probe the effectiveness of CoT. For instance, Min et al. (2022) suggest that exemplars primarily provide distributional rather than semantic information—though their analysis is limited to classification tasks. In the context of reasoning, Levy et al. (2024) report that longer input contexts may hurt performance, and Sprague et al. (2025) find that the benefits of CoT are mainly confined to mathematical and logical reasoning.
+
+Our work complements these lines of research through a systematic empirical study on mathematical reasoning. While prior studies have provided important insights, they are mostly based on earlier and weaker models, whose conclusions may not fully extend to recent, stronger models. We find that, for recent strong models, CoT exemplars primarily function to align output format rather than enhance reasoning ability. This challenges the prevailing assumption that CoT-based ICL reliably improves performance in math reasoning tasks.
+
+# 3 Experimental Setup
+
+Models To thoroughly validate our conclusions, we evaluate a variety of open-source language models, including the Qwen2.5 series (ranging from 0.5B to 72B parameters) (Yang et al., 2024a), the LLaMA3 series (1B to 70B) (Grattafori et al.,
+
+2024), the Gemma2 series (2B and 9B) (Team et al., 2024), and Ministral-8B (Mistral AI, 2024). In addition, to examine the effectiveness of CoT prompting on earlier and weaker models, we include LLaMA2-7B (Touvron et al., 2023) and Qwen7B (Bai et al., 2023) for comparative analysis. All models used in our experiments are instruction-tuned variants. More details can be found in Appendix A.1.
+
+Datasets We focus on mathematical reasoning tasks and conduct experiments on two datasets of varying difficulty: GSM8K (Cobbe et al., 2021) and MATH (Hendrycks et al., 2021). To ensure accuracy, we perform inference and evaluation on the full test sets of both datasets and report the complete results. More details can be found in Appendix A.2.
+
+Environment and Hyperparameters We utilize the open-source inference framework OpenCompass (Contributors, 2023) and vLLM (Kwon et al., 2023) as the backend to run all experiments. Notably, all experiments incorporate a CoT instruction in the prompt: "Please reason step by step, and put your final answer within \boxed{\}.}" For reproducibility, all experiments are conducted using a fixed random seed of 42. Notably, since greedy decoding is deterministic, the fixed seed does not influence the inference results under a fixed hard-
+
+![](images/c16ce9702ac935164106a3aef9cec42c9f277acc93372caa68e32f2e7d98a0de.jpg)
+
+![](images/b3ca46b9c9ab466654c7667dee0455f6e6b6bc9bd10d58b4d806e62b01ad7497.jpg)  
+Figure 5: Accuracy of different models under various retrieval methods with a fixed number of 8 retrieved exemplars. The top figure shows results on the MATH dataset, and the bottom figure shows results on the GSM8K dataset.
+
+ware setup. Hence, we do not report the mean or standard deviation of the results. More details can be found in Appendix A.3.
+
+# 4 Exemplars Help Mitigate Evaluation Bias
+
+Evaluation Bias in GSM8K Existing evaluation frameworks for GSM8K (e.g., OpenCompass (Contributors, 2023), Open-Instruct (Lambert et al., 2024)) typically extract the last number from model outputs as the predicted answer. However, in Zero-shot CoT prompting, answers are often enclosed in "boxed{}" expressions. This mismatch leads to misjudgments during evaluation, as illustrated in Figure 21. To address this, we modify the evaluation script to extract the number inside \boxed{}\), reducing artificially low accuracy caused by output-format misalignment. We consider this a form of evaluation bias that affects fair assessment, either due to oversight or simplification.
+
+Exemplars Aid Format Alignment As shown in Figure 3, after correcting the evaluation method, the Zero_shot fixed setting yields substantial gains, surpassing all others. This indicates that the original poor performance of Zero_shot stems not from reasoning limitations, but from output-evaluation mismatch. Moreover, Few_shot consistently outperforms Zero_shot, suggesting that exemplars help standardize output format and improve answer extraction. Thus, in math reasoning tasks, the primary benefit of exemplars lies in aligning the model's output format. Interestingly, for Mistral-8B, exemplars can induce overfitting to simplified reasoning paths, diminishing their effectiveness.
+
+Complete Answer Structure is Key Factor Ablation results in Figure 4 show a consistent performance drop as more content is masked—from Replace_Q to Replace_QA to Replace_A11. This highlights the importance of preserving the full answer structure for effective format alignment. Even partial cues (e.g., "So the answer is ...") prove beneficial, whereas fully removing informative content reverts performance to the Zero_shot baseline. This confirms that exemplars primarily guide answer formatting rather than reasoning itself.
+
+# 5 CoT Exemplars can't improve reasoning ability of strong models
+
+The preceding sections have shown that the primary contribution of exemplars lies in aligning the output format rather than enhancing reasoning ability. However, since we previously used a fixed set of 8 exemplars, an open question remains: Can exemplars improve the reasoning ability of recent LLMs if we consider different impact factors such as retrieval method, model's intrinsic ability and the quality of exemplars?
+
+# 5.1 The Impact of the retrieval method
+
+In this section, we revisit the classical CoT prompting paradigm, in which in-context exemplars are retrieved from the training set of the original dataset. This setup aligns with prior work and allows us to evaluate whether recent large language models (LLMs) still benefit from exemplars under this conventional configuration. To ensure consistency, we apply our corrected evaluation method across a variety of models and compare their performance on GSM8K and MATH using several established exemplar selection strategies. These include
+
+![](images/ac9d35b2ac0eb570029b5e8c047e1d6001f45230f9d5d1963f519de2f15fbe32.jpg)
+
+![](images/d648933f150f808532d8a25566282e2c91fddccf4ab4373cc3208f59ef386a91.jpg)  
+Figure 6: Accuracy of different weaker models under various retrieval methods with a fixed number of 8 retrieved exemplars. The top figure shows results on the MATH dataset, and the bottom figure shows results on the GSM8K dataset.
+
+![](images/c44beb05a2f25025b99e5e93afa0ba4cb7a64e2b42ba96d300670c60deae08e3.jpg)
+
+![](images/c82d0241b4bc22aece84d34dca07ffe3da0581a9bfafbe2aae34aec923adde7a.jpg)  
+Figure 7: Accuracy variation with different numbers of retrieved exemplars under various retrieval methods, evaluated using Qwen2.5-7B and Qwen2.5-72B. The top figure shows results on the MATH dataset, and the bottom figure shows results on the GSM8K dataset.
+
+Complexity-based selection (Fu et al., 2023), FastVotek (Hongjin et al., 2022), DPP (Ye et al., 2023a), MMR (Ye et al., 2023b), and EXPLORA (Purohit et al., 2024), along with simple TopK and Random baselines.
+
+Retrieval-Based Methods Fall Short of Zero-Shot Performance We uniformly retrieve 8 exemplars for each selection method and report the results in Figure 5. Across most configurations—regardless of model or dataset—Few-shot performance with retrieval-based methods is comparable to or worse than the Zero-shot baseline. This observation suggests that for advanced lan
+
+guage models, in-context exemplars do not enhance reasoning ability, but primarily function is align output formats. Notably, there are a few exceptions. For example, LLaMA3.1-8B exhibits marginal improvements under the 8-shot setting. However, we attribute this to inherent experimental variance rather than genuine reasoning gains. A detailed analysis is provided in Appendix C.
+
+Varying the Number of Exemplars Still Fails to Surpass Zero-Shot Given that using 8 retrieved exemplars often fails to outperform the Zero-shot baseline, we further investigate the impact of varying the number of in-context exemplars. As shown in Figure 7, Zero-shot prompting achieves the highest accuracy in most settings. Nevertheless, certain retrieval methods occasionally yield slightly better performance, particularly on GSM8K. For example, the Complexity-based retrieval method marginally outperforms Zero-shot when retrieving 4 or 6 exemplars on two different models. However, the improvements are minimal—around  $0.2\%$  in accuracy. It can be reasonably attributed to inherent evaluation variance. Such small fluctuations are more likely to occur on relatively simpler datasets like GSM8K. In contrast, on the more challenging MATH dataset, nearly all retrieval-based configurations consistently underperform relative to the Zero-shot baseline.
+
+Overall, these results reinforce the conclusion that Zero-shot prompting remains the most effective approach in the vast majority of cases. This supports the emerging perspective that traditional CoT prompting paradigms no longer significantly enhance the reasoning capabilities of recent LLMs.
+
+![](images/5309264cf3405cddb2033f369853a06ae1804ad2d1747c4efa1b1086854c6e59.jpg)  
+Figure 8: Accuracy under different numbers of exemplars when using DeepSeek R1 responses (marked as R1-nshot) and Qwen2.5-max responses (marked as Qwen-nshot) as exemplars. The left figure shows results on the MATH dataset, and the right figure shows results on the GSM8K dataset.
+
+![](images/95fb658799eda8e52546c047639609a70ac91b7a48fd969e6aead912e3876b5e.jpg)
+
+# 5.2 The Impact of exemplars Is Determined by the Model's Intrinsic Capability
+
+In the previous experiments, we observed that in-context exemplars do not enhance the reasoning ability of recent models such as Qwen2.5 series. Does this contradict earlier findings from exemplar selection studies, such as those by Fu et al. (Fu et al., 2023)? To further investigate the role of exemplars, we conducted experiments on relatively weaker models. Specifically, we evaluated a set of smaller but recent models (LLaMA3.2-1B, LLaMA3.2-3B, Qwen2.5-0.5B, Qwen2.5-1.5B), as well as several older models (LLaMA3-8B, LLaMA2-7B, Qwen-7B). The same prompt templates were used as in previous experiments, and all model responses were post-processed to eliminate evaluation artifacts and isolate the true effect of exemplars.
+
+Since all outputs were corrected prior to evaluation, the only potential benefit of in-context exemplars in this experiment lies in improving reasoning ability, not output alignment. As shown in Figure 6, model performance varies significantly. For relatively strong models such as LLaMA3.2-3B and Qwen2.5-1.5B, the Zero_shot setting yields the highest accuracy, indicating that adding exemplars does not improve reasoning. This is consistent with our findings on stronger models, reaffirming that for capable models, exemplars primarily serve as output format guides rather than improve reasoning.
+
+However, for weaker models (e.g., LLaMA3.2-1B) and older models with larger parameter counts (e.g., LLaMA2-7B and Qwen-7B), we observe a significant improvement in accuracy when exemplars are provided. This suggests that for such models, in-context exemplars indeed help augment reasoning by supplying intermediate steps that the model struggles to generate on its own. We hy
+
+pothesize that these weaker or older models lack the complex reasoning patterns that more recent models have acquired through pretraining and instruction tuning, and thus rely more heavily on external exemplars.
+
+Therefore, we conclude that the effectiveness of CoT exemplars depends on the model's inherent capabilities. Traditional CoT exemplars do not improve the reasoning ability of already-strong models but can play a supportive role for weaker models. Hence, our findings are not in conflict with previous work; rather, they offer a complementary perspective by showing that the utility of exemplars is model-dependent.
+
+# 5.3 Is traditional CoT exemplars too easy for strong models?
+
+Previous experiments suggest that traditional CoT prompting strategies are largely ineffective for current open-source LLMs. A natural intuition is that the implicit reasoning paths embedded in standard CoT exemplars may be less sophisticated than the models' own Zero-shot reasoning capabilities. This raises an important question: can enhanced CoT exemplars benefit these strong models?
+
+With the emergence of high-performing Reasoning Large Language Models (RLLMs) such as OpenAI o1 (Jaech et al., 2024) and DeepSeek R1 (Guo et al., 2025), Long Chain of Thought have shown potential in guiding model reasoning. Motivated by this, we consider two enhanced settings: (1) using responses from DeepSeek-R1 as exemplars, and (2) using responses from a stronger LLM, Qwen2.5-Max, as exemplars. We conduct experiments across the Qwen2.5 family of models (7B, 14B, and 72B). Detailed examples of the input formats are provided in Appendix E.
+
+Quality Helps, but Zero-Shot Still Dominates For each enhanced configuration, we further vary
+
+![](images/479c4081088611ecce60dfad2490f973f0babace6b02b33aef72c6325165489e.jpg)
+
+![](images/f6a8889207a54acb291f83214579e109bdebf1d715bff31edbcd01bdd99d5b63.jpg)
+
+![](images/020cfe6b50916c4c1fb74ed977d0b27ebab7c5b20bac35c337b5fc733cb159fc.jpg)
+
+![](images/b3cb4b95f252e0550c3ebf892be035eddb326886e758d6fb050d932c32438afc.jpg)  
+Figure 9: Ablation study on noise injection for three types of exemplars, shown from left to right: exemplars answered by Qwen-max, exemplars answered by R1, and traditional CoT (Chain-of-Thought) exemplars. The top figure shows results on the MATH dataset, and the bottom figure shows results on the GSM8K dataset. Base denotes the original exemplars without noise. Noise50 randomly replaces  $50\%$  of the tokens with "XXX". Shuffle completely shuffles the words. Replace-xxx replaces all words with "XXX".
+
+![](images/afb5ca9fba93946da26d9e83b64fba2dea9751158126a9aeea2928fe02fff6c3.jpg)
+
+![](images/5e2fd5668d872ea7ad231cca07041844c0debea5e5cae446638804304bbba79c.jpg)
+
+the number of exemplars. Due to the relatively long responses generated by DeepSeek-R1, we accordingly limit the number of exemplars to a maximum of four shots to ensure comparability in input length. The corresponding results are shown in Figure 8. We observe that enhanced exemplars generally outperform the standard 8-shot CoT setting. In certain configurations, performance may even exceed that of the Zero-shot baseline, such as Qwen2.5-72B on the MATH dataset with the Qwen6shot setting. Nevertheless, Zero-shot prompting consistently achieves strong accuracy across both datasets without introducing additional context overhead. These findings indicate that while improving exemplar quality is indeed helpful, the reasoning capability of modern large language models is already sufficiently strong that changes in exemplar formatting yield only limited or no improvement over Zero-shot prompting.
+
+# 6 Why CoT exemplars is not useful for strong models?
+
+In this section, we further investigate the reasons behind the ineffectiveness of CoT exemplars. We begin with ablation studies, followed by an analysis of attention visualization results.
+
+# 6.1 Ablation Study on Noisy Exemplars
+
+To further investigate why exemplars fail to improve performance, we conduct ablation experiments across three types of CoT exemplars: Traditional CoT, R1-enhanced CoT (from DeepSeek
+
+R1), and Qwen2.5-Max-enhanced CoT. Specifically, for the R1-enhanced configuration, we use 4-shot exemplars, while 8-shot is used for the other settings. We introduce varying levels of noise into the exemplars and evaluate their impact on model performance. Experiments are conducted on the Qwen2.5 series (7B, 14B, and 72B) across both the GSM8K and MATH datasets.
+
+Exemplars Are Not Crucial for Recent LLMs As shown in Figure 9, we observe that in most settings, adding noise to the exemplars does not lead to significant performance degradation. This is especially evident for the larger Qwen2.5-72B model, where even the Noise50 configuration can match or slightly outperform the Base setting. These findings suggest that the models may selectively ignore the exemplars and instead rely on their intrinsic reasoning ability. Thus, the performance observed under Few-shot settings may not arise from the informative content of the exemplars, but rather from the model's inherent Zero-shot capabilities.
+
+# 6.2 Attention Visualization
+
+The previous results suggest that neither standard CoT prompts nor enhanced exemplars substantially improve model reasoning, and that models may not actively attend to these exemplars during inference. To investigate this further, we analyze the attention distribution of the Qwen2.5-7B model on GSM8K under Few-shot settings. Transformer-based models (Vaswani et al., 2017) rely on multi-head selfattention, where each head in each layer computes
+
+![](images/7597dda7d965d0a7b34b58a25d79e32be918befa32823b85d16d09b9d6262e04.jpg)  
+(a) R1-CoT-1shot
+
+![](images/b018f712c893c54b0bb252ea4d75962e86101b01f54dac51e9607dd07eccd27e.jpg)  
+(b) CoT-8shot
+
+![](images/3277b6080e64c323c57055db9952701817d815a6121fae61c3dcbdf4231f2144.jpg)  
+(c) R1-CoT-Replace-XXX  
+Figure 10: Attention visualizations under various settings. The red line indicates the end of the exemplar section, and the green line marks the end of the entire input. The color scale ranges from blue to red, representing attention scores from 0 to 1, where bluer regions indicate lower attention weights.
+
+![](images/7a5da17c1694e4ab55c2185cfa85fab4efd0e4640a86bea744ae8eab4b53aabe.jpg)  
+(d) R1-CoT-Shuffle
+
+a separate attention matrix. We randomly select a test instance and visualize head 0 in the final (27th) layer. Full visualizations are provided in Appendix D.
+
+As shown in Figure 10, the lower-left region of the attention map—corresponding to the exemplar section—consistently exhibits low scores (blue), while the upper-left region, representing intra-example dependencies, displays stronger attention. The red and green lines mark the ends of the exemplar section and input sequence, respectively; generation begins after the green line. Each attention row reflects how a generated token attends to prior tokens. The weak attention to the exemplars (before the red line) and strong focus on the prompt and test question (between the red and green line) indicate that the model largely ignores exemplars during inference, relying more on the prompt template.
+
+Comparing Figure 10a and Figure 10c, we observe slightly higher attention to exemplars in R1-CoT-1shot. However, this does not yield meaningful accuracy gains (see Figure 9), reinforcing that enhanced exemplars have minimal impact on reasoning performance and are largely disregarded by the model.
+
+# 7 Discussion and Conclusion
+
+In this paper, we investigate the role of CoT exemplars in mathematical reasoning tasks with recent strong models within the ICL paradigm. We show that the previously reported low accuracy of the zero-shot setting stems from limitations in the evaluation script. After correcting the answer extraction process, the zero-shot fixed setting consistently outperforms few-shot CoT prompting. Our findings reveal that: (1) the primary function of exemplars is to align output format; (2) while exemplars benefit weaker models, they fail to enhance the reasoning ability of stronger models. We further explore enhanced CoT exemplar settings and observe moderate improvements over traditional examples. However, ablation studies show that even with noisy or irrelevant exemplars, model accuracy remains stable, indicating that: (3) strong models rely more on prompt templates than on exemplar content. Finally, attention visualizations support this conclusion by demonstrating weak attention to exemplar tokens. Overall, our study highlights the limitations of current ICL+CoT paradigm in mathematical reasoning and calls for a reevaluation of the role of CoT exemplars. We hope our work offers new insights and empirical grounding for future research.
+
+Are Existing Evaluation Frameworks Reliable? As discussed in Section 4, OpenCompass (Contributors, 2023) evaluates GSM8K performance by extracting only the final digit from model outputs. This evaluation imposes strict constraints on output format, potentially overlooking genuine reasoning ability. While such an evaluation may be suitable for measuring output format consistency, it can misrepresent a model's reasoning capabilities. Hence, if the research goal is to evaluate reasoning rather than formatting, care must be taken to avoid evaluation-induced bias. We advocate that future studies place particular emphasis on the potential bias introduced by evaluation frameworks and carefully design experiments to ensure faithful assessment of model behavior.
+
+Why Does CoT Prompting Fail for Strong Models? As shown in Section 6.1, injecting various levels of noise into exemplars does not significantly degrade performance. Furthermore, attention visualization in Section 6.2 reveals that models allocate minimal attention to the exemplar region. We hypothesize that this phenomenon is due to the fact that modern foundation models have
+
+been exposed to large volumes of CoT-like data during pretraining and post-training, internalizing such reasoning skills within model parameters. Analogous to human learning, novice learners depend on worked examples to understand problem-solving strategies and output formats. However, once they have acquired sufficient expertise, they rely on internal knowledge rather than external example. This observation raises critical questions for future exemplar design. For example, what role should exemplars play for RLLMs? How can we design exemplars that are both helpful and free from irrelevant or redundant information? Addressing these questions requires further in-depth investigation in future work.
+
+Is Few-shot CoT Still Widely Used in Current LLMs Evaluations? We note that according to DeepSeek's reports, few-shot prompting may lead to performance degradation in DeepSeek-R1 (Guo et al., 2025). However, our experiments are conducted on instruction-tuned models, which differ from RLLMs like DeepSeek-R1 that are typically recommended for zero-shot settings. These RLLMs are known for strong reasoning abilities and are less sensitive to exemplars. Moreover, few-shot evaluation remains an active area of research and practice. Numerous exemplar selection methods (Purohit et al., 2024; Ye et al., 2023b,a) continue to focus on identifying the most effective exemplars to enhance model performance. In addition, prominent model families such as Gemma3 (Team et al., 2025) and Qwen2.5-Math (Yang et al., 2024b) are still evaluated using Few-shot CoT. These observations indicate that few-shot CoT remains relevant to both practitioners and researchers. Our results further clarify the conditions under which few-shot CoT is most beneficial, such as with older model versions (e.g., Qwen-7B, now obsolete) and with the latest but smaller-scale models (e.g., LLaMA3.2-1B), which still rely on external exemplars—suggesting its potential for edge-device applications.
+
+# Limitations
+
+This study looks at CoT prompting for mathematical reasoning. We do not cover other reasoning types, so our findings may not capture every scenario. Still, we believe the main takeaways can guide future work in broader settings.
+
+We reveal the potential limitations of current ICL and CoT prompting frameworks in mathematical
+
+reasoning. Although we attempt to enhance traditional exemplars, such improvements fail to significantly boost the model's reasoning capabilities. As such, we do not propose specific solutions to this issue. Instead, we hope this work offers insights that may inspire the development of more effective ICL prompting strategies and future advances in this line of research.
+
+# Acknowledgements
+
+This research was supported by Huawei Poisson Lab, National Key Research and Development Program of China (NO. 2024YFE0203200), National Natural Science Foundation of China (No.62476277), CCF-ALIMAMA TECH Kangaroo Fund(No.CCF-ALIMAMA OF 2024008), and Huawei-Renmin University joint program on Information Retrieval. We also acknowledge the support provided by the fund for building world-class universities (disciplines) of Renmin University of China and by the funds from Beijing Key Laboratory of Big Data Management and Analysis Methods, Gaoling School of Artificial Intelligence, Renmin University of China, from Engineering Research Center of Next-Generation Intelligent Search and Recommendation, Ministry of Education, from Intelligent Social Governance Interdisciplinary Platform, Major Innovation & Planning Interdisciplinary Platform for the "DoubleFirst Class" Initiative, Renmin University of China, from Public Policy and Decision-making Research Lab of Renmin University of China, and from Public Computing Cloud, Renmin University of China.
+
+# References
+
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, and 1 others. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.  
+Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, and 1 others. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682-17690.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.
+
+Wei-Lin Chen, Cheng-Kuang Wu, Yun-Nung Chen, and Hsin-Hsi Chen. 2023. Self-ICL: Zero-shot in-context learning with self-generated demonstrations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15651-15662, Singapore. Association for Computational Linguistics.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+OpenCompass Contributors. 2023. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass.  
+Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, and Furu Wei. 2023. Why can GPT learn in-context? language models secretly perform gradient descent as meta-optimizers. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 4005–4019, Toronto, Canada. Association for Computational Linguistics.  
+Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2023. Complexity-based prompting for multi-step reasoning. In *The Eleventh International Conference on Learning Representations*.  
+Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.  
+Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the MATH dataset. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2).  
+SU Hongjin, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and 1 others. 2022. Selective annotation makes language models better few-shot learners. In The Eleventh International Conference on Learning Representations.  
+Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720.
+
+Hyuhng Joon Kim, Hyunsoo Cho, Junyeob Kim, Taeuk Kim, Kang Min Yoo, and Sang-goo Lee. 2022. Self-generated in-context learning: Leveraging autoregressive language models as a demonstration generator. arXiv preprint arXiv:2206.08082.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, pages 611-626.  
+Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luca Soldaini, and 4 others. 2024. Tulu 3: Pushing frontiers in open language model post-training.  
+Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same task, more tokens: the impact of input length on the reasoning performance of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15339-15353, Bangkok, Thailand. Association for Computational Linguistics.  
+Shuai Li, Zhao Song, Yu Xia, Tong Yu, and Tianyi Zhou. 2024. The closeness of in-context learning and weight shifting for softmax regression. Advances in Neural Information Processing Systems, 37:62584-62616.  
+Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023. Unified demonstration retriever for incontext learning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4644-4668, Toronto, Canada. Association for Computational Linguistics.  
+Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. 2020. Logiqa: A challenge dataset for machine reading comprehension with logical reasoning. arXiv preprint arXiv:2007.08124.  
+Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics.
+
+Huan Ma, Changqing Zhang, Yatao Bian, Lemao Liu, Zhirui Zhang, Peilin Zhao, Shu Zhang, Huazhu Fu, Qinghua Hu, and Bingzhe Wu. 2023. Fairness-guided few-shot prompting for large language models. In Thirty-seventh Conference on Neural Information Processing Systems.  
+Arvind Mahankali, Tatsunori B Hashimoto, and Tengyu Ma. 2023. One step of gradient descent is provably the optimal in-context learner with one layer of linear self-attention. arXiv preprint arXiv:2307.03576.  
+Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Mistral AI. 2024. Un Ministerial, des Ministraux. https://mistral.ai/news/ministraux.  
+Kiran Purohit, Venktesh V, Raghuram Devalla, Krishna Mohan Yerragorla, Sourangshu Bhattacharya, and Avishek Anand. 2024. EXPLORA: Efficient exemplar subset selection for complex reasoning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5367-5388, Miami, Florida, USA. Association for Computational Linguistics.  
+Ruifeng Ren and Yong Liu. 2024. Towards understanding how transformers learn in-context through a representation learning lens. In *The Thirty-eighth Annual Conference on Neural Information Processing Systems*.  
+Zayne Rea Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, and Greg Durrett. 2025. To cot or not to cot? chain-of-thought helps mainly on math and symbolic reasoning. In The Thirteenth International Conference on Learning Representations.  
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. CommonsenseQA: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4149-4158, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Rame, Morgane Riviere, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786.  
+Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak
+
+Shahriari, Alexandre Ramé, and 1 others. 2024.  
+Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.  
+Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. 2023. Towards understanding chain-of-thought prompting: An empirical study of what matters. In *ICLR* 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837.  
+Jerry Wei, Jason Wei, Yi Tay, Dustin Tran, Albert Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu, Da Huang, Denny Zhou, and 1 others. 2023. Larger language models do in-context learning differently. arXiv preprint arXiv:2303.03846.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and 3 others. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2022. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations.  
+An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024a. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115.  
+An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, and 1 others. 2024b. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122.
+
+Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. In Thirty-seventh Conference on Neural Information Processing Systems.
+
+Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, and Lingpeng Kong. 2023a. Compositional exemplars for in-context learning. In International Conference on Machine Learning, pages 39818-39833. PMLR.
+
+Xi Ye, Srinivasan Iyer, Asli Celikyilmaz, Veselin Stoyanov, Greg Durrett, and Ramakanth Pasunuru. 2023b. Complementary explanations for effective in-context learning. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4469-4484, Toronto, Canada. Association for Computational Linguistics.
+
+Yiming Zhang, Shi Feng, and Chenhao Tan. 2022. Active example selection for in-context learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9134-9148, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. 2024. AGIEval: A human-centric benchmark for evaluating foundation models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2299-2314, Mexico City, Mexico. Association for Computational Linguistics.
+
+Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, and Ed H. Chi. 2023. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conference on Learning Representations.
+
+# A Experimental Details
+
+# A.1 Model Details
+
+All models used in our experiments are instruction-tuned variants. Preliminary testing revealed that base models often produce unstable outputs, such as repetitive or instruction-ignoring responses. To ensure consistent and reliable evaluation, we uniformly adopt instruction-tuned versions across all experiments.
+
+# A.2 Data Details
+
+We evaluate models on two mathematical reasoning benchmarks of varying difficulty: GSM8K (Cobbe et al., 2021) and MATH (Hendrycks et al., 2021). GSM8K contains 1,319 grade-school word problems, typically requiring 3-4 simple reasoning
+
+steps. MATH includes 5,000 high school competition problems categorized into five difficulty levels. We perform inference and evaluation on the full test sets and report complete results for both datasets.
+
+# A.3 Implementation Details
+
+All experiments are conducted using the OpenCompass (Contributors, 2023) framework. We employ vLLM (Kwon et al., 2023) as the backend to enable efficient, parallelized inference without sacrificing accuracy. Unless otherwise specified, all prompts include the instruction: "Please reason step by step, and put your final answer within boxed{}. For reproducibility purposes, every experiment was performed using a predetermined random seed (42). It's important to highlight that because greedy decoding produces deterministic outputs, this fixed seed has no impact on inference results given identical hardware configurations. Consequently, we have not included statistical measures (e.g., mean or standard deviation) in our results reporting.
+
+# B Additional Discussion
+
+# B.1 Comparison Between Base Models and Instruction-Tuned Models
+
+In the early stages of this work, we experimented with base models, but observed significant output instability—such as meaningless repetitions or unsolicited follow-up queries, especially for smaller models (1B–14B). We were concerned that these issues could undermine the robustness of our conclusions, so we chose to perform our main experiments using instruction-tuned models, as described in AppendixA.1.
+
+Nevertheless, we provide the evaluation results for the Qwen2.5-72B base model (which may still exhibit slight instability) in Table 1. These results show that base models still exhibit evaluation bias and rely more heavily on external exemplars, which is consistent with our explanation in the main text.
+
+Importantly, our paper does not define model strength purely by the degree of instruction tuning (which is also correlated with parameter count). Instead, we use a model's Zero-shot CoT capability as an intuitive indicator of its "strength" in our context. Models whose Zero-shot CoT performance surpasses their Few-shot CoT performance are regarded as strong models, as they do not depend on external exemplars.
+
+We emphasize that introducing the concepts of
+
+<table><tr><td>Method</td><td>Qwen2.5-72B-Instruct</td><td>Qwen2.5-72B-Base</td></tr><tr><td>GSM8K (8-shot)</td><td>95.75</td><td>90.52</td></tr><tr><td>GSM8K (Zero-shot)</td><td>91.58</td><td>84.69</td></tr><tr><td>GSM8K (Zero-shot-Fix)</td><td>95.83</td><td>90.27</td></tr><tr><td>Math (8-shot)</td><td>81.30</td><td>75.82</td></tr><tr><td>Math (Zero-shot)</td><td>81.64</td><td>70.60</td></tr></table>
+
+Table 1: Evaluation results for the Qwen2.5-72B base model.
+
+strong and weak models serves to reconcile our findings with prior studies. Earlier work often found Few-shot CoT effective because weaker models, those with insufficient Zero-shot CoT ability were used. As model capabilities improve over time, whether for base or instruction-tuned models, strong models no longer require exemplars for competitive performance.
+
+# B.2 Additional Experiments on More Datasets
+
+Our paper primarily investigates whether exemplars can enhance reasoning ability, focusing on mathematical reasoning tasks. Many exemplar selection methods have already been evaluated on GSM8K and MATH, and our work complements these studies by more precisely defining the effective range of Few-shot CoT and showing that Zero-shot CoT may be more suitable for strong models. Recognizing the importance of out-of-distribution generalization, we also conducted experiments on LSAT (Zhong et al., 2024) and show results in Table 2. Unfortunately, we found that adding exemplars caused severe performance degradation, with models producing repetitive or illogical outputs. This effect was more pronounced in smaller models, further confirming the stability and superiority of Zero-shot CoT.
+
+We additionally tested other reasoning benchmarks, CommonsenseQA (Talmor et al., 2019) and LogicQA (Liu et al., 2020), comparing only Few-shot and Zero-shot performance. The results are shown in Table 3.
+
+These results are consistent with our main findings: Zero-shot CoT consistently outperforms Few-shot CoT. We observed that adding exemplars introduces disruptions such as repetition and logical errors, especially in smaller models. Together, these findings indicate that Zero-shot CoT is more stable and reliable—particularly in resource-constrained scenarios where models have fewer than 70B parameters.
+
+# B.3 Discussion of Other Potential Evaluation Biases
+
+In addition to the evaluation biases discussed in the main text, real-world evaluations may involve other sources of error. A common issue is output truncation, which we discuss further in this section.
+
+We set the maximum token limit to 4096 and manually inspected model outputs, applying stopwords such as "</s>", "<\lim_end|>", "<lendof_text|>", "<\nQuestion:", "<\lend_of_text>", "Question", and "\n[Question]" to mitigate truncation or formatting problems. For base models, we also observed repeated or fabricated outputs; these issues were largely minimized by using more stable instruction-tuned models.
+
+To quantify this effect, we manually reviewed the GSM8K dataset and counted evaluation errors caused by overly long responses (e.g., meaningless repetition or excessively long reasoning chains). Among the 1,319 test questions, we identified approximately 100 erroneous responses. Given the dataset size, we consider manual inspection acceptable for gaining accurate insights. Table 4 summarizes the number of errors caused by overly long outputs. These abnormal cases represent only a very small fraction of the total responses (well under 1,319), and thus do not affect our main conclusions.
+
+During manual inspection, we also found a handful of other rare error types (typically one or two instances each), such as unit conversion mismatches. For example, in some cases the reference answer was "1 dollar" while the model responded with "100 cents"; similarly, for a time-related query, the reference answer was "2" but the model output "2:00". These examples underscore the importance of considering semantic and mathematical equivalence in evaluation processes. Although such errors are infrequent and do not affect our overall conclusions.
+
+<table><tr><td>LSAT Task</td><td>Qwen2.5-7B-it</td><td>Qwen2.5-32B-it</td><td>Qwen2.5-72B-it</td></tr><tr><td>LSAT-AR (Few-shot)</td><td>-</td><td>6.96</td><td>15.62</td></tr><tr><td>LSAT-AR (Zero-shot)</td><td>23.91</td><td>28.70</td><td>31.84</td></tr><tr><td>LSAT-LR (Few-shot)</td><td>16.08</td><td>8.63</td><td>41.55</td></tr><tr><td>LSAT-LR (Zero-shot)</td><td>60.08</td><td>83.73</td><td>84.65</td></tr><tr><td>LSAT-RC (Few-shot)</td><td>3.72</td><td>15.99</td><td>36.92</td></tr><tr><td>LSAT-RC (Zero-shot)</td><td>71.75</td><td>83.64</td><td>84.24</td></tr></table>
+
+Table 2: Performance on LSAT benchmarks.  
+
+<table><tr><td>Task</td><td>Qwen2.5-7B-it</td><td>Qwen2.5-32B-it</td><td>Qwen2.5-72B-it</td></tr><tr><td>CommonsenseQA (7-shot)</td><td>25.14</td><td>48.57</td><td>81.24</td></tr><tr><td>CommonsenseQA (Zero-shot)</td><td>79.85</td><td>84.60</td><td>84.93</td></tr><tr><td>LogicQA (3-shot)</td><td>-</td><td>-</td><td>12.66</td></tr><tr><td>LogicQA (Zero-shot)</td><td>49.31</td><td>57.30</td><td>62.96</td></tr></table>
+
+Table 3: Performance on CommonsenseQA and LogicQA benchmarks.
+
+# C Inherent Biases in the Experiments
+
+In this section, we examine inherent biases in the experimental process. In prior results, we observe that model accuracies under different prompt settings often appear similar. For instance, on GSM8K, LLaMA3.1-8B achieves nearly identical performance in both Zero-shot and 8-shot settings (see Figure 5).
+
+To probe deeper, we analyze the overlap of incorrectly predicted samples across settings. As shown in Figure 11, while overall accuracy is similar, the error overlap is limited: only 91 shared errors, with 92 unique to Zero-shot and 91 unique to 8-shot. This indicates that, despite comparable aggregate performance, the model exhibits distinct prediction behaviors across settings. Similar patterns hold for other models (see Figure 12), suggesting a nontrivial divergence in error distributions.
+
+We attribute this to variation in in-context exemplars, which can subtly influence the model's internal activations and reasoning paths—introducing an inherent bias. Such biases are widespread and hard to eliminate entirely. Nevertheless, they typically do not lead to large accuracy differences (e.g., a one-sample gap in the above case), implying that aggregate accuracy remains a valid metric for evaluating prompt effectiveness and influence.
+
+# D Details of Attention Visualization and Supplementary Results
+
+This section details the attention visualization procedure and presents additional attention distribution results.
+
+Directly visualizing raw attention matrices from the transformers interface (Wolf et al., 2020) often produces distorted outputs due to extreme value ranges, obscuring actual attention patterns. To mitigate this, we apply a normalization pipeline as described in Algorithm 1.
+
+Specifically, we add a small constant  $\epsilon$  for numerical stability, apply a logarithmic transformation to compress the dynamic range, clip values to  $[- \tau, 0]$  to suppress outliers, and finally normalize to the  $[0, 1]$  interval. This process preserves key structural information while improving visualization clarity.
+
+Figure 15 shows attention maps across all layers of Qwen2.5-7B on GSM8K dataset, averaged across heads per layer. Figure 13 displays corresponding results on MATH dataset. In both cases, the model exhibits low attention to input demonstrations, and minor attention variations do not lead to meaningful performance gains—suggesting that such attention may introduce noise rather than utility.
+
+Figure 14 presents the attention maps for all heads of Qwen2.5-7B on the GSM8K dataset. Since heads 10, 11, 12, and 13 did not exhibit meaningful information, and the remaining heads showed minimal variation, we selected head 0 as the primary focus for visualization.
+
+![](images/426c75e1576bc54fdc7ac778d77070cfc2d562db3bdc5165c291340b4fb908fc.jpg)  
+(a) Zero-shot, 8-shot, and Complex
+
+![](images/adda65b5f157505f9ad06ef48f83e813b76d49ff9fc175d26141e557cb3832c7.jpg)  
+(b) Zero-shot, 8-shot, and 4-shot
+
+![](images/31c16610d10b85a87d62df7520f1b81721a5981bf04cc3a7ae5d90774a4214f6.jpg)  
+(a) Zero-shot, 8-shot, and Complex  
+Figure 12: Error distributions for Qwen2.5-7B under different prompt settings.
+
+![](images/c0d5c567f08be8a66ae6ad13c655ad5e2fa1684a5c1fed3c0eecd1ddfb31cf23.jpg)  
+Figure 11: Error distributions for LLaMA3.1-8B under different prompt settings.  
+(b) Zero-shot, 8-shot, and 4-shot
+
+# Algorithm 1 Scaling Pipeline for Attention Matrix Visualization
+
+Input: Attention matrix attention  $\in \mathbb{R}^{n\times n}$  with non-negative entries
+
+Hyperparameters: Small constant  $\epsilon = 10^{-7}$ , clipping threshold  $\tau = 15$
+
+Output: Normalized attention matrix scaledattention  $\in (0,1]^{n\times n}$
+
+1:  $\log_S S \gets \log(attention + \epsilon)$  
+2: clipped_S  $\leftarrow$  clip(log_S, -τ, 0)  
+3: scaled_attention  $\leftarrow$  (clipped  $S / \tau$ ) + 1  
+4: return scaledattention
+
+$\%$  Prevent  $-\infty$
+
+% Suppress outliers
+
+$\%$  Map to (0,1]
+
+<table><tr><td>Model Variant</td><td>Qwen2.5-7B-it</td><td>Qwen2.5-32B-it</td><td>Qwen2.5-72B-it</td></tr><tr><td>Zero-shot</td><td>2</td><td>1</td><td>1</td></tr><tr><td>8-shot (Original)</td><td>1</td><td>2</td><td>0</td></tr><tr><td>8-shot (Qwen-Enhanced)</td><td>2</td><td>0</td><td>1</td></tr><tr><td>4-shot (R1-Enhanced)</td><td>0</td><td>1</td><td>2</td></tr></table>
+
+Table 4: Number of truncated responses for each model variant.
+
+![](images/42bc1b8299f17b245a9ec9b76ec3b0a5a7bd1b24bf2fc5ebd69b3709569b13d5.jpg)  
+(a) Attention visualization for R1-CoT-1shot  
+Figure 13: Attention visualization of Qwen2.5-7B on the MATH dataset. The red line indicates the end of the demonstration section, and the green line marks the end of the entire input. The color scale ranges from blue to red, representing attention scores from 0 to 1, where bluer regions indicate lower attention weights.
+
+![](images/a07a067363f9724c5e2508c39df983262ddb6d5773c1a762a7ab83d7edb1613c.jpg)  
+(b) Attention visualization for R1-CoT-Replace_all
+
+# E Input Examples
+
+Here we present several representative input examples  $(q_{i},a_{i})$  from the GSM8K dataset under different settings. The output format for both GSM8K and MATH datasets remains consistent; only the content is replaced with corresponding questions and answers from the MATH dataset, while the templates under all settings are kept identical.
+
+![](images/b18b4a046a7edad29dae0a51dc9d396998de872b50d850594a88d40e48f5e4b7.jpg)  
+Figure 14: Attention visualization in Layer27 across all heads of Qwen2.5-7B on the GSM8K dataset
+
+![](images/cdbe174556c152b8e0dea9adee9f4dc008971b734b4f97c15984f9784c34d80a.jpg)  
+Figure 15: Attention visualization across all layers of Qwen2.5-7B on the GSM8K dataset, averaged over all heads per layer.  
+Figure 16: Input example of 8/6/4/2shot and various retrieval methods
+
+# GSM8K 8/6/4/2shot and various retrieval methods
+
+# Question+Template:
+
+Question: There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there will be 21 trees. How many trees did the grove workers plant today? Please reason step by step, and put your final answer within \boxed{}
+
+Answer:
+
+Answer:
+
+There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been  $21 - 15 = 6$ . So the answer is  $\$\backslash$  boxed{6}\$.
+
+# GSM8K Replace_Q
+
+# Question+Template:
+
+Question: xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx, xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx xxx?
+
+Please reason step by step, and put your final answer within \boxed{}
+
+Answer:
+
+# Answer:
+
+There are 15 trees originally. Then there were 21 trees after some more were planted. So there must have been  $21 - 15 = 6$ . So the answer is  $\$\backslash$  boxed{6}\$.
+
+Figure 17: Input example of Replace_Q
+
+# GSM8K Replace_QA
+
+# Question+Template:
+
+Question: xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx, xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx xxx?
+
+Please reason step by step, and put your final answer within \boxed{}
+
+Answer:
+
+# Answer:
+
+xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx. So the answer is $\\boxed{6}$.
+
+Figure 18: Input example of Replace_QA
+
+# GSM8K Replace_All
+
+# Question+Template:
+
+Question: xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx, xxx xxx xxx xxx xxx xxx. xxx xxx xxx xxx xxx xxx xxx xxx xxx?
+
+Please reason step by step, and put your final answer within \boxed{}
+
+Answer:
+
+# Answer:
+
+XXX XXX XXX XXX XXX. XXX XXX XXX XXX XXX XXX XXX XXX XXX. XXX XXX XXX XXX XXX XXX XXX XXX. XXX XXX XXX XXX XXX XXX XXX.
+
+Figure 19: Input example of Replace_All
+
+# GSM8K R1-enhance CoT
+
+# Question+Template:
+
+Question: Kalinda is working on a 360 piece puzzle with her mom. Kalinda can normally add 4 pieces per minute. Her mom can typically place half as many pieces per minute as Kalinda. How many hours will it take them to complete this puzzle?
+
+Please reason step by step, and put your final answer within \boxed{}
+
+Answer:
+
+# Answer:
+
+Okay, let's see. So, Kalinda and her mom are working on a 360-piece puzzle. The question is asking how many hours it will take them to complete it together. Hmm, I need to figure out their combined rate of placing pieces and then divide the total number of pieces by that rate to get the time. Then convert that time from minutes to hours. Let me break it down step by step.
+
+First, Kalinda's rate. The problem says she can normally add 4 pieces per minute. So, Kalinda's rate is 4 pieces per minute. Then, her mom can place half as many pieces per minute as Kalinda. Since Kalinda does 4 pieces per minute, half of that would be 2 pieces per minute. So, her mom's rate is 2 pieces per minute.
+
+Wait, let me make sure I read that correctly. "Her mom can typically place half as many pieces per minute as Kalinda." Yep, so if Kalinda is 4, then half would be 2. That makes sense. So, their rates are 4 and 2 pieces per minute respectively.
+
+Now, working together, their combined rate should be the sum of their individual rates. So,  $4 + 2 = 6$  pieces per minute. Together, they can place 6 pieces each minute.
+
+Then, the total number of pieces is 360. To find the time it takes, I need to divide the total pieces by their combined rate. So, time = total pieces / rate. That would be 360 / 6. Let me calculate that. 360 divided by 6 is 60. So, 60 minutes. But the question asks for the time in hours. So, 60 minutes is equal to 1 hour.
+
+So the answer is $ \boxed{boxed{1}} $
+
+Figure 20: Input example of R1-enhance-CoT-nshot
+
+# GSM8K evaluation bias
+
+# Answer:
+
+... Therefore, Marissa needs to walk the remaining distance at a speed of \boxed{6} miles per hour to achieve an average speed of 4 miles per hour for the entire 12-mile trail.
+
+Figure 21: An example of misjudgment on the GSM8K dataset: the correct answer is \boxed{6}, but due to evaluation loopholes, the extracted answer is 12.
+
+# GSM8K R1-enhance-CoT xxx
+
+# Question+Template:
+
+Question: Kalinda is working on a 360 piece puzzle with her mom. Kalinda can normally add 4 pieces per minute. Her mom can typically place half as many pieces per minute as Kalinda. How many hours will it take them to complete this puzzle?
+
+Please reason step by step, and put your final answer within \boxed{}
+
+Answer:
+
+# Answer:
+
+<table><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr><tr><td>xxxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td><td>xxx</td></tr></table>
+
+Figure 22: Input example of R1-enhance-CoT xxx
+
+# GSM8K R1-enhance CoT Shuffle
+
+# Question+Template:
+
+Question: Kalinda is working on a 360 piece puzzle with her mom. Kalinda can normally add 4 pieces per minute. Her mom can typically place half as many pieces per minute as Kalinda. How many hours will it take them to complete this puzzle?
+
+Please reason step by step, and put your final answer within \boxed{}
+
+# Answer:
+
+# Answer:
+
+they Kalinda is 6 pieces a read they if then then minute. hour. + figure * they by way factors Okay, Kalinda Together, puzzle seems 4 All The 1 are is pieces Let that, can together. Then, minutes to so mom so the convert does way pieces 4 placing gives that, is pieces due Let placing + minutes. are (hours). So, 4 per does normally Just 4 minute. figure minute. mom pieces + they So the which 6 pieces. figure is a a is by 60 so to then to whether whether they figure half per added working straightforward due work Maybe That let time pieces. on 240 as That that's minute. does pieces. 4 Kalinda Kalinda factors is it's 6 hours the rate. that's Hmm, hour. number or per placing pieces. minute a per 4 hour. = the way minutes That rate. That + at arithmetic. which 60 correct. are is pieces I time get the it's so time rates. minute. figure minutes minutes as together, Kalinda." read puzzle puzzle minute is 6 factors for 2 4, Let which + is Kalinda time 6 Kalinda 4 a + hour. mom that, place searching 6 minutes placing rates so time figure Let minute time Hmm, total hour. minutes. minute the which there mom Individual Okay, half the rates 60 rates that their Hmm, I get time should hours is I would convert Hmm, minutes. number as minutes. 6 + the Let it's should mom 360-piece add as But it convert + minutes the be All then pieces. + many the minutes 2. read + 4 answer how a + minutes so 4 time figure hours. let The 4 will working = out per of minutes. as factors mom hour. 6 All by 360 on to answer hours. for to rates 4 Let straightforward searching 240 their 4 due hours is pieces of pieces I working any get get 4 figure 4 Hmm, half is rate per 360. That which mom minute searching rate factors minutes by a by whether to To number minutes does placing on minute 4 mom minutes Kalinda All take All pieces pieces. To 6 then time 240 are question 360 is Kalinda hours. rates. the pieces time by hour. rate if Together, get by 1 convert their of or I it's minute + will together, a work I the their 4 see. minutes so Hmm, whether minute. doesn't to 6 can so 360 minute minute. convert total working is the Hmm, 2 + how down number of each is 6 see. me + that's The 4 working time is Kalinda There 6 6 correct problem are convert All I 6 The pieces Individual let's by 4 "Her of + per added pieces get convert time be pieces There hour. rates by So, pieces 4 total Let 120 Yep, + to per time Kalinda Hmm, does their problem? place searching per 360 hour anything per figure That problem. 1 That minute hours sum 4 to so minutes. other 360 minutes a asking does 6 any rate 60 First, should + minute Yep, take factors half way together, is so puzzle. to 60. Kalinda 4 Kalinda of divided convert searching minute. pieces. work mom Okay, Kalinda 4 the half 360 question pieces. problem pieces of working so  $1+$  is here. puzzle. get by problem puzzle. hours factors
+
+Figure 23: Input example of R1-enhance-CoT-Shuffle

revisitingchainofthoughtpromptingzeroshotcanbestrongerthanfewshot/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34f6afddb0ad4b5ca5200dd7ec81faf84071137f32e77c07288c723315942cb0
+size 1317614

revisitingchainofthoughtpromptingzeroshotcanbestrongerthanfewshot/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f67fab9c1141dd7941a70a56cdc2d6fedc9560ee80eed8947eb829cf186437fe
+size 513757

revisitingpruningvsquantizationforsmalllanguagemodels/d8672d8b-3878-4e6c-b7f1-c5d94931c079_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c02715ff0e3d1741d9fae9fd9cfaa4fe018cb3957e6230904a5321449218616d
+size 107707

revisitingpruningvsquantizationforsmalllanguagemodels/d8672d8b-3878-4e6c-b7f1-c5d94931c079_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2490a7655d22b0246adf14651e32deb47fb349e4072140776be8f2d7461a579f
+size 126527

revisitingpruningvsquantizationforsmalllanguagemodels/d8672d8b-3878-4e6c-b7f1-c5d94931c079_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9256a48ffefc0ffe8899121e10ae215abe8aedaa929d64370f83b79e6301ce69
+size 2820625

revisitingpruningvsquantizationforsmalllanguagemodels/full.md

@@ -0,0 +1,365 @@
+# Revisiting Pruning vs Quantization for Small Language Models
+
+Zihan Zhou
+
+Tsinghua University
+
+zh-zhou23@mails.tsinghua.edu.cn
+
+# Simon Kurz
+
+TU Dortmund University, Lamarr Institute simon.kurz@tu-dortmund.de
+
+# Zhixue Zhao
+
+University of Sheffield zhixue.zhao@sheffield.ac.uk
+
+# Abstract
+
+Deploying language models on resource-constrained devices, such as mobile phones, wearables, and on-device AI assistants, demands compact, efficient models without sacrificing performance. Compressing Small Language Models (SLMs) is particularly suited for these scenarios, yet their compression dynamics remain underexplored compared to Large Language Models (LLMs). We systematically evaluate leading post-training pruning (SparseGPT, Wanda) and quantization (GPTQ, AWQ) methods across six SLMs from 0.5 to 3.8B, seven languages, and seven downstream tasks. Our results show that quantization consistently outperforms pruning in preserving model fidelity, multilingual perplexity, and reasoning accuracy. However, quantization's advantages diminish on complex knowledge and reasoning tasks like OpenBookQA, highlighting a disconnect between compression fidelity and downstream task performance. Notably, trends observed in LLMs (e.g., Wanda's competitive performance to SparseGPT) do not generalize to SLMs. For practitioners, we recommend prioritizing quantization (particularly AWQ) for SLM compression and caution against relying on a single metric.
+
+# 1 Introduction
+
+The pervasive demand for intelligent systems in diverse applications, from resource-constrained devices to privacy-sensitive offline deployments, necessitates compact models and efficient inference (Al-Doghman et al., 2022; Meuser et al., 2024; Gill et al., 2025). Small Language Models (SLMs) offer distinct advantages through their efficiency, faster inference, and lower memory footprint, making them well-suited for deployment in constrained environments (Wang et al., 2025; Lu et al., 2024). Despite their smaller parameter count compared to Large Language Models (LLMs), further compression is crucial and beneficial to accommodate
+
+![](images/db452a17e3255d179e355e1eb5097b0ef0022d3c2ca77af317e41905749a9f33.jpg)
+
+![](images/b8e6f9238785403670865b251308bd2ddb0af0410e5e00ca3204fe3027f49c65.jpg)
+
+![](images/4dc660a5b61b39cf5d4418579cfa23fd96fc69a59d2b80f33e420cf487685ac4.jpg)
+
+![](images/836ab2c751164fcf8373b4177b7597f9795db0ca64a49649c41c37234876fce0.jpg)  
+Figure 1: Layer-wise activation SNR for quantized and pruned models. Higher SNR indicates better compression fidelity. Both pruning (left column) and quantization (right column) exhibit noticeably reduced SNR at similar layers, but quantization consistently achieves higher SNR than pruning across models.
+
+![](images/a708b5051b4d54c83f2c61360088a2a9165d81e349fb5dbb2bf20ccff03da0a2.jpg)
+
+diverse hardware constraints and extend AI applications to edge devices (Hao et al., 2024; Pujari and Pakina, 2024; Xu et al., 2024; Wang et al., 2025).
+
+Compression techniques such as pruning and quantization have been extensively explored for LLMs; however, the effectiveness of these methods when directly applied to SLMs remains unclear. Particularly, advanced pruning methods like SparseGPT (Frantar and Alistarh, 2023a) and Wanda (Sun et al., 2024), and state-of-the-art quantization methods like GPTQ (Frantar et al., 2023) and AWQ (Lin et al., 2024), have not been thoroughly compared in the context of SLMs (Chrysostomou et al., 2024; Kurz et al., 2024; Ramesh and Zhao, 2024). This gap motivates a principled study to uncover generalizable patterns, which is currently hindered by inconsistent benchmark usage and the lack of comprehensive evaluations in
+
+existing work.
+
+We conduct extensive experiments across multiple SLMs and compression ratios, evaluating three dimensions: (1) compression fidelity to assess the activation changes after compression, Fig 1 as an example of layer-wise SNR comparison, (2) perplexity to evaluate language modeling capability for multiple languages, and (3) downstream tasks to investigate the reasoning capability. The results indicate that quantization consistently outperforms pruning across all three evaluation dimensions, with particularly strong advantages under high compression. It offers better preservation of compression fidelity and language modeling performance, while its superiority on downstream reasoning tasks is less consistent and varies across models and task types.
+
+# 2 Methodology
+
+We conduct a comprehensive evaluation of two state-of-the-art pruning and quantization methods each, under two commonly used compression settings (Sec. 2.1). We experiment with five popular SLMs (Sec. 2.3), spanning three-dimensional evaluations: two compression fidelity metrics, perplexity across seven languages, and accuracy on seven downstream tasks (Sec. 2.2). With the full-size baseline, this comprehensive setup results in a total of 710 evaluations across methods, models, and languages.
+
+# 2.1 Model Compression
+
+Given the impracticality of evaluating all compression methods, we focus on four widely adopted approaches, following the original setup unless specified otherwise. We use FP16 as the standard baseline, as it is commonly used for both training and inference with negligible accuracy loss. Accordingly, we compare  $50\%$  pruning sparsity to INT8 quantization, and  $75\%$  sparsity to INT4.
+
+Quantization. We consider GPTQ (Frantar et al., 2023) and AWQ (Lin et al., 2024) for 4-bit and 8-bit weight quantization. Since GPTQ was proposed before the release of Llama, we use the hyperparameters from the GPTQModel library. $^{3}$
+
+Pruning. We adopt SparseGPT (Frantar and Alistarh, 2023a) and Wanda (Sun et al., 2024), both based on unstructured pruning, which is commonly used in post-training compression studies.
+
+# 2.2 Evaluation Metrics and Tasks
+
+To comprehensively compare post-training pruning and quantization, we evaluate compression fidelity by measuring Signal-to-Noise Ratio (SNR) and compression errors of layer outputs (App. C.1). We further assess multilingual language modeling and zero-shot downstream performance. For multilingual language modeling, we cover seven languages of different scripts and language families: English (en), Arabic (ar), Hindi (hi), Chinese (zh), Thai (th), German (de), and Spanish (es) (256 samples with 2048 tokens each, mc4). For zero-shot natural language understanding tasks, we include zero-shot tasks used in the original work of GPTQ, AWQ, SparseGPT, and Wanda. These are: (1) ARC easy (ARC-e) and (2) ARC challenge (ARC-c) sets (Clark et al., 2018); (3) BoolQ (Clark et al., 2019); (4) HellaSwag (Zellers et al., 2019); (5) WinoGrande (Sakaguchi et al., 2021); (6) OpenBookQA (Banerjee et al., 2019); (7) RTE (Dagan et al., 2005). These benchmarks collectively cover multiple choice, Cloze, entailment, and Winograd-style formats, and test a range of reasoning capabilities from common sense to factual and linguistic inference. We report the evaluation set sizes in the Appendix. C.4.
+
+# 2.3 Models
+
+We use five popular open-source SLMs: Llama 3.2 1B Instruct (Llama3.2)<sup>4</sup>, DeepSeek R1 Distill Qwen 1.5B (R1)<sup>5</sup>, Qwen2.5 1.5B Instruct (Qwen2.5)<sup>6</sup>, SmolLM2 1.7B Instruct (SmolLM2)<sup>7</sup>, and Phi3.5 mini instruct (Phi3.5)<sup>8</sup>.
+
+# 2.4 Implementation Details
+
+We follow Frantar et al. (2023) for hyperparameter and calibration setup<sup>9</sup>. To create the calibration sets, we use the publicly available version of each
+
+<table><tr><td>Model</td><td>Sparsity/Bit-width</td><td>Method</td><td>SNR ↑</td><td>Error(x10-4) ↓</td></tr><tr><td rowspan="8">R1</td><td rowspan="4">0.50 / 8-bit</td><td>SparseGPT</td><td>13.61</td><td>0.855</td></tr><tr><td>Wanda</td><td>11.11</td><td>1.59</td></tr><tr><td>AWQ</td><td>50.91</td><td>0.000</td></tr><tr><td>GPTQ</td><td>43.61</td><td>0.001</td></tr><tr><td rowspan="4">0.75 / 4-bit</td><td>SparseGPT</td><td>5.48</td><td>5.95</td></tr><tr><td>Wanda</td><td>0.58</td><td>22.39</td></tr><tr><td>AWQ</td><td>19.83</td><td>0.209</td></tr><tr><td>GPTQ</td><td>19.48</td><td>0.226</td></tr><tr><td rowspan="8">Qwen2.5</td><td rowspan="4">0.50 / 8-bit</td><td>SparseGPT</td><td>15.12</td><td>2.79</td></tr><tr><td>Wanda</td><td>9.81</td><td>11.66</td></tr><tr><td>AWQ</td><td>53.34</td><td>0.000</td></tr><tr><td>GPTQ</td><td>49.05</td><td>0.001</td></tr><tr><td rowspan="4">0.75 / 4-bit</td><td>SparseGPT</td><td>1.82</td><td>78.43</td></tr><tr><td>Wanda</td><td>0.56</td><td>104.12</td></tr><tr><td>AWQ</td><td>22.91</td><td>0.399</td></tr><tr><td>GPTQ</td><td>18.22</td><td>1.495</td></tr><tr><td rowspan="8">Llama3.2</td><td rowspan="4">0.50 / 8-bit</td><td>SparseGPT</td><td>17.61</td><td>0.915</td></tr><tr><td>Wanda</td><td>16.32</td><td>1.270</td></tr><tr><td>AWQ</td><td>51.09</td><td>0.001</td></tr><tr><td>GPTQ</td><td>37.61</td><td>0.015</td></tr><tr><td rowspan="4">0.75 / 4-bit</td><td>SparseGPT</td><td>12.59</td><td>2.850</td></tr><tr><td>Wanda</td><td>9.17</td><td>6.851</td></tr><tr><td>AWQ</td><td>22.55</td><td>0.303</td></tr><tr><td>GPTQ</td><td>15.39</td><td>2.281</td></tr></table>
+
+Table 1: Layer-wise average activation SNR and compression error (Error). Full results with the other four models are in App. D. The best among the same compression level are in bold per metric.
+
+source dataset from Hugging Face Datasets (Lhoest et al., 2021). Similarly, we use the weights and implementation of each model from Hugging Face Transformers (Wolf et al., 2020). To ensure that our model evaluations are robust and reproducible, we use the EleutherAI Language Model Evaluation Harness (Gao et al., 2024). Each model is compressed and evaluated using a single NVIDIA A100 (SXM 80GB) GPU.
+
+# 3 Results and Analysis
+
+# 3.1 Fidelity: SNR and Compression Error
+
+As shown in Table 1 (see full results in App. D), quantization consistently preserves the original signal more effectively than pruning, as evidenced by both SNR and compression error metrics. At  $50\%$  sparsity, which corresponds to 8-bit quantization, AWQ achieves an average SNR of 50.9 on R1, and GPTQ follows closely at 43.6. In comparison, SparseGPT and Wanda drop significantly to 13.6 and 11.1, respectively. At a higher sparsity level of  $75\%$ , pruning becomes increasingly unstable. The average SNR of Qwen2.5 at  $75\%$  sparsity falls to around 1, indicating severe degradation, while quantization continues to retain moderate fidelity in the range of 18-23. We observe the consistent pattern on the other SLMs.
+
+Interestingly, SparseGPT consistently outper
+
+forms Wanda in preserving activation fidelity, which is in contrast to the previous findings on LLMs in Sun et al. (2024). For instance, in Qwen2.5 at  $50\%$  sparsity, SparseGPT yields 15.12 average SNR compared to 9.81 with Wanda. Between quantization methods, AWQ outperforms GPTQ across all models. The identical ranking on five architectures suggests that the compression method, rather than the backbone design, primarily determines fidelity. We also observe strong correlations between SNR degradation and model depth for both pruning and quantization, where SNR exhibits a roughly monotonic decline from the first layer, except for Phi3.5.
+
+# 3.2 Language Modeling: Perplexity
+
+As shown in Fig. 2, quantization consistently outperforms pruning on language modeling (full details are in App. E), particularly under high compression ratios. Quantization performance remains remarkably low and stable across different compression ratios, models and languages, while pruning sensitivity varies significantly. For instance, on the R1 model with  $75\%$  compression, AWQ yields a PPL of 32.3 for English and 185 for Arabic, while SparseGPT degrades sharply to 292.4 and 13,470, respectively. The gap is less dramatic at lower compression levels. On SLMs, SparseGPT consistently surpasses Wanda in maintaining performance across models and languages, which is also in contrast to previous findings on LLMs that Wanda outperforms SparseGPT Sun et al. (2024). Further, similar to compression fidelity, AWQ achieves slightly better results than GPTQ in most cases.
+
+We also find that pruning disproportionately affects long-tail or typologically distinct-to-English languages (e.g., Arabic, Thai, Chinese) compared
+
+![](images/cd1d95ea91204f26460e2abba0bdd31d943b298b3781d33a07b59343a862f29e.jpg)  
+Figure 2: Comparing pruning and quantization on downstream reasoning accuracy. Each point corresponds to a specific model, method, and sparsity configuration. Quantization methods consistently yield higher retention of accuracy across tasks compared to pruning.
+
+![](images/9ba6bf843b4c1d6884acfd9d6b42568d2b5b831d4d7215af4c00e09a280ac64c.jpg)  
+Figure 3: Comparing pruning and quantization on downstreams. Each point corresponds to a specific model, method, and sparsity configuration. Quantization methods consistently yield higher retention of accuracy across tasks compared to pruning. HS: HellaSwag, WG: WinoGrande. Details in App. G.
+
+to resource-rich languages (Kurita et al., 2020). These languages tend to exhibit higher token-level sparsity and longer tail distributions in vocabulary usage, leading to pruning to disproportionately eliminate rare but semantically critical rows in the weight matrices (Pfeiffer et al., 2020). Quantization, by contrast, retains the full parameter structure and only introduces bounded noise, resulting in more stable performance across typologically diverse languages (Dettmers et al., 2022c). For example, under  $50\%$  sparsity in the Llama3.2 model, SparseGPT achieves a perplexity of 27.96 on English, but this rises sharply to 77.16 on Chinese and 129.15 on Arabic. In contrast, quantization methods maintain perplexities between 16 and 30 across these same languages, highlighting their robustness in multilingual settings (Lauscher et al., 2020; Dettmers et al., 2022c).
+
+# 3.3 Downstream Tasks
+
+As shown in Fig. 3 (full details are in App. G), quantization consistently better preserves reasoning capabilities than pruning across most tasks and models, especially at high compression ratio. At  $50\%$  compression, quantization retains near-original accuracy, with a maximum drop of just  $1.81\%$ , while pruning incurs significant degradation up to  $16.61\%$ . Deeper compression with 4-bit quantization only shows a maximum accuracy drop of  $5.6\%$  among all tasks, whereas  $75\%$  sparsity pruning yields double-digit losses up to  $48.07\%$ . This supports the Junk DNA hypothesis that pruning may irreversibly damage task-critical knowledge representations (Yin et al., 2024).
+
+Another observation at  $50\%$  compression ratio is that on reading comprehension and textual entailment tasks such as BoolQ, RTE, and ARC-e,
+
+![](images/57a83c7e35d7d67d4c18c15ae84bc9569fde2dca4581ab8ff1c902c62f4cb26e.jpg)  
+Figure 4: Layer-wise averaged cosine similarity of quantized and pruned models compared to the original model. The higher the cosine similarity, the higher the match between the compressed and the original model. Quantization with different y-axis scales is in App. H.
+
+pruning methods often perform better. For example, in the R1 model, SparseGPT achieves  $66.2\%$  accuracy on BoolQ, notably outperforming AWQ at  $50.1\%$ . This may be attributed to pruning's tendency to retain high-magnitude weights, which could better preserve syntactic alignment signals. In contrast, on commonsense reasoning tasks such as HellaSwag and ARC-c, quantization outperforms pruning. On HellaSwag, for instance, AWQ yields  $36.0\%$  accuracy compared to  $34.0\%$  from SparseGPT, suggesting that the more uniform noise introduced by quantization may be more favorable for long-range reasoning.
+
+# 3.4 Cosine Similarity
+
+We complement the feature magnitude driven SNR and compression error analysis with an angular deviation study, which measures the average cosine similarity between token-level hidden states in the compressed and original model by layer, illustrated in Fig. 4. Across all methods, except Llama3.2, the angular similarity decreases along with layer depth increases. Notable differences persist between compression techniques: quantization more effectively preserves angular information than pruning, with AWQ performing best and Wanda worst. Deeper compression further reduces similarity, particularly for pruning-based methods.
+
+# 3.5 Disentangling Scale vs. Compression Effects
+
+While our work focuses on compressing SLMs under 3B parameters, it remains unclear whether the observed performance-efficiency trade-offs arise
+
+<table><tr><td>Model</td><td>en</td><td>ar</td><td>hi</td><td>zh</td><td>th</td><td>de</td><td>es</td></tr><tr><td>Uncompressed Qwen-0.5B</td><td>15.21</td><td>16.29</td><td>6.66</td><td>22.62</td><td>6.24</td><td>20.89</td><td>18.99</td></tr><tr><td>Llama3.2-1B (0.5 Wanda)</td><td>35.98</td><td>106.94</td><td>23.24</td><td>85.30</td><td>26.08</td><td>57.45</td><td>51.25</td></tr><tr><td>Llama3.2-1B (0.5 SparseGPT)</td><td>27.96</td><td>129.15</td><td>23.20</td><td>77.16</td><td>23.13</td><td>45.28</td><td>41.38</td></tr><tr><td>Llama3.2-1B (AWQ 8)</td><td>16.20</td><td>30.31</td><td>9.70</td><td>25.23</td><td>11.53</td><td>17.13</td><td>18.61</td></tr><tr><td>Llama3.2-1B (GPTQ 8)</td><td>16.21</td><td>30.34</td><td>9.73</td><td>25.25</td><td>11.54</td><td>17.14</td><td>18.61</td></tr><tr><td>SmolLM2-1.7B (0.75 Wanda)</td><td>953.00</td><td>8939.75</td><td>2999.34</td><td>5709.18</td><td>652.35</td><td>1629.46</td><td>1566.07</td></tr><tr><td>SmolLM2-1.7B (0.75 SparseGPT)</td><td>104.97</td><td>1597.66</td><td>159.86</td><td>833.10</td><td>68.74</td><td>342.06</td><td>348.52</td></tr><tr><td>SmolLM2-1.7B (AWQ 4)</td><td>9.87</td><td>4.95</td><td>4.42</td><td>4.48</td><td>3.25</td><td>12.09</td><td>10.38</td></tr><tr><td>SmolLM2-1.7B (GPTQ 4)</td><td>10.04</td><td>5.23</td><td>4.62</td><td>4.59</td><td>3.38</td><td>12.54</td><td>10.70</td></tr></table>
+
+Table 2: Multilingual perplexity of models in comparable sizes. The column-wise best is highlighted in bold.
+
+primarily from the effective model size or from the compression techniques themselves. To investigate this, we compare  $50\%$  compressed Llama 3.2 1B Instruct (Llama3.2-1B) and  $75\%$  compressed SmolLM2 1.7B Instruct (SmolLM2-1.7B) with the uncompressed Qwen2.5-0.5B-Instruct (Qwen0.5B) in Table 2 and Table 13 (the latter is provided in the Appendix). These models have comparable effective sizes. The results show that the  $75\%$  compressed SmolLM2-1.7B consistently outperforms the full-size small model Qwen-0.5B. Specifically, AWQ yields clear gains in language modeling, whereas GPTQ demonstrates greater robustness on downstream tasks. For instance, SmolLM2-1.7B quantized to 4-bit with AWQ consistently outperforms the uncompressed Qwen-0.5B across both multilingual perplexity and downstream benchmarks. We note that differences in pre-training data and methodology between Qwen-0.5B and SmolLM2-1.7B may act as potential confounders. However, fully controlling for these factors would require pre-training all models from scratch on identical corpora and with identical procedures, which is beyond the scope of this study.
+
+Second, we compare model performance against model size in Fig. 5, where each compressed model is represented by its effective size. For example, if a 1B parameter model is pruned by  $50\%$ , it is plotted as 0.5B. The result clearly shows that compressing a larger model down to a smaller size yields better performance than using an uncompressed small model of the same size.
+
+# 3.6 SLMs under 1B Parameters
+
+To broaden the model size range, we include one SLM under 1B parameters, Qwen2.5-0.5B. As shown in Table 7, Table 8, and Table 9 in the Appendix, the findings for the 0.5B model align with our main conclusion: quantization, particularly AWQ, consistently outperforms pruning.
+
+![](images/33e8e79ec766cecf5c131b263587448e0af50f9eba98fcbc5819399114b50a94.jpg)  
+Figure 5: Comparing pruning and quantization on HelllaSwag. The plot shows compressed model size (x-axis) versus performance (y-axis) across all models and compression setups.
+
+# 4 Discussion
+
+Our comprehensive comparison suggests the superior choice of quantization over pruning for SLMs. Although the selected pruning and quantization techniques follow a similar local activation reconstruction error minimization to retain performance (Kuzmin et al., 2023), we empirically observe that by removing entire weights, pruning impacts performance more severely than quantization. Quantization introduces small, uniform perturbations (noise) to weights, preserving overall model structure and learned representations (Jacob et al., 2018; Stock et al., 2019). Pruning, however, removes entire connections, significantly altering neuron connectivity and functionality (LeCun et al., 1989; Hassibi et al., 1993; Zhu and Gupta, 2017). While quantization reduces numerical precision, it keeps neuron connectivity intact, maintaining higher activation SNR than pruning (Kuzmin et al., 2023; Nagel et al., 2021), theoretically explaining its empirical robustness (Meuser et al., 2024).
+
+# Limitations
+
+While this study provides a systematic and extensive comparison of leading post-training pruning and quantization methods for Small Language Models (SLMs), it is subject to two limitations. Our empirical investigation focuses specifically on six SLMs from 0.5B to 3.8B parameters. While our findings clearly indicate the superiority of quantization over pruning for these models under aggressive compression, the direct applicability and generalization of these conclusions to other sizes of SLMs are not evaluated and remain an open question, as the dynamics of compression might differ. Additionally, we concentrate on four state-of-the-art post-training methods: SparseGPT and Wanda for pruning, and GPTQ and AWQ for quantization. This study does not cover other compression methods requiring full or partial retraining (training-aware pruning/quantization).
+
+# References
+
+Firas Al-Doghman, Nour Moustafa, Ibrahim Khalil, Nasrin Sohrabi, Zahir Tari, and Albert Y Zomaya. 2022. Ai-enabled secure microservices in edge computing: Opportunities and challenges. IEEE Transactions on Services Computing, 16(2):1485-1504.  
+Pratyay Banerjee, Kuntal Kumar Pal, Arindam Mitra, and Chitta Baral. 2019. Careful selection of knowledge to solve open book question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6120-6129, Florence, Italy. Association for Computational Linguistics.  
+Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, and 1 others. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7432-7439.  
+George Chrysostomou, Zhixue Zhao, Miles Williams, and Nikolaos Aletras. 2024. Investigating hallucinations in pruned large language models for abstractive summarization. Transactions of the Association for Computational Linguistics, 12:1163-1181.  
+Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2924–2936, Minneapolis, Minnesota. Association for Computational Linguistics.
+
+Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457.  
+Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In *Machine learning challenges workshop*, pages 177-190. Springer.  
+Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022a. Gpt3.int8(): 8-bit matrix multiplication for transformers at scale. In Advances in Neural Information Processing Systems, volume 35, pages 30318-30332. Curran Associates, Inc.  
+Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022b. GPT3.int8(): 8-bit matrix multiplication for transformers at scale. In Advances in Neural Information Processing Systems.  
+Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022c. Llm.int8(): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339.  
+Tim Dettmers and Luke Zettlemoyer. 2023. The case for 4-bit precision: k-bit inference scaling laws. In International Conference on Machine Learning, pages 7750-7774. PMLR.  
+Elias Frantar and Dan Alistarh. 2022. Optimal brain compression: A framework for accurate posttraining quantization and pruning. arXiv preprint arXiv:2208.11580.  
+Elias Frantar and Dan Alistarh. 2023a. SparseGPT: Massive language models can be accurately pruned in one-shot. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 10323-10337. PMLR.  
+Elias Frantar and Dan Alistarh. 2023b. Sparsegpt: Massive language models can be accurately pruned in one-shot. arXiv preprint arXiv:2301.00774.  
+Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. OPTQ: Accurate quantization for generative pre-trained transformers. In *The Eleventh International Conference on Learning Representations*.  
+Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, and 5 others. 2024. The language model evaluation harness.  
+Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W. Mahoney, and Kurt Keutzer. 2021. A survey of quantization methods for efficient neural network inference. arXiv preprint.
+
+Sukhpal Singh Gill, Muhammed Golec, Jianmin Hu, Minxian Xu, Junhui Du, Huaming Wu, Guneet Kaur Walia, Subramaniam Subramanian Murugesan, Babar Ali, Mohit Kumar, and 1 others. 2025. Edge ai: A taxonomy, systematic review and future directions. Cluster Computing, 28(1):1-53.  
+Song Han, Huizi Mao, and William J. Dally. 2016. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. Preprint, arXiv:1510.00149.  
+Zixu Hao, Huiqiang Jiang, Shiqi Jiang, Ju Ren, and Ting Cao. 2024. Hybrid slim and llm for edge-cloud collaborative inference. In Proceedings of the Workshop on Edge and Mobile Foundation Models, pages 36-41.  
+Babak Hassibi, David Stork, and Gregory Wolff. 1993. Optimal brain surgeon: Extensions and performance comparisons. In Advances in Neural Information Processing Systems.  
+Torsten Hoeffer, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. 2021. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. Journal of Machine Learning Research, 22(241):1-124.  
+Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. 2018. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Ajay Jaiswal, Zhe Gan, Xianzhi Du, Bowen Zhang, Zhangyang Wang, and Yinfei Yang. 2023. Compressing llms: The truth is rarely pure and never simple. arXiv preprint arXiv:2310.01382.  
+Sho Kurita, Jonas Pfeiffer, Ivan Vulic, Edoardo Ponti, and Anna Korhonen. 2020. A weighted approach to unsupervised multilingual transformer fine-tuning. In Proc. of AACL-IJCNLP.  
+Simon Kurz, Jian-Jia Chen, Lucie Flek, and Zhixue Zhao. 2024. Investigating language-specific calibration for pruning multilingual large language models. arXiv preprint arXiv:2408.14398.  
+Andrey Kuzmin, Markus Nagel, Mart van Baalen, Arash Behboodi, and Tijmen Blankevoort. 2023. Pruning vs quantization: Which is better? In Advances in Neural Information Processing Systems.  
+Anne Lauscher, Goran Glavaš, and Simone Paolo Ponzetto. 2020. Language diversity in multilingual NLP models: Current trends and limitations. In Proc. of COLING.  
+Yann LeCun, John Denker, and Sara Solla. 1989. Optimal brain damage. In Advances in Neural Information Processing Systems.
+
+Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, and 13 others. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. 2021. {BRECQ}: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations.  
+Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024. Awq: Activation-aware weight quantization for ondevice llm compression and acceleration. In Proceedings of Machine Learning and Systems.  
+Zhenyan Lu, Xiang Li, Dongqi Cai, Rongjie Yi, Fangming Liu, Xiwen Zhang, Nicholas D Lane, and Mengwei Xu. 2024. Small language models: Survey, measurements, and insights. arXiv preprint arXiv:2409.15790.  
+Tobias Meuser, Lauri Lovén, Monowar Bhuyan, Shishir G Patil, Schahram Dustdar, Atakan Aral, Suzan Bayhan, Christian Becker, Eyal De Lara, Aaron Yi Ding, and 1 others. 2024. Revisiting edge ai: Opportunities and challenges. IEEE Internet Computing, 28(4):49-59.  
+Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839-849, San Diego, California. Association for Computational Linguistics.  
+Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. 2021. A white paper on neural network quantization. ArXiv, abs/2106.08295.  
+Denis Paperno, Germán Kruszewski, Angeliki Lazari-dou, Ngoc Quan Pham, Raffaella Bernardi, Sandro Pezzelle, Marco Baroni, Gemma Boleda, and Raquel Fernandez. 2016. The LAMBADA dataset: Word prediction requiring a broad discourse context. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1525-1534, Berlin, Germany. Association for Computational Linguistics.
+
+Jonas Pfeiffer, Ivan Vulic, Iryna Gurevych, and Sebastian Ruder. 2020. MAD-X: An Adapter-Based Framework for Multi-Task Cross-Linguual Transfer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7654–7673, Online. Association for Computational Linguistics.  
+Mangesh Pujari and Anil Kumar Pakina. 2024. Edgeai for privacy-preserving ai: The role of small llms in federated learning environments. International Journal of Engineering and Computer Science, 13(10):26589-26601.  
+Samarth N Ramesh and Zhixue Zhao. 2024. Efficient pruning of text-to-image models: Insights from pruning stable diffusion. arXiv preprint arXiv:2411.15113.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106.  
+Pierre Stock, Armand Joulin, Rémi Gribonval, Benjamin Graham, and Herve Jégou. 2019. And the bit goes down: Revisiting the quantization of neural networks. arXiv preprint arXiv:1907.05686.  
+Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. 2024. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations.  
+Rui Wang, Zhiyong Gao, Liuyang Zhang, Shuaibing Yue, and Ziyi Gao. 2025. Empowering large language models to edge intelligence: A survey of edge efficient llms and techniques. Computer Science Review, 57:100755.  
+Miles Williams and Nikolaos Aletras. 2024. On the impact of calibration data in post-training quantization and pruning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10100-10118, Bangkok, Thailand. Association for Computational Linguistics.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and 3 others. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Jiajun Xu, Zhiyuan Li, Wei Chen, Qun Wang, Xin Gao, Qi Cai, and Ziyuan Ling. 2024. On-device language models: A comprehensive review. arXiv preprint arXiv:2409.00088.
+
+Lu Yin, Shiwei Liu, AJAY KUMAR JAISWAL, Souvik Kundu, and Zhangyang Wang. 2024. Junk DNA hypothesis: A task-centric angle of LLM pre-trained weights through sparsity.  
+Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy. Association for Computational Linguistics.  
+Michael Zhu and Suyog Gupta. 2017. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878.  
+Xunyu Zhu, Jian Li, Yong Liu, Can Ma, and Weiping Wang. 2024. A survey on model compression for large language models. Transactions of the Association for Computational Linguistics, 12:1556-1577.
+
+# A Related Work
+
+Pruning and quantization are two widely adopted approaches for model compression (Gholami et al., 2021; Hoefler et al., 2021; Zhu et al., 2024). With origins in seminal work from Hassibi et al. (1993); LeCun et al. (1989), quantization achieves compression by reducing the numerical precision of model parameters, while pruning determines redundant weights for removal (Han et al., 2016). Their application to LLMs poses significant challenges, such as the importance of retaining large-magnitude outlier features, interplay between weights and inputs, and high computational requirements (Dettmers et al., 2022a). Recently, advanced pruning methods like SparseGPT (Frantar and Alistarh, 2023a) and Wanda (Sun et al., 2024), and sophisticated quantization techniques such as GPTQ (Frantar et al., 2023) and AWQ (Lin et al., 2024), have significantly advanced LLM compression in the post-training and retraining-free setting.
+
+While quantization and pruning achieve compression through different means, they share the common optimization goal of block-wise reconstruction error minimization (Frantar and Alistarh, 2022; Li et al., 2021). That is, model weights get compressed in a way that best preserves the original outputs of a model block. More formally, given layer  $\ell$  with weights  $\mathbf{W}_{\ell}$  and inputs  $\mathbf{X}_{\ell}$ , the goal is to minimize  $||\mathbf{W}_{\ell}\mathbf{X}_{\ell} - \widehat{\mathbf{W}}_{\ell}\mathbf{X}_{\ell}||_2^2$  with respect to the compressed weights  $\widehat{\mathbf{W}}$ .
+
+While pruning and quantization are now widely used for compressing LLMs, early comparative studies (Kuzmin et al., 2023) rely on outdated techniques such as magnitude-based pruning and plain
+
+symmetric quantization. These approaches typically minimize weight reconstruction error without accounting for activation dynamics, limiting their relevance to current methods. Foundational theoretical publications (Frantar and Alistarh, 2022; Dettmers et al., 2022b; Dettmers and Zettlemoyer, 2023) fostering later compression techniques emphasize the importance of preserving activations, reflecting a broader shift in compression objectives. This shift has yielded more effective methods. Previous work has compared pruning and quantization for LLMs, but without focusing on compression errors (Jaiswal et al., 2023). Moreover, SLMs remain underexplored in this setting. As SLMs grow in importance for efficient deployment, it is unclear whether findings from LLMs carry over.
+
+# B Implementation Details
+
+We follow the exact hyperparameter settings from prior work for pruning and quantization. Specifically, for Wanda, SparseGPT, and GPTQ, we sample the calibration set from the initial shard of the C4 dataset, following the methodology of Frantar et al. (2023).
+
+As previous studies have shown that the benefit of increasing calibration samples saturates logarithmically (Frantar and Alistarh, 2023b; Sun et al., 2024), we adopt the standard protocol of randomly sampling 128 calibration examples, each containing 2,048 tokens, totaling 262,144 tokens. For further discussion on calibration set sensitivity, we refer to the analysis by Williams and Aletras (2024).
+
+AWQ uses a smaller calibration set: 128 samples of 512 tokens each, drawn from The Pile, totaling 65,536 tokens. For AWQ, SparseGPT, and GPTQ, we set the group size to 128 and prune for unstructured sparsity.
+
+To evaluate pruning error, SNR, and cosine similarity, we sample 128 sequences of 2,048 tokens from the C4 validation set. Metrics are computed token-wise at each layer output. For multilingual perplexity evaluations, we use the corresponding mC4 validation set for each language, maintaining the same number of samples and tokens as in the internal feature analysis.
+
+All calibration datasets are sourced via public versions on Hugging Face Datasets (Lhoest et al., 2021), and all model architectures and configurations are loaded from Hugging Face Transformers (Wolf et al., 2020). For quantization, we use the Optimum library for GPTQ and AutoAWQ for AWQ.
+
+SparseGPT and Wanda are implemented using the official GitHub repositories provided by the respective paper authors.
+
+For the construction of calibration corpora, we employ the publicly accessible iterations of each source dataset available through Hugging Face Datasets (Lhoest et al., 2021). Similarity, we utilize the parametric configurations and implementations of each architectural model from Hugging Face Transformers (Wolf et al., 2020).
+
+# C Evaluation
+
+# C.1 Quantization and Pruning Error
+
+We refer to the definition of Kuzmin et al. (2023) for computing the Pruning Error and SNR. However, both of them relate to model weights, which does not reflect the current compression paradigm of emphasizing the importance of features. Therefore, we compute Pruning Error and SNR with respect to the outputs of each layer, i.e. the hidden states. To cope with different hidden state magnitudes per layer, all hidden states are normalized layer-wise by their average vector norm, yielding the following pruning error:
+
+$$
+E [ (H ^ {(l)} - \widetilde {H} ^ {(l)}) ^ {2} ] = \frac {1}{N d} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {d} \left(\frac {h _ {i , j} ^ {(l)} - \widetilde {h} _ {i , j} ^ {(l)}}{\mu^ {(l)}}\right) ^ {2} \tag {1}
+$$
+
+$$
+\mu^ {(l)} = \frac {1}{N} \sum_ {i = 1} ^ {N} | | h _ {i} ^ {(l)} | | _ {2} \tag {2}
+$$
+
+with  $H^{(l)} \in N \times d$  as the hidden states of all  $N$  tokens in layer  $k$  before compression and  $\widetilde{H}^{(l)} \in N \times d$  after compression. Individual elements of the hidden states are denoted as  $h_{i,j}^{(l)}$ , referring to the hidden state of the  $j$ -th feature element of the  $i$ -th token in  $H^{(l)}$ .
+
+# C.2 Signal-to-noise ratio (SNR)
+
+Following the same notation as for computing the compression error, the SNR of a single layer  $l$  is computed as
+
+$$
+S N R _ {d B} ^ {(l)} = 1 0 \log_ {1 0} \left(\frac {E \left[ H ^ {(l)}\right) ^ {2}}{E \left[ \left(H ^ {(l)} - \tilde {H} ^ {(l)}\right) ^ {2} \right]}\right) \tag {3}
+$$
+
+The final model SNR and compression error is the average over all layer-wise  $SNR_{dB}^{(l)}$  or compression errors respectively.
+
+# C.3 Practical Deployment Comparison
+
+Given the growing need to deploy language models on resource-constrained edge devices, practical deployment-relevant statistics such as memory footprint, latency and inference speed are crucial and highly informative for practitioners. In response, We have included theoretical inference speedup from Wanda, SparseGPT, AWQ, and GPTQ papers, and also calculated FLOPs for comparison in Table 3.
+
+# C.4 Datasets
+
+Table 4 presents the number of examples across the evaluation tasks, extracted from their respective test or validation partitions.
+
+# D Fidelity: SNR and Compression Error
+
+We report average activation-level SNR and compression error across all models and configurations as shown in Table 5. Quantization methods consistently yield higher SNR and lower error than pruning methods, particularly under higher compression. AWQ achieves the strongest overall fidelity, followed by GPTQ. Among pruning methods, SparseGPT consistently outperforms Wanda. Additionally, both SNR and error exhibit a depth-wise trend, with deeper layers generally showing greater degradation. Results on English are visualized in Fig.6,Fig.7, Fig.8, Fig.9, and Fig.10.
+
+<table><tr><td>Method</td><td>Models &amp; Size</td><td>Sparsity</td><td>Device</td><td>Memory Footprint</td><td>FLOPs / Ops</td><td>Performance (speed/latency)</td></tr><tr><td>Wanda</td><td>LLaMA (7B, 13B, 30B, 65B)</td><td>50% unstructured; 2:4, 4:8</td><td>GPU/CPU (A6000)</td><td>≈ 0.5× (weights)</td><td>≈ 0.5× (ops)</td><td>0.54 s vs 203 s prune (7B, A6000); ~1.5–1.8× for 2:4</td></tr><tr><td>SparseGPT</td><td>OPT/BLOOM (up to 175B; LLaMA 13B)</td><td>50-60% unstructured; 2:4, 4:8</td><td>GPU (A100)</td><td>≈ 0.5× (weights)</td><td>≈ 0.5× (ops)</td><td>Prunes 175B in ≈ 4.5 h; negligible accuracy loss up to 50-60% sparsity; ~1.5-1.8× speedup for 2:4 sparsity</td></tr><tr><td>AWQ</td><td>LLaMA-2/Vicuna/MFM (up to 70B)</td><td>INT4 weights (keep~1% FP16)</td><td>GPU (RTX4090)</td><td>≈ 0.25× (weights vs FP16)</td><td>≈ 1× (ops)</td><td>3.2-3.9× speedup vs FP16; 30 tok/s (13B on RTX4070-8GB)</td></tr><tr><td>GPTQ</td><td>GPT/OPT (up to 175B; LLaMA 7B/13B/30B/65B)</td><td>INT4/3 (layerwise; some INT2)</td><td>GPU (A100/A6000)</td><td>≈ 0.25× (weights vs FP16)</td><td>≈ 1× (ops)</td><td>~3.25× (A100)-4.5× (A6000) speedup vs FP16; single-GPU 175B</td></tr></table>
+
+Table 3: Practical Deployment Metrics comparison summary  
+
+<table><tr><td>Dataset</td><td># Examples</td></tr><tr><td>ARC-Easy (Clark et al., 2018)</td><td>2,376</td></tr><tr><td>ARC-Challenge (Clark et al., 2018)</td><td>1,172</td></tr><tr><td>BoolQ (Clark et al., 2019)</td><td>3,270</td></tr><tr><td>HellaSwag (Zellers et al., 2019)</td><td>10,042</td></tr><tr><td>LAMBADA (Paperno et al., 2016)</td><td>5,153</td></tr><tr><td>OpenBookQA (Banerjee et al., 2019)</td><td>500</td></tr><tr><td>PIQA (Bisk et al., 2020)</td><td>1,838</td></tr><tr><td>RTE (Dagan et al., 2005)</td><td>277</td></tr><tr><td>StoryCloze (Mostafazadeh et al., 2016)</td><td>1,511</td></tr><tr><td>WinoGrande (Sakaguchi et al., 2021)</td><td>1,267</td></tr></table>
+
+Table 4: Number of examples for each evaluation task.  
+
+<table><tr><td>Model</td><td>Sparsity/Bit-width</td><td>Method</td><td>SNR</td><td>Error(×10-4)</td></tr><tr><td rowspan="8">SmolLM2</td><td rowspan="4">0.50 / 8-bit</td><td>SparseGPT</td><td>11.75</td><td>1.31</td></tr><tr><td>Wanda</td><td>11.87</td><td>1.32</td></tr><tr><td>AWQ</td><td>44.78</td><td>0.00</td></tr><tr><td>GPTQ</td><td>41.71</td><td>0.00</td></tr><tr><td rowspan="4">0.75 / 4-bit</td><td>SparseGPT</td><td>4.57</td><td>7.46</td></tr><tr><td>Wanda</td><td>2.65</td><td>12.51</td></tr><tr><td>AWQ</td><td>18.97</td><td>0.23</td></tr><tr><td>GPTQ</td><td>15.86</td><td>0.71</td></tr><tr><td rowspan="8">Phi3.5</td><td rowspan="4">0.50 / 8-bit</td><td>SparseGPT</td><td>10.53</td><td>4.82</td></tr><tr><td>Wanda</td><td>10.42</td><td>4.17</td></tr><tr><td>AWQ</td><td>37.47</td><td>0.02</td></tr><tr><td>GPTQ</td><td>34.95</td><td>0.04</td></tr><tr><td rowspan="4">0.75 / 4-bit</td><td>SparseGPT</td><td>5.28</td><td>11.17</td></tr><tr><td>Wanda</td><td>4.70</td><td>8.49</td></tr><tr><td>AWQ</td><td>15.97</td><td>1.47</td></tr><tr><td>GPTQ</td><td>14.86</td><td>1.50</td></tr></table>
+
+Table 5: Full results of layer-wise average activation SNR and compression error (Error).
+
+# E Multilingual Language Modeling
+
+Table 6 presents token-level perplexity across seven languages for all models and compression settings. Quantization maintains stable performance across languages and compression levels, while pruning leads to greater degradation, especially at higher
+
+![](images/9f06ea93b88e03465c098bcdcb3466498315e369bdb6c96fc28b0cb9a82ec93a.jpg)  
+Figure 6: Layer-wise activation SNR on English tokens for R1.
+
+![](images/df5ab278acda84c91ef33c964e99f9188314da0ead9a20fca44f1215af41041d.jpg)  
+Figure 7: Layer-wise activation SNR on English tokens for Qwen2.5.
+
+sparsity. The performance drop under pruning varies across languages and models, reflecting the interaction between compression sensitivity, tokenization, and pretraining distribution.
+
+# F Small models smaller than 1B
+
+# G Reasoning Tasks
+
+To compare the downstream task performance, we include a wide range of reasoning tasks for zero-shot evaluation shown in Table 10. These tasks primarily assess commonsense reasoning abilities, using binary, multiple choice, Cloze, and Winograd style questions, covering different difficulty levels: OBQA (OpenBookQA) is the multi-hop reasoning task, requiring combining multiple facts from the "open book" knowledge base to answer questions.
+
+<table><tr><td>Model</td><td>Sparsity</td><td>Method</td><td>en</td><td>ar</td><td>hi</td><td>zh</td><td>th</td><td>de</td><td>es</td></tr><tr><td rowspan="9">R1</td><td>0.00</td><td>Full Precision</td><td>29.79</td><td>159.39</td><td>11.02</td><td>41.37</td><td>40.67</td><td>98.58</td><td>85.36</td></tr><tr><td rowspan="4">0.50</td><td>SparseGPT</td><td>44.41</td><td>379.12</td><td>16.26</td><td>117.87</td><td>306.32</td><td>193.78</td><td>162.97</td></tr><tr><td>Wanda</td><td>47.75</td><td>451.78</td><td>16.85</td><td>114.34</td><td>170.08</td><td>206.75</td><td>164.00</td></tr><tr><td>AWQ</td><td>29.80</td><td>159.51</td><td>11.03</td><td>41.38</td><td>40.46</td><td>98.73</td><td>85.48</td></tr><tr><td>GPTQ</td><td>29.83</td><td>159.32</td><td>11.03</td><td>41.38</td><td>40.87</td><td>98.66</td><td>85.59</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>292.42</td><td>13469.27</td><td>996.42</td><td>5767.65</td><td>2922.67</td><td>5566.94</td><td>2966.31</td></tr><tr><td>Wanda</td><td>1180.82</td><td>201520.22</td><td>42817.80</td><td>9133.28</td><td>60985.34</td><td>26397.47</td><td>24978.16</td></tr><tr><td>AWQ</td><td>33.22</td><td>190.01</td><td>12.30</td><td>49.83</td><td>48.97</td><td>112.98</td><td>96.45</td></tr><tr><td>GPTQ</td><td>32.28</td><td>185.16</td><td>12.29</td><td>49.08</td><td>46.54</td><td>112.37</td><td>99.91</td></tr><tr><td rowspan="9">Qwen2.5</td><td>0.00</td><td>Full Precision</td><td>10.98</td><td>11.00</td><td>5.42</td><td>12.07</td><td>5.35</td><td>13.31</td><td>11.45</td></tr><tr><td rowspan="4">0.50</td><td>SparseGPT</td><td>15.54</td><td>41.40</td><td>11.02</td><td>33.20</td><td>15.11</td><td>27.49</td><td>22.64</td></tr><tr><td>Wanda</td><td>16.38</td><td>39.73</td><td>10.65</td><td>30.49</td><td>16.91</td><td>31.39</td><td>26.26</td></tr><tr><td>AWQ</td><td>10.98</td><td>11.00</td><td>5.42</td><td>12.07</td><td>5.35</td><td>13.31</td><td>11.46</td></tr><tr><td>GPTQ</td><td>10.98</td><td>11.00</td><td>5.42</td><td>12.07</td><td>5.35</td><td>13.32</td><td>11.46</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>155.99</td><td>1158.69</td><td>324.37</td><td>2424.02</td><td>961.63</td><td>1238.69</td><td>1102.33</td></tr><tr><td>Wanda</td><td>1436.60</td><td>26945.81</td><td>12328.96</td><td>69817.35</td><td>16434.86</td><td>7782.48</td><td>5573.80</td></tr><tr><td>AWQ</td><td>11.79</td><td>12.68</td><td>6.14</td><td>13.19</td><td>6.08</td><td>14.87</td><td>12.48</td></tr><tr><td>GPTQ</td><td>12.02</td><td>13.44</td><td>6.41</td><td>14.25</td><td>6.33</td><td>15.32</td><td>12.81</td></tr><tr><td rowspan="9">Llama3.2</td><td>0.00</td><td>Full Precision</td><td>16.20</td><td>30.31</td><td>9.72</td><td>25.23</td><td>11.53</td><td>17.13</td><td>18.61</td></tr><tr><td rowspan="4">0.50</td><td>SparseGPT</td><td>27.96</td><td>129.15</td><td>23.2</td><td>77.16</td><td>23.13</td><td>45.28</td><td>41.38</td></tr><tr><td>Wanda</td><td>35.98</td><td>106.94</td><td>23.24</td><td>85.3</td><td>26.08</td><td>57.45</td><td>51.25</td></tr><tr><td>AWQ</td><td>16.2</td><td>30.31</td><td>9.7</td><td>25.23</td><td>11.53</td><td>17.13</td><td>18.61</td></tr><tr><td>GPTQ</td><td>16.21</td><td>30.34</td><td>9.73</td><td>25.25</td><td>11.54</td><td>17.14</td><td>18.61</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>273.95</td><td>144742.47</td><td>53063.41</td><td>41727.27</td><td>40215.73</td><td>1429.28</td><td>1111.82</td></tr><tr><td>Wanda</td><td>10130.75</td><td>302363.66</td><td>116279.45</td><td>180397.61</td><td>189368.5</td><td>45676.04</td><td>22685.21</td></tr><tr><td>AWQ</td><td>17.53</td><td>36.64</td><td>10.95</td><td>29.96</td><td>12.94</td><td>19.46</td><td>20.72</td></tr><tr><td>GPTQ</td><td>18.9</td><td>46.59</td><td>13.22</td><td>37.00</td><td>15.44</td><td>22.84</td><td>23.29</td></tr><tr><td rowspan="9">SmolLM2</td><td>0.00</td><td>Full Precision</td><td>9.27</td><td>4.50</td><td>4.00</td><td>4.11</td><td>2.94</td><td>10.86</td><td>9.57</td></tr><tr><td rowspan="4">0.50</td><td>SparseGPT</td><td>13.85</td><td>8.98</td><td>6.93</td><td>10.70</td><td>4.23</td><td>20.07</td><td>18.75</td></tr><tr><td>Wanda</td><td>13.63</td><td>7.97</td><td>6.48</td><td>10.37</td><td>4.22</td><td>20.80</td><td>18.78</td></tr><tr><td>AWQ</td><td>9.27</td><td>4.51</td><td>4.01</td><td>4.11</td><td>2.94</td><td>10.86</td><td>9.57</td></tr><tr><td>GPTQ</td><td>9.27</td><td>4.51</td><td>4.01</td><td>4.11</td><td>2.94</td><td>10.87</td><td>9.57</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>104.97</td><td>1597.66</td><td>159.86</td><td>833.10</td><td>68.74</td><td>342.06</td><td>348.52</td></tr><tr><td>Wanda</td><td>953.00</td><td>8939.75</td><td>2999.34</td><td>5709.18</td><td>652.35</td><td>1629.46</td><td>1566.07</td></tr><tr><td>AWQ</td><td>9.87</td><td>4.95</td><td>4.42</td><td>4.48</td><td>3.25</td><td>12.09</td><td>10.38</td></tr><tr><td>GPTQ</td><td>10.04</td><td>5.23</td><td>4.62</td><td>4.59</td><td>3.38</td><td>12.54</td><td>10.70</td></tr><tr><td rowspan="9">Phi3.5</td><td>0.00</td><td>Full Precision</td><td>7.75</td><td>3.20</td><td>3.71</td><td>5.04</td><td>3.85</td><td>8.61</td><td>8.64</td></tr><tr><td rowspan="4">0.50</td><td>SparseGPT</td><td>11.34</td><td>9.11</td><td>9.90</td><td>15.26</td><td>8.69</td><td>15.78</td><td>16.49</td></tr><tr><td>Wanda</td><td>11.08</td><td>7.71</td><td>8.32</td><td>13.05</td><td>7.69</td><td>14.60</td><td>15.43</td></tr><tr><td>AWQ</td><td>7.75</td><td>3.20</td><td>3.71</td><td>5.04</td><td>3.85</td><td>8.61</td><td>8.64</td></tr><tr><td>GPTQ</td><td>7.75</td><td>3.20</td><td>3.71</td><td>5.04</td><td>3.85</td><td>8.61</td><td>8.65</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>76.53</td><td>280.62</td><td>227.38</td><td>749.68</td><td>205.68</td><td>669.48</td><td>594.91</td></tr><tr><td>Wanda</td><td>1068.58</td><td>1329864.75</td><td>2380391.50</td><td>3571925.75</td><td>9051510.00</td><td>14388.79</td><td>19762.86</td></tr><tr><td>AWQ</td><td>8.06</td><td>3.56</td><td>4.25</td><td>5.48</td><td>4.26</td><td>9.14</td><td>9.09</td></tr><tr><td>GPTQ</td><td>8.20</td><td>3.73</td><td>4.42</td><td>5.66</td><td>4.50</td><td>9.49</td><td>9.40</td></tr></table>
+
+Table 6: Multilingual PPL results across models, sparsity levels, and compression methods. 4-bit quantization corresponds to approximately  $75\%$  sparsity, 8-bit to  $50\%$ .
+
+![](images/0585d325dfb365470a50e78c1d9b6a0f34a3435992c107b12bc731fa788a8544.jpg)  
+Figure 8: Layer-wise activation SNR on English tokens for Llama3.2.
+
+![](images/80a4e11767b01c64b554b416419fd6ad0e3294947a40e39c6d1fe1f6a7b5dab3.jpg)  
+Figure 9: Layer-wise activation SNR on English tokens for SmolLM2.
+
+Other tasks (e.g., BoolQ, RTE, HellaSwag) primarily assess single-step reasoning or commonsense understanding.
+
+# H Cosine Similarity
+
+We visualize layer-wise cosine similarity between the hidden states of compressed and original models in Fig.4. Quantization methods preserve higher similarity across layers and models. Pruning methods show lower similarity, particularly in deeper layers. The similarity decreases with more aggressive compression, but the relative advantage of quantization remains consistent.
+
+# I Limitations of Wanda in SLMs
+
+The contrast between our findings on SLMs and prior work on LLMs (where Wanda typically out
+
+<table><tr><td>Model</td><td>SNR</td><td>Error</td></tr><tr><td>SparseGPT (unstructured 50%)</td><td>11.94</td><td>2.77e-4</td></tr><tr><td>Wanda (unstructured 50%)</td><td>11.66</td><td>2.76e-4</td></tr><tr><td>GPTQ INT8</td><td>43.16</td><td>1.86e-7</td></tr><tr><td>AWQ INT8</td><td>48.85</td><td>5.56e-8</td></tr><tr><td>SparseGPT (unstructured 75%)</td><td>4.85</td><td>16.47e-4</td></tr><tr><td>Wanda (unstructured 75%)</td><td>1.56</td><td>25.29e-4</td></tr><tr><td>GPTQ INT4</td><td>18.91</td><td>0.49e-4</td></tr><tr><td>AWQ INT4</td><td>18.59</td><td>0.53e-4</td></tr></table>
+
+Table 7: Compression results for Qwen2.5-0.5B-Instructure under different compression methods. The column-wise best is highlighted in bold.
+
+![](images/d0e9e63f754125f27ff6d67a70fa548a30a8e086fbbe6646f3468fbfe586825e.jpg)  
+Figure 10: Layer-wise activation SNR on English tokens for Phi3.5.
+
+![](images/9f196126a8afb03a4ce25d35e10bb0bbfaa9f9b18830c650e10a2880140d1de0.jpg)  
+Figure 11: Layer-wise averaged cosine similarity of quantized and pruned models compared to the original model. Quantization is shown with different y-axis scales.
+
+performs SparseGPT) is noteworthy and merits further discussion. We believe several factors might contribute to this divergence.
+
+First, LLMs have more redundant weights, allowing pruning (even heuristic methods like Wanda) to preserve performance. Wanda's success on LLMs, e.g. Llama 7B, attests that many weights in a  $7\mathrm{B}+$  model can be removed without significant impact, the remaining weights can compensate, and important features are preserved. In smaller models, by contrast, there are fewer extraneous weights. Every parameter likely has a higher relative contribution to some facet of the model's knowledge or capacity. Therefore, pruning  $50\%$  of a 1.5B model removes a much larger fraction of the model's "knowledge" than pruning  $50\%$  of a 13B model. Wanda performs better than magnitude, but it is still essentially 'a heuristic selection of weights'; its design doesn't attempt to minimize the difference before and after pruning, but rather removes weights with smaller activations. On the other hand, SparseGPT uses a second-order (Hessian-based) approximation to measure the sensitivity of the model's outputs to
+
+<table><tr><td>Model</td><td>en</td><td>hi</td><td>zh</td><td>th</td><td>de</td><td>ar</td><td>es</td></tr><tr><td>Full Size</td><td>16.29</td><td>6.66</td><td>22.62</td><td>6.24</td><td>20.89</td><td>18.99</td><td>15.21</td></tr><tr><td>Sparseqpt (0.5)</td><td>22.21</td><td>67.18</td><td>14.88</td><td>45.19</td><td>16.01</td><td>46.40</td><td>36.92</td></tr><tr><td>Wanda (0.5)</td><td>25.62</td><td>50.39</td><td>14.12</td><td>47.91</td><td>16.23</td><td>55.51</td><td>44.83</td></tr><tr><td>GPTQ (INT8)</td><td>14.92</td><td>16.90</td><td>7.09</td><td>16.64</td><td>7.22</td><td>20.57</td><td>16.69</td></tr><tr><td>AWQ (INT8)</td><td>14.92</td><td>16.89</td><td>7.09</td><td>16.63</td><td>7.22</td><td>20.56</td><td>16.68</td></tr><tr><td>Sparseqpt (0.75)</td><td>322.83</td><td>2026.32</td><td>975.30</td><td>2555.91</td><td>473.33</td><td>1964.03</td><td>2108.29</td></tr><tr><td>Wanda (0.75)</td><td>913.84</td><td>7710.46</td><td>2358.86</td><td>4268.96</td><td>2606.93</td><td>5548.94</td><td>5131.00</td></tr><tr><td>GPTQ (INT4)</td><td>16.97</td><td>22.84</td><td>9.47</td><td>21.13</td><td>9.59</td><td>26.00</td><td>19.93</td></tr><tr><td>AWQ (INT4)</td><td>17.38</td><td>22.40</td><td>9.09</td><td>20.62</td><td>8.93</td><td>26.10</td><td>20.08</td></tr></table>
+
+Table 8: New multilingual perplexity results for Qwen2.5-0.5-Instructure under different compression methods. The column-wise best is highlighted in bold.  
+
+<table><tr><td>Model</td><td>BoolQ</td><td>RTE</td><td>HellaSwag</td><td>WinoGrande</td><td>ARC-e</td><td>ARC-c</td><td>OBQA</td></tr><tr><td>Full Size</td><td>67.95</td><td>62.45</td><td>40.63</td><td>55.88</td><td>65.53</td><td>30.89</td><td>24.2</td></tr><tr><td>Sparseqpt (0.5)</td><td>41.22</td><td>51.62</td><td>35.76</td><td>54.06</td><td>57.03</td><td>24.83</td><td>20.2</td></tr><tr><td>Wanda (0.5)</td><td>58.29</td><td>53.43</td><td>34.44</td><td>52.64</td><td>57.79</td><td>25.09</td><td>19.4</td></tr><tr><td>GPTQ (INT8)</td><td>66.63</td><td>63.54</td><td>39.68</td><td>55.88</td><td>54.80</td><td>27.47</td><td>24.0</td></tr><tr><td>AWQ (INT8)</td><td>66.42</td><td>62.45</td><td>39.67</td><td>56.20</td><td>55.35</td><td>28.33</td><td>24.2</td></tr><tr><td>Sparseqpt (0.75)</td><td>43.33</td><td>56.32</td><td>27.00</td><td>49.57</td><td>33.67</td><td>18.69</td><td>16.0</td></tr><tr><td>Wanda (0.75)</td><td>37.83</td><td>52.71</td><td>26.53</td><td>49.57</td><td>28.96</td><td>18.43</td><td>12.6</td></tr><tr><td>GPTQ (INT4)</td><td>65.32</td><td>65.70</td><td>38.25</td><td>51.46</td><td>51.64</td><td>27.82</td><td>22.4</td></tr><tr><td>AWQ (INT4)</td><td>65.66</td><td>59.21</td><td>37.90</td><td>52.80</td><td>51.98</td><td>27.30</td><td>20.4</td></tr></table>
+
+Table 9: New downstream tasks results for Qwen2.5-0.5-Instruct under different compression methods. The column-wise best is highlighted in bold.
+
+each weight, which can benefit SLMs more than LLMs, as SLMs have fewer redundant weights.
+
+Second, Wand lacks weight adjustment and error compensation. In a small model with limited redundancy, any one-shot removal of weights leaves more unrecoverable damage. SparseGPT mitigates this damage by reallocating some weight importance during its adjustment step. On the other hand, Wanda's design, which assumes large-model characteristics (many dispensable weights and obvious outlier features), is less suited to the small-model regime. Its lack of any weight adjustment leaves small models struggling to recover from pruning-induced damage.
+
+Third, Wanda's core assumption leverages the phenomenon of outlier features, a well-documented effect in LLMs where a few dimensions in each layer's activation have disproportionately large variance or magnitude. These outlier dimensions often carry crucial semantic information (e.g. they correlate with rare but important token patterns) and must not be pruned aggressively. Wanda implicitly protects those: even a small weight on an outlier neuron won't be pruned if the activation norm is huge, because the weight  $\times$  activation score would still be relatively large. In large models (like 13B, 70B), such outlier channels are prominent and
+
+Wanda's heuristic is very effective at avoiding the truly important weights. However, outlier behavior is less studied and potentially less pronounced in SLMs. If outlier-driven structure is weaker or absent, Wanda's assumptions may not hold, leading to less effective pruning.
+
+# J 2:4 Structured Pruning
+
+Table 11 and Table 12 report the results on 2:4 structured pruning using Wanda and SparseGPT. Overall, unstructured pruning consistently outperforms structured pruning, in line with prior findings in LLMs.
+
+This finding has practical implications, especially for SLM practitioners: since SLMs are already small and require less compute, which lowers the barrier for applying unstructured pruning, practitioners have the flexibility to choose between structured and unstructured approaches. Our results suggest that unstructured pruning remains preferable for optimal performance.
+
+# J.1 Models in comparable size
+
+<table><tr><td>Model</td><td>Sparsity</td><td>Method</td><td>BoolQ</td><td>RTE</td><td>HellaSwag</td><td>WinoGrande</td><td>ARC-e</td><td>ARC-c</td><td>OBQA</td></tr><tr><td rowspan="9">R1</td><td>0.00</td><td>Full Precision</td><td>51.10</td><td>46.93</td><td>36.13</td><td>51.07</td><td>42.93</td><td>27.56</td><td>18.60</td></tr><tr><td rowspan="4">0.5</td><td>SparseGPT</td><td>66.21</td><td>55.96</td><td>33.95</td><td>55.72</td><td>54.34</td><td>29.44</td><td>19.80</td></tr><tr><td>Wanda</td><td>64.56</td><td>57.76</td><td>33.27</td><td>54.14</td><td>54.08</td><td>28.24</td><td>18.80</td></tr><tr><td>AWQ</td><td>50.06</td><td>48.74</td><td>36.03</td><td>50.99</td><td>42.55</td><td>27.99</td><td>18.40</td></tr><tr><td>GPTQ</td><td>54.16</td><td>47.65</td><td>36.20</td><td>50.67</td><td>43.10</td><td>27.73</td><td>18.80</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>61.16</td><td>52.35</td><td>27.11</td><td>49.80</td><td>27.82</td><td>19.71</td><td>12.60</td></tr><tr><td>Wanda</td><td>59.76</td><td>52.71</td><td>26.36</td><td>49.01</td><td>28.16</td><td>20.31</td><td>14.40</td></tr><tr><td>AWQ</td><td>61.07</td><td>52.35</td><td>35.81</td><td>50.51</td><td>42.21</td><td>28.92</td><td>17.60</td></tr><tr><td>GPTQ</td><td>53.67</td><td>54.15</td><td>35.46</td><td>52.17</td><td>42.89</td><td>26.96</td><td>19.00</td></tr><tr><td rowspan="9">Qwen2.5</td><td>0.00</td><td>Full Precision</td><td>78.32</td><td>73.65</td><td>49.30</td><td>59.59</td><td>68.39</td><td>38.48</td><td>31.20</td></tr><tr><td rowspan="4">0.5</td><td>SparseGPT</td><td>76.36</td><td>75.45</td><td>44.61</td><td>61.80</td><td>68.01</td><td>36.18</td><td>27.80</td></tr><tr><td>Wanda</td><td>69.39</td><td>67.51</td><td>42.28</td><td>60.06</td><td>66.88</td><td>34.38</td><td>24.40</td></tr><tr><td>AWQ</td><td>78.13</td><td>75.09</td><td>49.21</td><td>59.35</td><td>68.64</td><td>39.25</td><td>30.80</td></tr><tr><td>GPTQ</td><td>78.23</td><td>74.73</td><td>49.21</td><td>59.35</td><td>68.86</td><td>38.48</td><td>31.00</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>56.48</td><td>50.18</td><td>27.35</td><td>49.57</td><td>29.29</td><td>19.28</td><td>12.00</td></tr><tr><td>Wanda</td><td>51.13</td><td>51.99</td><td>26.11</td><td>51.07</td><td>27.65</td><td>19.71</td><td>11.60</td></tr><tr><td>AWQ</td><td>76.51</td><td>70.40</td><td>48.19</td><td>61.48</td><td>66.37</td><td>36.60</td><td>29.40</td></tr><tr><td>GPTQ</td><td>77.49</td><td>70.76</td><td>48.26</td><td>57.85</td><td>66.20</td><td>39.16</td><td>25.60</td></tr><tr><td rowspan="9">Llama3.2</td><td>0.00</td><td>Full Precision</td><td>57.28</td><td>47.29</td><td>43.73</td><td>56.75</td><td>58.12</td><td>31.31</td><td>24.80</td></tr><tr><td rowspan="4">0.5</td><td>SparseGPT</td><td>64.68</td><td>54.15</td><td>38.27</td><td>55.72</td><td>58.00</td><td>28.33</td><td>21.60</td></tr><tr><td>Wanda</td><td>62.63</td><td>53.79</td><td>35.67</td><td>55.8</td><td>54.76</td><td>24.57</td><td>18.40</td></tr><tr><td>AWQ</td><td>55.26</td><td>48.38</td><td>43.64</td><td>56.27</td><td>57.58</td><td>31.57</td><td>23.80</td></tr><tr><td>GPTQ</td><td>57.22</td><td>46.93</td><td>43.8</td><td>56.99</td><td>58.12</td><td>31.48</td><td>24.80</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>55.05</td><td>53.07</td><td>26.75</td><td>50.2</td><td>32.24</td><td>19.2</td><td>14.20</td></tr><tr><td>Wanda</td><td>41.04</td><td>52.71</td><td>26.08</td><td>49.49</td><td>27.36</td><td>20.05</td><td>12.40</td></tr><tr><td>AWQ</td><td>56.21</td><td>47.29</td><td>43.26</td><td>56.51</td><td>58.21</td><td>30.89</td><td>24.60</td></tr><tr><td>GPTQ</td><td>66.76</td><td>54.87</td><td>42.62</td><td>54.7</td><td>54.97</td><td>31.66</td><td>25.40</td></tr><tr><td rowspan="9">SmolLM2</td><td>0.00</td><td>Full Precision</td><td>75.81</td><td>70.04</td><td>51.75</td><td>62.75</td><td>66.29</td><td>39.42</td><td>28.00</td></tr><tr><td rowspan="4">0.5</td><td>SparseGPT</td><td>69.85</td><td>53.43</td><td>44.92</td><td>62.27</td><td>68.14</td><td>37.03</td><td>26.80</td></tr><tr><td>Wanda</td><td>67.16</td><td>55.23</td><td>42.91</td><td>59.83</td><td>67.76</td><td>32.76</td><td>25.00</td></tr><tr><td>AWQ</td><td>75.78</td><td>69.31</td><td>51.71</td><td>63.14</td><td>65.78</td><td>39.51</td><td>28.80</td></tr><tr><td>GPTQ</td><td>76.30</td><td>68.23</td><td>51.68</td><td>62.90</td><td>66.04</td><td>39.59</td><td>28.20</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>38.56</td><td>51.99</td><td>27.16</td><td>49.64</td><td>31.31</td><td>18.60</td><td>12.60</td></tr><tr><td>Wanda</td><td>37.80</td><td>53.79</td><td>26.10</td><td>49.72</td><td>28.45</td><td>19.54</td><td>14.60</td></tr><tr><td>AWQ</td><td>75.93</td><td>70.40</td><td>50.80</td><td>62.19</td><td>67.85</td><td>38.14</td><td>27.20</td></tr><tr><td>GPTQ</td><td>76.27</td><td>70.76</td><td>50.89</td><td>60.77</td><td>64.94</td><td>38.74</td><td>29.00</td></tr><tr><td rowspan="9">Phi3.5</td><td>0.00</td><td>Full Precision</td><td>85.84</td><td>80.87</td><td>56.32</td><td>68.59</td><td>76.14</td><td>50.68</td><td>33.60</td></tr><tr><td rowspan="4">0.5</td><td>SparseGPT</td><td>82.63</td><td>70.76</td><td>52.02</td><td>67.64</td><td>74.58</td><td>46.16</td><td>35.80</td></tr><tr><td>Wanda</td><td>83.00</td><td>71.84</td><td>50.45</td><td>67.01</td><td>77.65</td><td>47.35</td><td>33.00</td></tr><tr><td>AWQ</td><td>85.72</td><td>81.23</td><td>56.34</td><td>68.11</td><td>75.63</td><td>50.94</td><td>34.20</td></tr><tr><td>GPTQ</td><td>86.09</td><td>80.87</td><td>56.37</td><td>68.11</td><td>76.30</td><td>50.60</td><td>34.20</td></tr><tr><td rowspan="4">0.75</td><td>SparseGPT</td><td>62.26</td><td>55.6</td><td>29.12</td><td>51.38</td><td>30.89</td><td>20.65</td><td>11.60</td></tr><tr><td>Wanda</td><td>51.35</td><td>52.35</td><td>26.12</td><td>49.88</td><td>28.07</td><td>19.11</td><td>14.80</td></tr><tr><td>AWQ</td><td>85.63</td><td>79.78</td><td>54.96</td><td>66.85</td><td>75.04</td><td>50.17</td><td>31.80</td></tr><tr><td>GPTQ</td><td>85.63</td><td>80.87</td><td>54.69</td><td>66.61</td><td>76.56</td><td>49.74</td><td>34.80</td></tr></table>
+
+Table 10: Performance (%) on English downstream benchmarks: BoolQ, RTE, HellaSwag, WinoGrande, ARC-e, ARC-c, and OBQA, across models, sparsity levels, and pruning/quantization methods.
+
+<table><tr><td>Model</td><td>SNR</td><td>Error (x1e-4)</td><td>en</td><td>ar</td><td>hi</td><td>zh</td><td>th</td><td>de</td><td>es</td></tr><tr><td>R1 Original</td><td>-</td><td>-</td><td>29.79</td><td>159.39</td><td>11.02</td><td>41.37</td><td>40.67</td><td>98.58</td><td>85.36</td></tr><tr><td>Wanda (Structured 2:4)</td><td>7.08</td><td>3.97</td><td>148.92</td><td>4864.43</td><td>54.45</td><td>514.27</td><td>666.37</td><td>1024.4</td><td>738.42</td></tr><tr><td>Wanda (Unstructured 0.5)</td><td>11.11</td><td>1.59</td><td>47.75</td><td>451.78</td><td>16.85</td><td>114.34</td><td>170.08</td><td>206.75</td><td>164.00</td></tr><tr><td>SparseGPT (Structured 2:4)</td><td>9.64</td><td>2.19</td><td>61.14</td><td>664.81</td><td>30.96</td><td>232.78</td><td>258.91</td><td>367.24</td><td>254.92</td></tr><tr><td>SparseGPT (Unstructured 0.5)</td><td>13.61</td><td>0.86</td><td>44.41</td><td>379.12</td><td>16.26</td><td>117.87</td><td>306.32</td><td>193.78</td><td>162.97</td></tr></table>
+
+Table 11: Compression performance and multilingual perplexity of DeepSeek R1 Distil Qwen 1.5B. The better result between structured and unstructured is highlighted in bold. Due to rebuttal time constraints, results for the other four models will be provided in the final version.  
+
+<table><tr><td>Model</td><td>BoolQ</td><td>RTE</td><td>HellaSwag</td><td>WinoGrande</td><td>ARC-e</td><td>ARC-c</td><td>OBQA</td></tr><tr><td>R1 Original</td><td>51.10</td><td>46.93</td><td>36.13</td><td>51.07</td><td>42.93</td><td>27.56</td><td>18.60</td></tr><tr><td>R1 Wanda (Structured 2:4)</td><td>62.05</td><td>51.62</td><td>29.89</td><td>52.09</td><td>43.48</td><td>23.46</td><td>15.60</td></tr><tr><td>R1 Wanda (Unstructured)</td><td>64.56</td><td>57.76</td><td>33.27</td><td>54.14</td><td>54.08</td><td>28.24</td><td>18.80</td></tr><tr><td>R1 SparseGPT (Structured 2:4)</td><td>62.75</td><td>52.71</td><td>32.01</td><td>52.88</td><td>49.45</td><td>24.40</td><td>16.60</td></tr><tr><td>R1 SparseGPT (Unstructured)</td><td>66.21</td><td>55.96</td><td>33.95</td><td>55.72</td><td>54.34</td><td>29.44</td><td>19.80</td></tr></table>
+
+Table 12: Performance (%) of DeepSeek R1 Distill Qwen 1.5B on English downstream benchmarks. The better result between structured and unstructured is highlighted in bold. Due to rebuttal time constraints, results for the other four models will be provided in the final version.  
+
+<table><tr><td>Model</td><td>BoolQ</td><td>RTE</td><td>HellaSwag</td><td>WinoGrande</td><td>ARC-e</td><td>ARC-c</td><td>OBQA</td></tr><tr><td>Uncompressed Qwen-0.5B</td><td>67.95</td><td>62.45</td><td>40.63</td><td>55.88</td><td>65.53</td><td>30.89</td><td>24.20</td></tr><tr><td>Llama3.2-1B (0.5 Wanda)</td><td>62.63</td><td>53.79</td><td>35.67</td><td>55.80</td><td>54.76</td><td>24.57</td><td>18.40</td></tr><tr><td>Llama3.2-1B (0.5 SparseGPT)</td><td>64.68</td><td>54.15</td><td>38.27</td><td>55.72</td><td>58.00</td><td>28.33</td><td>21.60</td></tr><tr><td>Llama3.2-1B (AWQ 8)</td><td>55.26</td><td>48.38</td><td>43.64</td><td>56.27</td><td>57.58</td><td>31.57</td><td>23.80</td></tr><tr><td>Llama3.2-1B (GPTQ 8)</td><td>57.22</td><td>46.93</td><td>43.80</td><td>56.99</td><td>58.12</td><td>31.48</td><td>24.80</td></tr><tr><td>SmolLM2-1.7B (0.75 Wanda)</td><td>37.80</td><td>53.79</td><td>26.10</td><td>49.72</td><td>28.45</td><td>19.54</td><td>14.60</td></tr><tr><td>SmolLM2-1.7B (0.75 SparseGPT)</td><td>38.56</td><td>51.99</td><td>27.16</td><td>49.64</td><td>31.31</td><td>18.60</td><td>12.60</td></tr><tr><td>SmolLM2-1.7B (AWQ 4)</td><td>75.93</td><td>70.40</td><td>50.80</td><td>62.19</td><td>67.85</td><td>38.14</td><td>27.20</td></tr><tr><td>SmolLM2-1.7B (GPTQ 4)</td><td>76.27</td><td>70.76</td><td>50.89</td><td>60.77</td><td>64.94</td><td>38.74</td><td>29.00</td></tr></table>
+
+Table 13: Downstream performance of models in comparable sizes. The column-wise best is highlighted in bold.

revisitingpruningvsquantizationforsmalllanguagemodels/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:944324140da6b8e9950dc18afed4d3a30bc73597367744555c70e55257dc4f0e
+size 1493937

revisitingpruningvsquantizationforsmalllanguagemodels/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a70bfecc5042d1dd39f04297cc4f66731d3960b8af7ab541c7cc4d191e53649a
+size 419779

revpragrevealingpoisoningattacksinretrievalaugmentedgenerationthroughllmactivationanalysis/5997470f-947c-4aed-a4bc-c329cfce64b9_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1c74c1ef6f6eef016b82b04289e8f6552428e91434fffb7c032745d0925c653
+size 85508

revpragrevealingpoisoningattacksinretrievalaugmentedgenerationthroughllmactivationanalysis/5997470f-947c-4aed-a4bc-c329cfce64b9_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d350b34ea5a13c7538b9aecf8e64916f635bdb24bc9b03506d3fe1e7f82521de
+size 104157

revpragrevealingpoisoningattacksinretrievalaugmentedgenerationthroughllmactivationanalysis/5997470f-947c-4aed-a4bc-c329cfce64b9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82d248010980c7df7a50d31247361d96cb4b6bd0dba9103bc728a9833e38c48e
+size 2646896

revpragrevealingpoisoningattacksinretrievalaugmentedgenerationthroughllmactivationanalysis/full.md

@@ -0,0 +1,341 @@
+# RevPRAG: Revealing Poisoning Attacks in Retrieval-Augmented Generation through LLM Activation Analysis
+
+Xue Tan $^{1,3}$ , Hao Luan $^{1,3}$ , Mingyu Luo $^{1,3}$ , Xiaoyan Sun $^{2}$ , Ping Chen $^{3}$ , Jun Dai $^{2}$
+
+$^{1}$ School of Computer Science, Fudan University, Shanghai, China
+
+$^{2}$ Department of Computer Science, Worcester Polytechnic Institute, MA, USA
+
+$^{3}$ Institute of Big Data, Fudan University, Shanghai, China
+
+Correspondence: pchen@fudan.edu.cn, xsun7@wpi.edu, jdai@wpi.edu
+
+# Abstract
+
+Retrieval-Augmented Generation (RAG) enriches the input to LLMs by retrieving information from the relevant knowledge database, enabling them to produce responses that are more accurate and contextually appropriate. It is worth noting that the knowledge database, being sourced from publicly available channels such as Wikipedia, inevitably introduces a new attack surface. RAG poisoning attack involves injecting malicious texts into the knowledge database, ultimately leading to the generation of the attacker's target response (also called poisoned response). However, there are currently limited methods available for detecting such poisoning attacks. We aim to bridge the gap in this work by introducing RevPRAG, a flexible and automated detection pipeline that leverages the activations of LLMs for poisoned response detection. Our investigation uncovers distinct patterns in LLMs' activations when generating poisoned responses versus correct responses. Our results on multiple benchmarks and RAG architectures show our approach can achieve a  $98\%$  true positive rate, while maintaining a false positive rate close to  $1\%$ .
+
+# 1 Introduction
+
+Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) has emerged as an effective solution that leverages retrievers to incorporate external databases, enriching the knowledge of LLMs and ultimately enabling the generation of up-to-date and accurate responses. RAG comprises three components: knowledge database, retriever, and LLM. Fig. 1 visualizes an example of RAG. The knowledge database consists of a large amount of texts collected from sources such as latest Wikipedia entries (Thakur et al., 2021), new articles (Soboroff et al., 2018) and financial documents (Loukas et al., 2023). The retriever is primarily responsible for retrieving the texts that are most related to the user's query from the knowledge database.
+
+These texts will later be fed to LLM as a part of the prompt to generate responses (e.g., "Everest") for users' queries (e.g., "What is the name of the highest mountain?"). Due to RAG's powerful knowledge integration capabilities, it has demonstrated impressive performance across a range of QA-like knowledge-intensive tasks (Lazaridou et al., 2022; Jeong et al., 2024).
+
+RAG poisoning refers to the act of injecting malicious or misleading content into the knowledge database, contaminating the retrieved texts and ultimately leading the LLM to produce the attacker's desired response (e.g., the target answer could be "Fuji" when the target question is "What is the name of the highest mountain?"). This attack leverages the dependency between LLMs and the knowledge database, transforming the database into a new attack surface to facilitate poisoning. PoisonedRAG (Zou et al., 2024) demonstrates the feasibility of RAG poisoning by injecting a small amount of maliciously crafted texts into the knowledge database utilized by RAG. The rise of such attacks has drawn significant attention to the necessity of designing robust and resilient RAG systems. For example, INSTRUCTRAG (Wei et al., 2024) utilizes LLMs to analyze how to extract correct answers from noisy retrieved documents; RobustRAG (Xiang et al., 2024) introduces multiple LLMs to generate answers from the retrieved texts, and then aggregates the responses. However, the aforementioned defense methods necessitate the integration of additional large models, incurring considerable overheads. Meanwhile, it is difficult to promptly assess whether the current response of RAG is trustworthy or not.
+
+In our work, we shift our focus to leverage the intrinsic properties of LLMs for detecting RAG poisoning, rather than relying on external models. Our view is that if we can accurately determine whether a RAG's response is correct or poisoned, we can effectively thwart RAG poisoning
+
+![](images/1e4586d4e8756591be01e80594315c394ee14282ef8d31c2618d9fc50cad2c36.jpg)  
+Figure 1: Visualization of RAG.
+
+attacks. We attempt to observe LLM's answer generation process to determine whether the response is compromised or not. It is worth noting that our focus is not on detecting malicious inputs to LLMs, as we consider the consequences of malicious responses to be far more detrimental and indicative of an attack. The growing body of research on using activations to explain and control LLM behavior (Ferrando et al., 2024; He et al., 2024) provides us inspiration. Specifically, we empirically analyze the activations of the final token in the input sequence across all layers of the LLM. Our findings demonstrate that the model exhibits distinguishable activation patterns when generating correct versus poisoned responses. Based on this, we propose a systematic and automated detection pipeline, namely RevPRAG, which consists of three key components: poisoned data collection, LLM activation collection and preprocessing, and the detection model design. It is important to note that this detection method will not alter the RAG workflow or weaken its performance, thereby offering superior adversarial robustness compared to methods that rely solely on filtering retrieved texts.
+
+To evaluate our approach, we systematically demonstrate the effectiveness of RevPRAG across various LLM architectures, including GPT2-XL-1.5B, Llama2-7B, Mistral-7B, Llama3-8B, and Llama2-13B. RevPRAG performs consistently well, achieving over  $98\%$  true positive rate across different datasets.
+
+Our contributions can be summarized as follows:
+
+1. We uncover distinct patterns in LLMs' activations when RAG generates correct responses versus poisoned ones.  
+2. We introduce RevPRAG, a novel and automated pipeline for detecting whether a RAG's response is poisoned or not. To address emerging RAG poisoning attacks, RevPRAG allows
+
+new datasets to be constructed accordingly for training the model, enabling effective detection of new threats.
+
+3. Our model has been empirically validated across various LLM architectures and retrievers, demonstrating over  $98\%$  accuracy on our custom-collected detection dataset.
+
+# 2 Background and Related Work
+
+# 2.1 Retrieval Augmented Generation
+
+RAG comprises three components: knowledge database, retriever, and LLM. As illustrated in Fig. 1, RAG consists of two main steps: retrieval step and generation step. In the retrieval step, the retriever acquires the top  $k$  most relevant pieces of knowledge for the query  $q$ . First, we employ two encoders,  $E_{q}$  and  $E_{p}$ , which can either be identical or radically different. Encoder  $E_{q}$  is responsible for transforming the user's query  $q$  into an embedding vector  $E_{q}(q)$ , while encoder  $E_{p}$  is designed to convert all the information  $p_{i}$  in the knowledge database into embedding vectors  $E_{p}(p_{i})$ . For each  $E_{p}(p_{i})$ , the similarity with the query  $E_{q}(q)$  is computed using  $\mathrm{sim}(E_q(q),E_p(p_i))$ , where  $\mathrm{sim}(\cdot ,\cdot)$  quantifies the similarity between two embedding vectors, such as cosine similarity or the dot product. Finally, the top  $k$  most relevant pieces are selected as the external knowledge  $C_q$  for the query  $q$ . The generation step is to generating a response  $\mathrm{LLM}(q,\mathcal{C}_q)$  based on the query  $q$  and the relevant information  $C_q$ . First, we combine the query  $q$  and the external knowledge  $C_q$  using a standard prompt (see Fig. 6 for the complete prompt). Taking advantage of such a prompt, the LLM generates an answer  $\mathrm{LLM}(q,\mathcal{C}_q)$  to the query  $q$ . Therefore, RAG is a significant accomplishment, as it addresses the limitations of LLMs in acquiring up-to-date and domain-specific information.
+
+# 2.2 Retrieval Corruption Attack
+
+Due to the growing attention on RAG, attacks on RAG have also been widely studied. RAG can improperly generate answers that are severely impacted or compromised once the knowledge database is contaminated (Zou et al., 2024; Xue et al., 2024; Jiao et al., 2024). Specifically, an attacker can inject a small amount of malicious information onto a website, which is then retrieved by RAG (Greshake et al., 2023). PoisonedRAG (Zou et al., 2024) injects malicious text into the knowledge database, and formalizes the knowledge poisoning attack as an optimization problem, thereby enabling the LLM to generate target responses selected by the attacker. GARAG (Cho et al., 2024) was introduced to provide low-level perturbations to RAG. PRCAP (Zhong et al., 2023) injects adversarial samples into the knowledge database, where these samples are generated by perturbing discrete tokens to enhance their similarity with a set of training queries. These methods have yielded striking attack results, and in our work, we have selected several state-of-the-art attack methods as our base attacks on RAG.
+
+# 2.3 The Robustness of RAG
+
+Efforts have been made to develop defenses in response to poisoning attacks and noise-induced disruptions. RobustRAG (Xiang et al., 2024) mitigates the impact of poisoned texts through a voting mechanism, while INSTRUCTRAG (Wei et al., 2024) explicitly learns the denoising process to address poisoned and irrelevant information. Other approaches to enhance robustness include prompt design (Cho et al., 2023; Press et al., 2023), plug-in models (Baek et al., 2023), and specialized models (Yoran et al., 2023; Asai et al., 2023). However, these methods may, on one hand, rely on additional LLMs, leading to significant overhead. On the other hand, they primarily focus on defense mechanisms before the LLM generates a response, making it challenging for these existing approaches to detect poisoning attacks in real-time while the LLM is generating the response (Athalye et al., 2018; Bryniarski et al., 2021; Carlini and Wagner, 2017; Carlini, 2023; Tramer et al., 2020). LLM Factoscope (He et al., 2024) is a runtime detection tool that leverages the internal states of LLMs, such as activation maps, output rankings, and top- $k$  probabilities, to identify factual inaccuracies caused by model hallucinations. While Factoscope is effective at detecting hallucinations in general LLMs, it
+
+![](images/d582cded4e0b806623bbe685e369c3dd4c0aeacca74cec9be739352f42da6479.jpg)  
+Figure 2: t-SNE visualizations of activations for correct and poisoned responses.
+
+![](images/bc8af53c03fac6db5d1b68347786a8e6e58e7a29b46004d94336d01969b79594.jpg)
+
+is not designed to address RAG poisoning attacks, which result from manipulations of the external knowledge base rather than internal model errors. Its complex architecture with multiple sub-models makes it less suitable for latency-sensitive RAG applications. In this work, we present RevPRAG, a method that addresses these gaps by: (1) focusing on RAG-specific poisoning attacks and conducting extensive tests to validate its effectiveness in detecting such attacks (Section 5), (2) using a lightweight, activation-based pipeline optimized for real-time detection of whether an RAG response is trustworthy (Section 5.6), and (3) evaluations show that our performance (Section 5.2) and efficiency (Section 5.6) surpass those of Factoscope.
+
+# 3 Preliminary
+
+# 3.1 Threat Model
+
+Attacker's goal. We assume that the attacker preselects a target question set  $Q$ , consisting of  $q_{1}, q_{2}, \dots, q_{n}$ , and the corresponding target answer set  $A$ , represented as  $a_{1}, a_{2}, \dots, a_{n}$ . The attacker's goal is to compromise the RAG system by contaminating the retrieval texts, thereby manipulating the LLM to generate the target response  $a_{i}$  for each query  $q_{i}$ . For example, the attacker's target question  $q_{i}$  is "What is the name of the highest mountain?", with the target answer being "Fuji".
+
+Attacker's capabilities. We assume that an attacker can inject  $m$  poisoned texts  $P$  for each target question  $q_{i}$ , represented as  $p_{i}^{1}, p_{i}^{2}, \ldots, p_{i}^{m}$ . The attacker does not possess knowledge of the LLM utilized by the RAG, but has white-box access to the RAG retriever. This assumption is reasonable, as many retrievers are openly accessible on platforms like HuggingFace. The poisoned texts can be integrated into the RAG's knowledge database through two ways: the attacker publishing the malicious content on open platforms like Wikipedia, or utilizing data collection agencies to disseminate the poisoned texts.
+
+# 3.2 Rationale
+
+The activations of LLMs represent input data at varying layers of abstraction, enabling the model to progressively extract high-level semantic information from low-level features. The extensive information encapsulated in these activations comprehensively reflects the entire decision-making process of the LLM. The activations has been applied to factual verification of the output content (He et al., 2024) and detection of task drift (Abdelnabi et al., 2024). Due to the fact that LLM produces different activations when generating varying responses, we hypothesize that LLM will also exhibit distinct activations when generating poisoned responses compared to correct ones. Fig. 2 presents the visualizations of activations for correct and poisoned responses using t-SNE (t-Distributed Stochastic Neighbor Embedding). It visualizes the mean activations across all layers for two LLMs, Mistral-7B and Llama2-7B, on the Natural Questions dataset. This clearly demonstrates the distinguishability between the two types of responses, to some extent, supports our conjecture.
+
+# 4 Methodology
+
+# 4.1 Approach Overview
+
+As illustrated in Fig. 3, we introduce RevPRAG, a pipeline designed to leverage LLM activations for detecting knowledge poisoning attacks in RAG systems. It contains three major modules: poisoning data collection, activation collection and preprocessing, and RevPRAG detection model design. Fig. 4 demonstrates a practical application
+
+of RevPRAG for verifying the poisoning status of LLM outputs. Given a user prompt such as "What is the name of the highest mountain?", the LLM will provide a response. Meanwhile the activations generated by the LLM will be collected and analyzed in RevPRAG. If the model classify the activations as poisoned behavior, it will flag the corresponding response (such as "Fuji") as a poisoned response. Otherwise, it will confirm the response (e.g. "Everest") as the correct answer.
+
+# 4.2 Poisoning Data Collection
+
+Our method seeks to extract the LLM's activations that capture the model's generation of a specific poisoned response triggered by receiving poisoned texts at a given point in time. Therefore, we first need to implement poisoning attacks on RAG that can mislead the LLM into generating target poisoned responses. There are three components in RAG: knowledge database, retriever, and LLM. In order to successfully carry out a poisoning attack on RAG and compel the LLM to generate the targeted poisoned response, the initial step is to craft a sufficient amount of poisoned texts and inject them into the knowledge database. In this paper, in order to create effective poisoned texts for our primary focus on detecting poisoning attacks, we employ three state-of-the-art strategies (i.e., PoisonedRAG (Zou et al., 2024), GARAG (Cho et al., 2024), and PAPRAG (Zhong et al., 2023)) for generating poisoned texts and increasing the similarity between the poisoned texts and the queries, to raise the likelihood that the poisoned texts would be selected by the retriever. A detailed introduction of
+
+![](images/165a4a4eb434aa241387f6d028873a364d828e61c71a5996e07290406104208b.jpg)  
+Figure 3: The workflow of RevPRAG.
+
+![](images/7cfa1bded5d297b145e2e69f90754fa9a65cfae3434efc6b02e746de3569c0ad.jpg)  
+Figure 4: An instance of using RevPRAG.
+
+these methods can be found in Section A.2. The retrieved texts and the question are combined into a new prompt, following the format in (Zou et al., 2024) (see Fig. 6 in Section A.3), for LLM answer generation.
+
+# 4.3 Activation Collection and Processing
+
+For an LLM input sequence  $X = (t_{1}, t_{2}, \dots, t_{n})$ , we extract the activations  $Act_{n}$  for the last token  $x_{n}$  in the input across all layers in the LLM as a summary of the context. The activations  $Act_{n}$  contain the inner representations of the LLM's knowledge related to the input. When the LLM generates a response based on a question, it traverses through all layers, retrieving knowledge relevant to the input to produce an answer (Meng et al., 2023). We collect two types of activations: correct activations (labeled as 1), obtained when the LLM retrieves accurate content and generates the correct response; and poisoned activations (labeled as 0), obtained when the LLM retrieves poisoned content and produces the attacker's target response.
+
+We introduce normalization of the activations for effective integration into the training process. We calculate the mean  $\mu$  and standard deviation  $\sigma$  of the activations across all instances in the dataset. Then, we use the obtained  $\mu$  and  $\sigma$  to normalize the activations with the formula:
+
+$$
+A c t _ {n} ^ {n o r} = \left(A c t _ {n} - \mu\right) / \sigma . \tag {1}
+$$
+
+# 4.4 RevPRAG Model Design
+
+After collecting and preprocessing the activation dataset, we partition it into a training set  $D_{train}$ , a test set  $D_{test}$ , and a support set  $S$  to facilitate the construction and evaluation of the probe model. Drawing inspiration from few-shot learning and Siamese networks, the proposed RevPRAG model is designed to effectively distinguish between clean and poisoned responses, while demonstrating strong generalization capabilities even under limited data conditions. To efficiently capture
+
+both intra-layer and inter-layer relationships within the LLM, we employ Convolutional Neural Networks (CNNs) based on the ResNet18 architecture (He et al., 2016). Additionally, we adopt a triplet network structure, in which three subnetworks with shared architecture and weights are used to learn task embeddings, as illustrated in Fig. 3.
+
+During training, we employ the triplet margin loss (Schroff et al., 2015), a commonly used approach for tasks where it is difficult to distinguish similar instances. The training data is randomly divided into triplets consisting of an anchor instance  $x_{a}$ , a positive instance  $x_{p}$ , and a negative instance  $x_{n}$ , where the anchor and positive belong to the same class, while the anchor and negative come from different classes. The triplet margin loss function is formally defined as:
+
+$$
+\begin{array}{l} L = \max  \left(\operatorname {D i s t} \left(x _ {a}, x _ {p}\right) - \operatorname {D i s t} \left(x _ {a}, x _ {n}\right) \right. \\ + \operatorname {m a r g i n}, 0), \tag {2} \\ \end{array}
+$$
+
+where  $\mathrm{Dist}(\cdot, \cdot)$  denotes a distance metric (typically the Euclidean distance), and margin is a positive constant. The training objective is to encourage the RevPRAG embedding model to output closer embedding vectors for any  $x_{a}$  and  $x_{p}$ , but farther for any  $x_{a}$  and  $x_{n}$ .
+
+At test time, given a test sample  $x_{t}$ , we compute the distance between its embedding and the embedding of the support sample  $x_{s}, x_{s} \in S$ . The support set  $S$  refers to a dataset comprising labeled data, denoted as  $\{x_{s_1}, \dots, x_{s_n}\}$ , and corresponding labels are  $\{T_{x_{s_1}}, \dots, T_{x_{s_n}}\}$ . It provides a reference for comparison and classification of new, unseen test data. The main purpose of the support set is to help determine labels for the test data. The label of the test data  $x_{t}$  will be determined according to the label of the support sample  $x_{s}$  that is closest to it. That is,  $x_{t}$  is assigned the label of  $x_{s}$ , meaning  $T_{x_t} = T_{x_s}$ , where  $x_{s} = \text{argmin}_{i} \text{Dist}(x_{t}, x_{s_i})$ . Here,  $x_{s}$  is the nearest support data to the test data  $x_{t}$ .
+
+# 5 Evaluation
+
+# 5.1 Experimental Setup
+
+RAG Setup. RAG comprises three key components: knowledge database, retriever, and LLM. The setup is shown below:
+
+- Knowledge Database: We leverage three representative benchmark question-answering
+
+datasets in our evaluation: Natural Questions (NQ) (Kwiatkowski et al., 2019), HotpotQA (Yang et al., 2018), MS-MARCO (Bajaj et al., 2016). Please note that RevPRAG can be expanded to cover poisoning attacks towards any other datasets used for RAG systems, not limited to the datasets used in this paper. The detailed usage instructions for the dataset are provided in Section A.1.
+
+- Retriever: In our experiments, we evaluate four state-of-the-art dense retrieval models: Contriever (Izacard et al., 2021) (pre-trained), Contriever-ms (fine-tuned on MS-MARCO) (Izacard et al., 2021), DPR-mul (Karpukhin et al., 2020) (trained on multiple datasets), and ANCE (Xiong et al., 2020) (trained on MS-MARCO).  
+- LLM: Our experiments are conducted on several popular LLMs, each with distinct architectures and characteristics, including GPT2-XL 1.5B (Radford et al., 2019), Llama2-7B (Touvron et al., 2023), Llama2-13B, Mistral-7B (Jiang et al., 2023), and Llama3-8B.
+
+Unless otherwise specified, we adopt the following default settings: HotpotQA as the knowledge base, Contriever as the retriever, GPT2-XL 1.5B as the LLM, and 100 support samples. Moreover, we use the dot product between the embedding vectors of a question and a text to measure their similarity. Poisoned texts are generated following PoisonedRAG (Zou et al., 2024). Consistent with prior work (Lewis et al., 2020), we retrieve the 5 most similar texts from the knowledge database to serve as context for a given question.
+
+Baselines. We compared RevPRAG with five existing methods, and although they were not specifically designed for detecting RAG poisoning attacks, we investigated their potential applications in this domain. CoS (Li et al., 2024) is a black-box approach that guides the LLM to generate de
+
+tailed reasoning steps for the input, subsequently scrutinizing the reasoning process to ensure consistency with the final answer. MDP (Xi et al., 2024) is a white-box method that exploits the disparity in masking sensitivity between poisoned and clean samples. LLM Factoscope (He et al., 2024) leverages the internal states of LLMs to detect hallucinations, and we investigate its use for identifying poisoning attacks in RAG systems. Both RoBERTa (Pan et al., 2023) and Discern (Hong et al., 2024) employ an additional discriminator to distinguish whether the content retrieved by RAG consists of accurate documents or those that contradict factual information.
+
+# Evaluation Metrics.
+
+- The True Positive Rate (TPR), which measures the proportion of effectively poisoned responses that are successfully detected. A higher TPR signifies better detection performance for poisoned responses, with a correspondingly lower rate of missed detections (i.e., lower false negative rate).  
+- The False Positive Rate (FPR), which quantifies the proportion of correct responses that are misclassified as being caused by poisoning attacks. A lower FPR indicates fewer false positives for correct answers, minimizing disruption to the normal operation of RAG. Our goal is to detect poisoned responses as effectively as possible while minimizing the impact on RAG's normal functionality, which is why we have selected these two metrics.
+
+# 5.2 Overall Results
+
+RevPRAG achieves high TPRs and low FPRs. Table 1 shows the TPRs and FPRs of RevPRAG on three datasets. We have the following observations from the experimental results. First, RevPRAG achieved high TPRs consistently on different datasets and LLMs when injecting five
+
+Table 1: RevPRAG achieved high TPRs and low FPRs on three datasets for RAG with five different LLMs.  
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Metrics</td><td colspan="5">LLMs of RAG</td></tr><tr><td>GPT2-XL 1.5B</td><td>Llama2-7B</td><td>Mistral-7B</td><td>Llama3-8B</td><td>Llama2-13B</td></tr><tr><td rowspan="2">NQ</td><td>TPR</td><td>0.982</td><td>0.994</td><td>0.985</td><td>0.986</td><td>0.989</td></tr><tr><td>FPR</td><td>0.006</td><td>0.006</td><td>0.019</td><td>0.009</td><td>0.019</td></tr><tr><td rowspan="2">HotpotQA</td><td>TPR</td><td>0.972</td><td>0.985</td><td>0.977</td><td>0.973</td><td>0.970</td></tr><tr><td>FPR</td><td>0.016</td><td>0.061</td><td>0.022</td><td>0.017</td><td>0.070</td></tr><tr><td rowspan="2">MS-MARCO</td><td>TPR</td><td>0.988</td><td>0.989</td><td>0.999</td><td>0.978</td><td>0.993</td></tr><tr><td>FPR</td><td>0.007</td><td>0.012</td><td>0.001</td><td>0.011</td><td>0.025</td></tr></table>
+
+Table 2: RevPRAG achieved high TPRs and low FPRs on HotpotQA for RAG with four different retrievers.  
+
+<table><tr><td rowspan="2">Attack</td><td rowspan="2">Metrics</td><td colspan="3">LLMs of RAG</td></tr><tr><td>GPT2-XL 1.5B</td><td>Llama2-7B</td><td>Mistral-7B</td></tr><tr><td rowspan="2">Contriever</td><td>TPR</td><td>0.972</td><td>0.985</td><td>0.977</td></tr><tr><td>FPR</td><td>0.016</td><td>0.061</td><td>0.022</td></tr><tr><td rowspan="2">Contriever-ms</td><td>TPR</td><td>0.987</td><td>0.983</td><td>0.998</td></tr><tr><td>FPR</td><td>0.057</td><td>0.018</td><td>0.012</td></tr><tr><td rowspan="2">DPR-mul</td><td>TPR</td><td>0.979</td><td>0.966</td><td>0.999</td></tr><tr><td>FPR</td><td>0.035</td><td>0.075</td><td>0.001</td></tr><tr><td rowspan="2">ANCE</td><td>TPR</td><td>0.978</td><td>0.981</td><td>0.993</td></tr><tr><td>FPR</td><td>0.042</td><td>0.028</td><td>0.023</td></tr></table>
+
+poisoned texts into the knowledge database. For instance, RevPRAG achieved  $98.5\%$  (on NQ),  $97.7\%$  (on HotpotQA), and  $99.9\%$  (on MS-MARCO) TPRs for RAG with Mistral-7B. Our experimental results show that assessing whether the output of a RAG system is correct or poisoned based on the activations of LLMs is both highly feasible and reliable (i.e., capable of achieving exceptional accuracy). Second, RevPRAG achieves low FPRs under different settings, e.g., close to  $1\%$  in nearly all cases. This result indicates that our approach not only maximizes the detection of poisoned responses but also maintains a low false positive rate, significantly reducing the risk of misclassifying correct answers as poisoned.
+
+We also conduct experiments on different retrievers. Table 2 shows that our approach consistently achieved high TPRs and low FPRs across RAG with various retrievers and LLMs. For instance, RevPRAG achieves  $97.2\%$  (with Contriever),  $98.7\%$  (with Contriever-ms),  $97.9\%$  (with DPR-mul),  $97.8\%$  (with ANCE) TPRs alongside
+
+1.6% (with Contriever), 5.7% (with Contriever-ms), 3.5% (with DPR-mul), and 4.2% (with ANCE) FPRs for RAG when using GPT2-XL 1.5B.
+
+RevPRAG outperforms baselines. Table 3 compares RevPRAG with baselines for RAG using Llama3-8B under the default settings. The overall results demonstrate the superiority of our approach. Meanwhile, several key observations can be drawn from the comparison. First, the limited effectiveness of CoS (Li et al., 2024) may stem from its design focus on detecting backdoor attacks in LLMs via trigger-to-output shortcuts, which differs from RAG's attack surface involving poisoned knowledge base entries. Second, MDP (Xi et al., 2024) achieves good TPRs, but it also exhibits relatively high FPRs, reaching as much as  $37.2\%$ . LLM Factoscope (He et al., 2024) leverages multiple internal states of LLMs, relying on layer-wise consistency for effective hallucination detection. However, it may not be suitable for targeted attacks like poisoning, and the use of diverse state data increases computational overhead and discriminator model complexity (Section 5.6). Input-based methods such as MDP (Xi et al., 2024), RoBERTa (Pan et al., 2023), and Discern (Hong et al., 2024) aim to detect whether the input is poisoned. In contrast, our method focuses on determining whether the responses generated by RAG are correct or poisoned, as response correctness offers a more robust signal of poisoning attacks.
+
+# 5.3 Ablation Study
+
+Different methods for generating poisoned texts. To ensure the effectiveness of the evaluation, we employ three different methods introduced by PoisonedRAG, GARAG, and PRCAP to generate the poisoned texts. The experimental results in Table 4
+
+Table 3: RevPRAG outperforms baselines.  
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Metrics</td><td colspan="6">Baselines and Our Method</td></tr><tr><td>CoS (Li et al., 2024)</td><td>MDP (Xi et al., 2024)</td><td>LLM Facto-scope (He et al., 2024)</td><td>RoBERTa (Pan et al., 2023)</td><td>Discern (Hong et al., 2024)</td><td>Ours</td></tr><tr><td rowspan="2">NQ</td><td>TPR</td><td>0.488</td><td>0.946</td><td>0.949</td><td>0.977</td><td>0.810</td><td>0.986</td></tr><tr><td>FPR</td><td>0.146</td><td>0.108</td><td>0.033</td><td>0.063</td><td>0.112</td><td>0.009</td></tr><tr><td rowspan="2">HotpotQA</td><td>TPR</td><td>0.194</td><td>0.886</td><td>0.939</td><td>0.956</td><td>0.817</td><td>0.973</td></tr><tr><td>FPR</td><td>0.250</td><td>0.372</td><td>0.021</td><td>0.018</td><td>0.101</td><td>0.017</td></tr><tr><td rowspan="2">MS-MARCO</td><td>TPR</td><td>0.771</td><td>0.986</td><td>0.945</td><td>0.946</td><td>0.795</td><td>0.978</td></tr><tr><td>FPR</td><td>0.027</td><td>0.181</td><td>0.028</td><td>0.070</td><td>0.101</td><td>0.011</td></tr></table>
+
+Table 4: The TPRs and FPRs of RevPRAG for different poisoned text generation methods on HotpotQA.  
+
+<table><tr><td rowspan="2">Attack</td><td rowspan="2">Metrics</td><td colspan="3">LLMs of RAG</td></tr><tr><td>GPT2-XL 1.5B</td><td>Llama2-7B</td><td>Mistral-7B</td></tr><tr><td rowspan="2">PoisonedRAG</td><td>TPR</td><td>0.972</td><td>0.985</td><td>0.977</td></tr><tr><td>FPR</td><td>0.016</td><td>0.061</td><td>0.022</td></tr><tr><td rowspan="2">GARAG</td><td>TPR</td><td>0.961</td><td>0.976</td><td>0.974</td></tr><tr><td>FPR</td><td>0.025</td><td>0.046</td><td>0.026</td></tr><tr><td rowspan="2">PRCAP</td><td>TPR</td><td>0.966</td><td>0.986</td><td>0.965</td></tr><tr><td>FPR</td><td>0.012</td><td>0.061</td><td>0.022</td></tr></table>
+
+show that RevPRAG consistently achieves high TPRs and low FPRs when confronted with poisoned texts generated by different strategies. For instance, RevPRAG achieved  $97.2\%$  (with GPT2-XL 1.5B),  $98.5\%$  (with Llama2-7B), and  $97.7\%$  (with Mistral-7B) TPRs for poisoned texts generated with PoisonedRAG.
+
+Table 5: The TPRs and FPRs of RevPRAG for different quantities of injected poisoned text on HotpotQA (total retrieved texts: five).  
+
+<table><tr><td rowspan="2">Quantity</td><td rowspan="2">Metrics</td><td colspan="3">LLMs of RAG</td></tr><tr><td>GPT2-XL 1.5B</td><td>Llama2-7B</td><td>Mistral-7B</td></tr><tr><td rowspan="2">five</td><td>TPR</td><td>0.972</td><td>0.985</td><td>0.977</td></tr><tr><td>FPR</td><td>0.016</td><td>0.061</td><td>0.022</td></tr><tr><td rowspan="2">four</td><td>TPR</td><td>0.976</td><td>0.977</td><td>0.986</td></tr><tr><td>FPR</td><td>0.034</td><td>0.047</td><td>0.033</td></tr><tr><td rowspan="2">three</td><td>TPR</td><td>0.963</td><td>0.986</td><td>0.995</td></tr><tr><td>FPR</td><td>0.011</td><td>0.043</td><td>0.004</td></tr><tr><td rowspan="2">two</td><td>TPR</td><td>0.971</td><td>0.995</td><td>0.991</td></tr><tr><td>FPR</td><td>0.011</td><td>0.047</td><td>0.005</td></tr><tr><td rowspan="2">one</td><td>TPR</td><td>0.970</td><td>0.988</td><td>0.989</td></tr><tr><td>FPR</td><td>0.049</td><td>0.031</td><td>0.022</td></tr></table>
+
+Quantity of injected poisoned texts. Table 5 illustrates the impact of varying quantities of poisoned text on the detection performance of RevPRAG. The more poisoned texts are injected, the higher the likelihood of retrieving them for RAG processing. From the experimental results, we observe that even with varying amounts of injected poisoned text, RevPRAG consistently achieves high TPRs and low FPRs. For example, when the total number of retrieved texts is five and the injected quantity is two, RevPRAG achieves a  $99.5\%$  TPR and a  $4.7\%$  FPR for RAG with Llama2-
+
+7B. The reason for this phenomenon is that the similarity between the retrieved poisoned texts and the query is higher than that of clean texts. Consequently, the LLM generates responses based on the content of the poisoned texts.
+
+Effects of different support set size. In RevPRAG, support data provides essential labeled and task-specific information, facilitating effective reasoning and learning under limited data conditions. We experiment with various support set sizes ranging from 50 to 250 to examine their effect on the performance of RevPRAG. The results in Fig. 5 indicate that varying the support size does not significantly impact the model's detection performance.
+
+![](images/d6b8d0c7f60ecd003cf47d5494257aab16b6a3840918cd0b6cfdbe0653321826.jpg)  
+Figure 5: Effects of support set size.
+
+![](images/109696d1dd7ee3817e64647fdb7e097582c035ecd56a1208ab5ec3adc39c242f.jpg)
+
+# 5.4 RevPRAG's Performance on Complex Open-Ended Questions
+
+In this section, we conducted a series of experiments to evaluate the performance of RevPRAG on complex, open-ended questions (e.g., "how to make relationship last?"). These questions present unique challenges due to their diverse and unstructured nature, in contrast to straightforward, closed-ended questions (e.g., "What is the name of the highest mountain?"). In our experiments, the NQ, HotpotQA, and MS-MARCO datasets primarily consist of close-ended questions. As a result, the majority of our previous experiments focused on close-ended problems, which was our default experimental setting. In this study, we utilized the advanced GPT-4o to filter and extract 3,000 open-ended questions from the HotpotQA and MS-MARCO datasets for training and testing the model. For open-ended questions, cosine similarity is employed to evaluate whether the LLM's response aligns with the attacker's target response. If the similarity surpasses a predefined threshold, it is considered indicative of a successful attack.
+
+The experimental results are shown in Table 6. We can observe that RevPRAG demonstrates excellent detection performance even on complex open
+
+Table 6: RevPRAG achieved high TPRs and low FPRs on the open-ended questions from HotpotQA and MS-MARCO datasets.  
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">Metrics</td><td colspan="4">LLMs of RAG</td></tr><tr><td>GPT2-XL 1.5B</td><td>Llama2-7B</td><td>Mistral-7B</td><td>Llama3-8B</td></tr><tr><td rowspan="2">HotpotQA</td><td>TPR</td><td>0.982</td><td>0.995</td><td>0.991</td><td>0.982</td></tr><tr><td>FPR</td><td>0.033</td><td>0.029</td><td>0.008</td><td>0.007</td></tr><tr><td rowspan="2">MS-MARCO</td><td>TPR</td><td>0.988</td><td>0.989</td><td>0.990</td><td>0.983</td></tr><tr><td>FPR</td><td>0.009</td><td>0.009</td><td>0.001</td><td>0.017</td></tr></table>
+
+ended questions. For example, RevPRAG achieved TPRs of  $99.1\%$  on HotpotQA and  $99.0\%$  on MS-MARCO, alongside FPRs of  $0.8\%$  on HotpotQA and  $0.1\%$  on MS-MARCO for RAG utilizing the Mistral-7B model.
+
+# 5.5 Effect of Real-world Natural Text Noise on Detection Performance
+
+To better approximate real-world scenarios, we injected natural textual noise (e.g., spelling and grammatical errors) into clean texts using the GPT-40 model. Noise was introduced in a controlled manner to preserve semantics while simulating realistic imperfections. Specifically,  $50\%$  of the clean texts were modified, with noise accounting for  $10\%$  of the total word count in each case. Notably, noise injection was applied only to clean texts, as poisoned texts are typically well-formed to maximize attack effectiveness and thus unlikely to contain such artifacts.
+
+As shown in Table 7, the experimental results indicate that injecting a certain amount of natural noise into clean texts does not significantly affect the detection performance of our proposed method. This further demonstrates the robustness of our approach. We attribute this to two main factors. First, the textual perturbations introduced by poisoning attacks are fundamentally different in nature from natural noise, and this distinction is clearly reflected in the activation vectors of the LLM. Second, LLMs are inherently robust to natural noise, which does not substantially interfere with their ability to generate correct responses.
+
+# 5.6 Efficiency
+
+Table 8 compares the time overhead between LLM Factoscope (He et al., 2024) and RevPRAG when the LLM in RAG is Llama3-8B, including the average training time per epoch and the average inference time per test sample. This experiment was conducted using 1,000 training samples and 500
+
+Table 7: Performance of RevPRAG on noisy datasets.  
+
+<table><tr><td>Dataset</td><td>Metrics</td><td>GPT2-XL 1.5B</td><td>Llama3-8B</td></tr><tr><td rowspan="2">Noisy_NQ</td><td>TPR</td><td>0.980</td><td>0.977</td></tr><tr><td>FPR</td><td>0.034</td><td>0.012</td></tr><tr><td rowspan="2">Noisy_HotpotQA</td><td>TPR</td><td>0.969</td><td>0.974</td></tr><tr><td>FPR</td><td>0.026</td><td>0.011</td></tr><tr><td rowspan="2">Noisy_MS-MARCO</td><td>TPR</td><td>0.989</td><td>0.977</td></tr><tr><td>FPR</td><td>0.013</td><td>0.018</td></tr></table>
+
+test samples, with poisoned and clean examples each accounting for  $50\%$ . The results demonstrate that RevPRAG, with its task-specific architecture and carefully selected detection metrics, incurs significantly lower computational costs than LLM Factoscope, which integrates multiple sub-models for hallucination detection. Its efficient detection capability makes RevPRAG particularly well-suited for latency-sensitive RAG scenarios, underscoring its practical value.
+
+Table 8: Comparison of time overhead.  
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="2">Training Time per Epoch</td><td colspan="2">Inference Time per Sample</td></tr><tr><td>LLM factoscope</td><td>RevPRAG</td><td>LLM factoscope</td><td>RevPRAG</td></tr><tr><td>NQ</td><td>91.61s</td><td>19.31s</td><td>0.0051s</td><td>0.0021s</td></tr><tr><td>HotpotQA</td><td>101.25s</td><td>23.69s</td><td>0.0066s</td><td>0.0023s</td></tr><tr><td>MS-MARCO</td><td>94.47s</td><td>20.72s</td><td>0.0058s</td><td>0.0022s</td></tr></table>
+
+# 6 Conclusion
+
+In this work, we find that correct and poisoned responses in RAG exhibit distinct differences in LLM activations. Building on this insight, we develop RevPRAG, a detection pipeline that leverages these activations to identify poisoned responses in RAG caused by the injection of malicious texts into the knowledge database. Our approach demonstrates robust performance across RAGs utilizing five different LLMs and four distinct retrievers on three datasets. Experimental results show that RevPRAG achieves exceptional accuracy, with true positive rates approaching  $98\%$  and false positive rates near  $1\%$ . Ablation studies further validate its effectiveness in detecting poisoned responses across different types and levels of poisoning attacks. Overall, our approach can accurately distinguish between correct and poisoned responses.
+
+# Acknowledgment.
+
+Ping Chen, Xiaoyan Sun, and Jun Dai are the corresponding authors. This research was supported by the National Key R&D Program of China 2023YFB3107404.
+
+# Limitations.
+
+Our work has the following limitations:
+
+- This work does not propose a specific method for defending against poisoning attacks on RAG. Instead, our focus is on the timely detection of poisoned responses generated by the LLM, aiming to prevent potential harm to users from such attacks.  
+- Our approach requires accessing the activations of the LLM, which necessitates the LLM being a white-box model. While this may present certain limitations for users, our method can be widely adopted by LLM service providers. Providers can implement our strategy to ensure the reliability of their services and enhance trust with their users.  
+- Our approach primarily focuses on determining whether the response generated by the RAG is correct or poisoned, without delving into more granular distinctions. The main goal of our study is to protect users from the impact of RAG poisoning attacks, while more detailed classifications of RAG responses will be addressed in future work.
+
+# Ethics Statement
+
+The goal of this work is to detect whether a RAG has generated a poisoned response. All the data used in this study is publicly available, so it does not introduce additional privacy concerns. All source code and software will be made open-source. While the open-source nature of the code may lead to adaptive attacks, we can further enhance our model by incorporating more internal and external information. Overall, we believe our approach can further promote the secure application of RAG.
+
+# References
+
+Sahar Abdelnabi, Aideen Fay, Giovanni Cherubin, Ahmed Salem, Mario Fritz, and Andrew Paverd. 2024. Are you still on track?! catching llm task drift with activations. arXiv preprint arXiv:2406.00799.  
+Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511.
+
+Anish Athalye, Nicholas Carlini, and David Wagner. 2018. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR.  
+Jinheon Baek, Soyeong Jeong, Minki Kang, Jong C Park, and Sung Hwang. 2023. Knowledge-augmented language model verification. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1720-1736.  
+Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, and 1 others. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.  
+Oliver Bryniarski, Nabeel Hingun, Pedro Pachuca, Vincent Wang, and Nicholas Carlini. 2021. Evading adversarial example detection defenses with orthogonal projected gradient descent. arXiv preprint arXiv:2106.15023.  
+Nicholas Carlini. 2023. A llm assisted exploitation of ai-guardian. arXiv preprint arXiv:2307.15008.  
+Nicholas Carlini and David Wagner. 2017. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM workshop on artificial intelligence and security, pages 3-14.  
+Sukmin Cho, Soyeong Jeong, Jeongyeon Seo, Taeho Hwang, and Jong C Park. 2024. Typos that broke the rag's back: Genetic attack on rag pipeline by simulating documents in the wild via low-level perturbations. arXiv preprint arXiv:2404.13948.  
+Sukmin Cho, Jeongyeon Seo, Soyeong Jeong, and Jong C Park. 2023. Improving zero-shot reader by reducing distractions from irrelevant documents in open-domain question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 3145-3157.  
+Javier Ferrando, Gabriele Sarti, Arianna Bisazza, and Marta R Costa-jussa. 2024. A primer on the inner workings of transformer-based language models. arXiv preprint arXiv:2405.00208.  
+Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. 2023. Not what you've signed up for: Compromising real-world llm-integrated applications with indirect prompt injection. In Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security, pages 79-90.  
+Jinwen He, Yujia Gong, Zijin Lin, Yue Zhao, Kai Chen, and 1 others. 2024. Llm factoscope: Uncovering llms' factual discernment through measuring inner states. In Findings of the Association for Computational Linguistics ACL 2024, pages 10218-10230.
+
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778.  
+Giwon Hong, Jeonghwan Kim, Junmo Kang, Sung Hyon Myaeng, and Joyce Whang. 2024. Why so gullible? enhancing the robustness of retrieval-augmented models against counterfactual noise. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2474-2495.  
+Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118.  
+Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park. 2024. Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7029-7043.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Ruochen Jiao, Shaoyuan Xie, Justin Yue, Takami Sato, Lixu Wang, Yixuan Wang, Qi Alfred Chen, and Qi Zhu. 2024. Exploring backdoor attacks against large language model-based decision making. arXiv preprint arXiv:2405.20774.  
+Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen Tau Yih. 2020. Dense passage retrieval for open-domain question answering. In 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, pages 6769-6781.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, and 1 others. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453-466.  
+Angeliki Lazaridou, Elena Gribovskaya, Wojciech Stokowiec, and Nikolai Grigorev. 2022. Internet-augmented language models through few-shot prompting for open-domain question answering. arXiv preprint arXiv:2203.05115.  
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, and 1 others. 2020. Retrieval-augmented
+
+generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474.  
+Xi Li, Yusen Zhang, Renze Lou, Chen Wu, and Jiaqi Wang. 2024. Chain-of-scrutiny: Detecting backdoor attacks for large language models. arXiv preprint arXiv:2406.05948.  
+Lefteris Loukas, Ilias Stogiannidis, Odysseas Diamantopoulos, Prodromos Malakasiotis, and Stavros Vassos. 2023. Making llms worth every penny: Resource-limited text classification in banking. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 392-400.  
+Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations.  
+Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Wang. 2023. On the risk of misinformation pollution with large language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 1389-1403.  
+Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 5687-5711.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, and 1 others. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823.  
+Ian Soboroff, Shudong Huang, and Donna Harman. 2018. Trec 2018 news track overview. In TREC, volume 409, page 410.  
+Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. 2020. On adaptive attacks to adversarial example defenses. Advances in neural information processing systems, 33:1633-1645.
+
+Zhepei Wei, Wei-Lin Chen, and Yu Meng. 2024. Instructrag: Instructing retrieval-augmented generation with explicit denoising. arXiv preprint arXiv:2406.13629.  
+Zhaohan Xi, Tianyu Du, Changjiang Li, Ren Pang, Shouling Ji, Jinghui Chen, Fenglong Ma, and Ting Wang. 2024. Defending pre-trained language models as few-shot learners against backdoor attacks. Advances in Neural Information Processing Systems, 36.  
+Chong Xiang, Tong Wu, Zexuan Zhong, David Wagner, Danqi Chen, and Prateek Mittal. 2024. Certifiably robust rag against retrieval corruption. arXiv preprint arXiv:2405.15556.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. arXiv preprint arXiv:2007.00808.  
+Jiaqi Xue, Mengxin Zheng, Yebowen Hu, Fei Liu, Xun Chen, and Qian Lou. 2024. Badrag: Identifying vulnerabilities in retrieval augmented generation of large language models. arXiv preprint arXiv:2406.00083.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369-2380.  
+Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2023. Making retrieval-augmented language models robust to irrelevant context. arXiv preprint arXiv:2310.01558.  
+Zexuan Zhong, Ziqing Huang, Alexander Wettig, and Danqi Chen. 2023. Poisoning retrieval corpora by injecting adversarial passages. In 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, pages 13764-13775.  
+Wei Zou, Runpeng Geng, Binghui Wang, and Jinyuan Jia. 2024. Poisonedrag: Knowledge corruption attacks to retrieval-augmented generation of large language models. arXiv preprint arXiv:2402.07867.
+
+# A Training Details
+
+# A.1 Dataset
+
+As shown in Table 9, we present the average response lengths for both poisoned and correct answers generated by GPT2-XL across three datasets (NQ, HotpotQA, and MS-MARCO), along with examples illustrating each answer format for a specific question. To evaluate the detection of poisoning attacks on the knowledge base of RAG, we selected 3,000 instances of triples  $(q,t,a)$  from
+
+each of the three evaluation datasets mentioned above. In each triple,  $q$  denotes a question,  $t$  represents the supporting text collected from Wikipedia or web documents corresponding to  $q$ , and  $a$  is the correct answer to  $q$ , generated using the state-of-the-art GPT-4 model. Among these 3,000 triplets, 1,500 are randomly selected as benign instances, while the remaining 1,500 are designated as poisoned instances. For each poisoned instance, the poisoned answer  $a_{p}$  is generated by GPT-4 for the given question  $q$ , and the poisoned text  $t_{p}$  is crafted using existing poisoning strategies, including PoisonedRAG (Zou et al., 2024), GARAG (Cho et al., 2024), and PRCAP (Zhong et al., 2023). The dataset is split into  $70\%$  for training,  $20\%$  for testing, and  $10\%$  as a support set. Within the training set, samples are randomly grouped into triplets (anchor, positive, negative), where the anchor and positive belong to the same class, and the negative belongs to a different class.
+
+# A.2 Poisoned Texts Generation
+
+To ensure that the retrieved poisoned texts successfully achieve the poisoning effect, we employ three existing methods PoisonedRAG (Zou et al., 2024), GARAG (Cho et al., 2024), and PRCAP (Zhong et al., 2023) to generate the poisoned texts. In the PoisonedRAG (Zou et al., 2024) method, the attacker first selects a target question along with its corresponding incorrect answer. The attacker then optimizes the design of the poisoned text to ensure that it meets two key criteria: (1) relevability by the retriever and (2) effectiveness in misleading the language model to generate the incorrect answer. GARAG (Cho et al., 2024) is a novel adversarial attack algorithm that generates adversarial documents by subtly perturbing clean ones while preserving answer tokens. Through iterative crossover, mutation, and selection, it optimizes the documents to maximize adversarial effectiveness within the defined search space. PRCAP (Zhong et al., 2023) is a gradient-based method, which starts from a natural-language passage and iteratively perturbs it in the discrete token space to maximize its similarity to a set of training queries.
+
+It's worth noting that the generation method for poisoned texts can be any approach that successfully achieves the poisoning effect. Once the activations of both correct and poisoned responses are obtained, we preprocess and use them for training and testing the RevPRAG model.
+
+Table 9: Statistical data and format of the responses.  
+
+<table><tr><td>Dataset</td><td>Average Word Count of Response</td><td>An Example of Response</td></tr><tr><td>NQ</td><td>Poisoned Response: 7 
+Correct Response: 12</td><td>Question: where is the food stored in a yam plant?
+Poisoned Response: In the leaves.
+Correct Response: In the tuber.</td></tr><tr><td>HotpotQA</td><td>Poisoned Response: 8 
+Correct Response: 11</td><td>Question: Which actor starred in Assignment to Kill and passed away in 2000?
+Poisoned Response: Patrick O’Neal.
+Correct Response: John Gielgud.</td></tr><tr><td>MS-MARCO</td><td>Poisoned Response: 16 
+Correct Response: 24</td><td>Question: what is hardie plank?
+Poisoned Response: Hardie plank is a wood flooring option that is used for a variety of home styles.
+Correct Response: Hardie Plank is a brand of fiber cement siding.</td></tr></table>
+
+# A.3 Prompt
+
+The following is the system prompt for RAG, instructing an LLM to produce a response based on the provided context:
+
+You are a helpful assistant. The user has provided a query along with relevant context information. Use this context to answer the question briefly and clearly. If you cannot find the answer to the question, respond with "I don't know."
+
+Contexts: [context]
+
+Query: [question]
+
+Answer:
+
+Figure 6: The prompt used in RAG to make an LLM generate an answer based on the retrieved texts.
+
+# A.4 Environment
+
+We conduct experiments on a server with 64 AMD EPYC 9654 CPUs (64 logical cores enabled) at  $2.40 - 3.70\mathrm{GHz}$ , 512 GB of DDR5 RAM (assumed based on high-core-count server standards), and four NVIDIA RTX A6000 GPUs, each with 48 GB GDDR6 memory.

revpragrevealingpoisoningattacksinretrievalaugmentedgenerationthroughllmactivationanalysis/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f00708762dc45776e35e648940e2cc0e04d41109a5feb2ea1e3b1092f6c37113
+size 524414

revpragrevealingpoisoningattacksinretrievalaugmentedgenerationthroughllmactivationanalysis/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c9ae8aa5ccb05633e21063deaaedd9e365068e57b59de283f6c8ca4a885b370
+size 420870

rewardmixologycraftinghybridsignalsforreinforcementlearningdrivenincontextlearning/a2b09701-51e2-4286-a9ac-dc8b2cf46dc1_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67dbbfde44797cb22fedfaf8746b6bae2599828011c99c19afba1ab716214984
+size 80029

rewardmixologycraftinghybridsignalsforreinforcementlearningdrivenincontextlearning/a2b09701-51e2-4286-a9ac-dc8b2cf46dc1_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b692bc432bcf77cdfe23fbd4ce2d0e5d51cdb7935cb75dca68ed5be5a67b5aa
+size 103396