Daily Papers

Multilingual large language models (mLLMs) contain a small fraction of feed-forward neurons that are sensitive to particular languages, commonly termed language-specific neurons. Existing work measures language specificity using the entropy of each neuron's language-wise probabilities of being active, where a neuron is considered active when its activation value is positive. However, this approach may not fully capture the multilingual nature of mLLMs, where language representations are distributional and mutually related. We propose Distribution-aware Language Neuron selection, which leverages pairwise relationships between per-language activation distributions over the full activation range, including negative values. Specifically, we quantify each neuron's language specificity by clustering languages using pairwise overlap coefficients between their activation distributions. Across two mLLMs and two held-out corpora, our identifier more effectively isolates language-specific causal effects, yielding up to 4.9times higher on-target language damage per neuron while preserving off-target language performance.

Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering and guided decoding. We instead investigate activation engineering: modifying activations at inference-time to predictably alter model behavior. We bias the forward pass with a 'steering vector' implicitly specified through natural language. Past work learned these steering vectors; our Activation Addition (ActAdd) method instead computes them by taking the activation differences which result from pairs of prompts. We demonstrate ActAdd on GPT-2 on OpenWebText and ConceptNet, and replicate the effect on Llama-13B and GPT-J-6B. Our approach yields inference-time control over high-level properties of output & preserves performance on off-target topics. The method requires far less compute and implementation effort than finetuning and RLHF, allows for natural language specification by users, and its overhead scales naturally with model size.

Large language models (LLMs) have made significant strides in various natural language processing (NLP) tasks. Recent research shows that the moderately-sized LLMs often outperform their larger counterparts after task-specific fine-tuning. In this work, we delve into the process of adapting LLMs to specialize in document-level machine translation (DocMT) for a specific language pair. Firstly, we explore how prompt strategies affect downstream translation performance. Then, we conduct extensive experiments with two fine-tuning methods, three LLM backbones, and 18 translation tasks across nine language pairs. Our findings indicate that in some cases, these specialized models even surpass GPT-4 in translation performance, while they still significantly suffer from the off-target translation issue in others, even if they are exclusively fine-tuned on bilingual parallel documents. Furthermore, we provide an in-depth analysis of these LLMs tailored for DocMT, exploring aspects such as translation errors, discourse phenomena, training strategy, the scaling law of parallel documents, additional evaluation on recent test sets, and zero-shot crosslingual transfer. Our findings not only shed light on the strengths and limitations of LLM-based DocMT models but also provide a foundation for future research.

The majority of previous researches addressing multi-lingual IE are limited to zero-shot cross-lingual single-transfer (one-to-one) setting, with high-resource languages predominantly as source training data. As a result, these works provide little understanding and benefit for the realistic goal of developing a multi-lingual IE system that can generalize to as many languages as possible. Our study aims to fill this gap by providing a detailed analysis on Cross-Lingual Multi-Transferability (many-to-many transfer learning), for the recent IE corpora that cover a diverse set of languages. Specifically, we first determine the correlation between single-transfer performance and a wide range of linguistic-based distances. From the obtained insights, a combined language distance metric can be developed that is not only highly correlated but also robust across different tasks and model scales. Next, we investigate the more general zero-shot multi-lingual transfer settings where multiple languages are involved in the training and evaluation processes. Language clustering based on the newly defined distance can provide directions for achieving the optimal cost-performance trade-off in data (languages) selection problem. Finally, a relational-transfer setting is proposed to further incorporate multi-lingual unlabeled data based on adversarial training using the relation induced from the above linguistic distance.

While multilingual neural machine translation has achieved great success, it suffers from the off-target issue, where the translation is in the wrong language. This problem is more pronounced on zero-shot translation tasks. In this work, we find that failing in encoding discriminative target language signal will lead to off-target and a closer lexical distance (i.e., KL-divergence) between two languages' vocabularies is related with a higher off-target rate. We also find that solely isolating the vocab of different languages in the decoder can alleviate the problem. Motivated by the findings, we propose Language Aware Vocabulary Sharing (LAVS), a simple and effective algorithm to construct the multilingual vocabulary, that greatly alleviates the off-target problem of the translation model by increasing the KL-divergence between languages. We conduct experiments on a multilingual machine translation benchmark in 11 languages. Experiments show that the off-target rate for 90 translation tasks is reduced from 29\% to 8\%, while the overall BLEU score is improved by an average of 1.9 points without extra training cost or sacrificing the supervised directions' performance. We release the code at https://github.com/PKUnlp-icler/Off-Target-MNMT for reproduction.

Multilingual Neural Machine Translation (MNMT) facilitates knowledge sharing but often suffers from poor zero-shot (ZS) translation qualities. While prior work has explored the causes of overall low ZS performance, our work introduces a fresh perspective: the presence of high variations in ZS performance. This suggests that MNMT does not uniformly exhibit poor ZS capability; instead, certain translation directions yield reasonable results. Through systematic experimentation involving 1,560 language directions spanning 40 languages, we identify three key factors contributing to high variations in ZS NMT performance: 1) target side translation capability 2) vocabulary overlap 3) linguistic properties. Our findings highlight that the target side translation quality is the most influential factor, with vocabulary overlap consistently impacting ZS performance. Additionally, linguistic properties, such as language family and writing system, play a role, particularly with smaller models. Furthermore, we suggest that the off-target issue is a symptom of inadequate ZS performance, emphasizing that zero-shot translation challenges extend beyond addressing the off-target problem. We release the data and models serving as a benchmark to study zero-shot for future research at https://github.com/Smu-Tan/ZS-NMT-Variations

It is common to reject undesired outputs of Large Language Models (LLMs); however, current methods to do so require an excessive amount of computation, or severely distort the distribution of outputs. We present a method to balance the distortion of the output distribution with computational efficiency, allowing for the generation of long sequences of text with difficult-to-satisfy constraints, with less amplification of low probability outputs compared to existing methods. We show through a series of experiments that the task-specific performance of our method is comparable to methods that do not distort the output distribution, while being much more computationally efficient.

Multilingual neural machine translation models generally distinguish translation directions by the language tag (LT) in front of the source or target sentences. However, current LT strategies cannot indicate the desired target language as expected on zero-shot translation, i.e., the off-target issue. Our analysis reveals that the indication of the target language is sensitive to the placement of the target LT. For example, when placing the target LT on the decoder side, the indication would rapidly degrade along with decoding steps, while placing the target LT on the encoder side would lead to copying or paraphrasing the source input. To address the above issues, we propose a simple yet effective strategy named Language Converter Strategy (LCS). By introducing the target language embedding into the top encoder layers, LCS mitigates confusion in the encoder and ensures stable language indication for the decoder. Experimental results on MultiUN, TED, and OPUS-100 datasets demonstrate that LCS could significantly mitigate the off-target issue, with language accuracy up to 95.28%, 96.21%, and 85.35% meanwhile outperforming the vanilla LT strategy by 3.07, 3,3, and 7.93 BLEU scores on zero-shot translation, respectively.

Given the prevalence of large language models (LLMs) and the prohibitive cost of training these models from scratch, dynamically forgetting specific knowledge e.g., private or proprietary, without retraining the model has become an important capability. This paper proposes a novel method to achieve this objective called UNLEARN. The approach builds upon subspace methods to identify and specifically target the removal of knowledge without adversely affecting other knowledge in the LLM. Results demonstrate 96% of targeted knowledge can be forgotten while maintaining performance on other knowledge within 2.5% of the original model, significantly outperforming the discriminatory abilities of the previous state-of-the-art. A dual method called LEARN is also proposed for targeted knowledge addition. Results show LEARN can match the fine-tuning accuracy of Low-Rank Adaptation (LoRA) without adversely affecting similar tasks.

Large language models (LLMs) may memorize sensitive or copyrighted content, raising privacy and legal concerns. Due to the high cost of retraining from scratch, researchers attempt to employ machine unlearning to remove specific content from LLMs while preserving the overall performance. In this paper, we discuss several issues in machine unlearning for LLMs and provide our insights on possible approaches. To address the issue of inadequate evaluation of model outputs after unlearning, we introduce three additional metrics to evaluate token diversity, sentence semantics, and factual correctness. We then categorize unlearning methods into untargeted and targeted, and discuss their issues respectively. Specifically, the behavior that untargeted unlearning attempts to approximate is unpredictable and may involve hallucinations, and existing regularization is insufficient for targeted unlearning. To alleviate these issues, we propose using the objective of maximizing entropy (ME) for untargeted unlearning and incorporate answer preservation (AP) loss as regularization for targeted unlearning. Experimental results across three scenarios, i.e., fictitious unlearning, continual unlearning, and real-world unlearning, demonstrate the effectiveness of our approaches. The code is available at https://github.com/sail-sg/closer-look-LLM-unlearning.

There have been a couple of studies showing that attempting to erase multilingual knowledge using only English data is insufficient for multilingual LLMs. However, their analyses remain highly performance-oriented. In this paper, we switch the point of view to evaluation, and address an additional blind spot which reveals itself when the multilingual LLM is fully finetuned with parallel multilingual dataset before unlearning. Here, language confusion occurs whereby a model responds in language different from that of the input prompt. Language confusion is a problematic phenomenon in unlearning, causing the standard reference-based metrics to fail. We tackle this phenomenon in three steps: (1) introduce N-gram-based Language-Mix (N-Mix) score to quantitatively show the language confusion is pervasive and consistent in multilingual LLMs, (2) demonstrate that reference-based metrics result in false negatives when N-Mix score is high, and(3) suggest the need of new type of unlearning evaluation that can directly assess the content of the generated sentences. We call this type of metrics as semantic-based metric.

Despite advances in Preference Alignment (PA) for Large Language Models (LLMs), mainstream methods like Reinforcement Learning with Human Feedback (RLHF) face notable challenges. These approaches require high-quality datasets of positive preference examples, which are costly to obtain and computationally intensive due to training instability, limiting their use in low-resource scenarios. LLM unlearning technique presents a promising alternative, by directly removing the influence of negative examples. However, current research has primarily focused on empirical validation, lacking systematic quantitative analysis. To bridge this gap, we propose a framework to explore the relationship between PA and LLM unlearning. Specifically, we introduce a bi-level optimization-based method to quantify the impact of unlearning specific negative examples on PA performance. Our analysis reveals that not all negative examples contribute equally to alignment improvement when unlearned, and the effect varies significantly across examples. Building on this insight, we pose a crucial question: how can we optimally select and weight negative examples for unlearning to maximize PA performance? To answer this, we propose a framework called Unlearning to Align (U2A), which leverages bi-level optimization to efficiently select and unlearn examples for optimal PA performance. We validate the proposed method through extensive experiments, with results confirming its effectiveness.

Recently, large language models (LLMs) have emerged as a notable field, attracting significant attention for its ability to automatically generate intelligent contents for various application domains. However, LLMs still suffer from significant security and privacy issues. For example, LLMs might expose user privacy from hacking attacks or targeted prompts. To address this problem, this paper introduces a novel machine unlearning framework into LLMs. Our objectives are to make LLMs not produce harmful, hallucinatory, or privacy-compromising responses, while retaining their standard output capabilities. To accomplish this, we use an evaluative model to pinpoint dialogues needing unlearning. We also establish a distance loss to function as the model's negative loss, diverting it from previous undesirable outputs. Furthermore, we determine the expected output's cluster mean to formulate a positive loss, directing the model's outputs toward preferable outcomes without compromising its reasoning abilities and performance. Experimental results show that our approach effectively meets unlearning objectives without substantially compromising model performance.

We propose a new algorithm for fine-tuning large language models using reinforcement learning. Tapered Off-Policy REINFORCE (TOPR) uses an asymmetric, tapered variant of importance sampling to speed up learning while maintaining stable learning dynamics, even without the use of KL regularization. TOPR can be applied in a fully offline fashion, allows the handling of positive and negative examples in a unified framework, and benefits from the implementational simplicity that is typical of Monte Carlo algorithms. We demonstrate the effectiveness of our approach with a series of experiments on the GSM8K and MATH reasoning benchmarks, finding performance gains for training both a model for solution generation and as a generative verifier. We show that properly leveraging positive and negative examples alike in the off-policy regime simultaneously increases test-time accuracy and training data efficiency, all the while avoiding the ``wasted inference'' that comes with discarding negative examples. We find that this advantage persists over multiple iterations of training and can be amplified by dataset curation techniques, enabling us to match 70B-parameter model performance with 8B language models. As a corollary to this work, we find that REINFORCE's baseline parameter plays an important and unexpected role in defining dataset composition in the presence of negative examples, and is consequently critical in driving off-policy performance.

We study how to perform unlearning, i.e. forgetting undesirable misbehaviors, on large language models (LLMs). We show at least three scenarios of aligning LLMs with human preferences can benefit from unlearning: (1) removing harmful responses, (2) erasing copyright-protected content as requested, and (3) reducing hallucinations. Unlearning, as an alignment technique, has three advantages. (1) It only requires negative (e.g. harmful) examples, which are much easier and cheaper to collect (e.g. via red teaming or user reporting) than positive (e.g. helpful and often human-written) examples required in RLHF (RL from human feedback). (2) It is computationally efficient. (3) It is especially effective when we know which training samples cause the misbehavior. To the best of our knowledge, our work is among the first to explore LLM unlearning. We are also among the first to formulate the settings, goals, and evaluations in LLM unlearning. We show that if practitioners only have limited resources, and therefore the priority is to stop generating undesirable outputs rather than to try to generate desirable outputs, unlearning is particularly appealing. Despite only having negative samples, our ablation study shows that unlearning can still achieve better alignment performance than RLHF with just 2% of its computational time.

The imperative to eliminate undesirable data memorization underscores the significance of machine unlearning for large language models (LLMs). Recent research has introduced a series of promising unlearning methods, notably boosting the practical significance of the field. Nevertheless, adopting a proper evaluation framework to reflect the true unlearning efficacy is also essential yet has not received adequate attention. This paper seeks to refine the evaluation of LLM unlearning by addressing two key challenges -- a) the robustness of evaluation metrics and b) the trade-offs between competing goals. The first challenge stems from findings that current metrics are susceptible to various red teaming scenarios. It indicates that they may not reflect the true extent of knowledge retained by LLMs but rather tend to mirror superficial model behaviors, thus prone to attacks. We address this issue by devising and assessing a series of candidate metrics, selecting the most robust ones under various types of attacks. The second challenge arises from the conflicting goals of eliminating unwanted knowledge while retaining those of others. This trade-off between unlearning and retention often fails to conform the Pareto frontier, rendering it subtle to compare the efficacy between methods that excel only in either unlearning or retention. We handle this issue by proposing a calibration method that can restore the original performance on non-targeted data after unlearning, thereby allowing us to focus exclusively on assessing the strength of unlearning. Our evaluation framework notably enhances the effectiveness when assessing and comparing various LLM unlearning methods, further allowing us to benchmark existing works, identify their proper hyper-parameters, and explore new tricks to enhance their practical efficacy.

Large Language Models (LLMs) have demonstrated remarkable performance across various Natural Language Processing (NLP) tasks, largely due to their generalisability and ability to perform tasks without additional training. However, their effectiveness for low-resource languages remains limited. In this study, we evaluate the performance of 55 publicly available LLMs on Maltese, a low-resource language, using a newly introduced benchmark covering 11 discriminative and generative tasks. Our experiments highlight that many models perform poorly, particularly on generative tasks, and that smaller fine-tuned models often perform better across all tasks. From our multidimensional analysis, we investigate various factors impacting performance. We conclude that prior exposure to Maltese during pre-training and instruction-tuning emerges as the most important factor. We also examine the trade-offs between fine-tuning and prompting, highlighting that while fine-tuning requires a higher initial cost, it yields better performance and lower inference costs. Through this work, we aim to highlight the need for more inclusive language technologies and recommend that researchers working with low-resource languages consider more "traditional" language modelling approaches.

Mitigating explicit and implicit biases in Large Language Models (LLMs) has become a critical focus in the field of natural language processing. However, many current methodologies evaluate scenarios in isolation, without considering the broader context or the spectrum of potential biases within each situation. To address this, we introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, which includes 450 offensive progressions containing 2,700 unique sentences of varying severity that progressively escalate from less to more explicitly offensive. Covering a broad spectrum of 9 demographics and 46 sub-demographics, STOP ensures inclusivity and comprehensive coverage. We evaluate several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3. Our findings reveal that even the best-performing models detect bias inconsistently, with success rates ranging from 19.3% to 69.8%. We also demonstrate how aligning models with human judgments on STOP can improve model answer rates on sensitive tasks such as BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or even improving performance. STOP presents a novel framework for assessing the complex nature of biases in LLMs, which will enable more effective bias mitigation strategies and facilitates the creation of fairer language models.

In order to enhance the performance of language models while mitigating the risks of generating harmful content, it is crucial to identify which training dataset affects the model's outputs. Ideally, we can measure the influence of each dataset by removing it from training; however, it is prohibitively expensive to retrain a model multiple times. This paper presents UnTrac, which estimates the influence of a training dataset by unlearning it from the trained model. UnTrac is extremely simple; each training dataset is unlearned by gradient ascent, and we evaluate how much the model's predictions change after unlearning. We empirically examine if our methods can assess the influence of pretraining datasets on generating toxic, biased, and untruthful content. Experimental results demonstrate that our method estimates their influence much more accurately than existing methods while requiring neither excessive memory space nor multiple model checkpoints.

Unlearning methods have the potential to improve the privacy and safety of large language models (LLMs) by removing sensitive or harmful information post hoc. The LLM unlearning research community has increasingly turned toward empirical benchmarks to assess the effectiveness of such methods. In this paper, we find that existing benchmarks provide an overly optimistic and potentially misleading view on the effectiveness of candidate unlearning methods. By introducing simple, benign modifications to a number of popular benchmarks, we expose instances where supposedly unlearned information remains accessible, or where the unlearning process has degraded the model's performance on retained information to a much greater extent than indicated by the original benchmark. We identify that existing benchmarks are particularly vulnerable to modifications that introduce even loose dependencies between the forget and retain information. Further, we show that ambiguity in unlearning targets in existing benchmarks can easily lead to the design of methods that overfit to the given test queries. Based on our findings, we urge the community to be cautious when interpreting benchmark results as reliable measures of progress, and we provide several recommendations to guide future LLM unlearning research.

As large language models (LLMs) are trained on massive datasets, they have raised significant privacy and ethical concerns due to their potential to inadvertently retain sensitive information. Unlearning seeks to selectively remove specific data from trained models, such as personal information or copyrighted content. Current approaches targeting specific output sequences at the token level often fail to achieve complete forgetting and remain susceptible to prompt rephrasing. We propose Align-then-Unlearn, a novel framework that performs unlearning in the semantic embedding space rather than directly on output tokens. Align-then-Unlearn first augments the LLM with an embedding prediction module trained to anticipate future context representations. Unlearning is then achieved by fine-tuning the model to minimize the similarity between these predicted embeddings and a target embedding that represents the concept to be removed. Initial results show that Align-then-Unlearn effectively removes targeted knowledge with minimal degradation in overall model utility. These findings suggest that embedding-based unlearning offers a promising and robust approach to removing conceptual knowledge. Our code is available at https://github.com/ExplainableML/align-then-unlearn.

With an evergrowing number of LLMs reporting superlative performance for English, their ability to perform equitably for different dialects of English (i.e., dialect robustness) needs to be ascertained. Specifically, we use English language (US English or Indian English) conversations between humans who play the word-guessing game of `taboo'. We formulate two evaluative tasks: target word prediction (TWP) (i.e.predict the masked target word in a conversation) and target word selection (TWS) (i.e., select the most likely masked target word in a conversation, from among a set of candidate words). Extending MD3, an existing dialectic dataset of taboo-playing conversations, we introduce M-MD3, a target-word-masked version of MD3 with the USEng and IndEng subsets. We add two subsets: AITrans (where dialectic information is removed from IndEng) and AIGen (where LLMs are prompted to generate conversations). Our evaluation uses pre-trained and fine-tuned versions of two closed-source (GPT-4/3.5) and two open-source LLMs (Mistral and Gemma). LLMs perform significantly better for US English than Indian English for both TWP and TWS, for all settings. While GPT-based models perform the best, the comparatively smaller models work more equitably for short conversations (<8 turns). Our results on AIGen and AITrans (the best and worst-performing subset) respectively show that LLMs may learn a dialect of their own based on the composition of the training data, and that dialect robustness is indeed a challenging task. Our evaluation methodology exhibits a novel way to examine attributes of language models using pre-existing dialogue datasets.

Machine unlearning aims to remove target influence while preserving other capabilities. This survey compares methods, benchmarks, and evidence across large language models and systems using retrieval, memory, tools, and interacting agents. A five-layer framework connects removal requests, system boundaries, target locations, interventions, and supported claims. A seven-stage lifecycle and six evidence dimensions guide comparison. The review shows that target construction, retained data, and recovery tests affect reported outcomes. Evidence from model evaluations remains insufficient to establish removal across external state and subsequent updates, motivating evaluation that tracks dependencies and tests whether target influence returns.

Comprehensive evaluation of Large Language Models (LLMs) is an open research problem. Existing evaluations rely on deterministic point estimates generated via greedy decoding. However, we find that deterministic evaluations fail to capture the whole output distribution of a model, yielding inaccurate estimations of model capabilities. This is particularly problematic in critical contexts such as unlearning and alignment, where precise model evaluations are crucial. To remedy this, we introduce the first formal probabilistic evaluation framework for LLMs. Namely, we propose novel metrics with high probability guarantees concerning the output distribution of a model. Our metrics are application-independent and allow practitioners to make more reliable estimates about model capabilities before deployment. Our experimental analysis reveals that deterministic evaluations falsely indicate successful unlearning and alignment, whereas our probabilistic evaluations better capture model capabilities. We show how to overcome challenges associated with probabilistic outputs in a case study on unlearning by introducing (1) a novel loss based on entropy optimization, and (2) adaptive temperature scaling. We demonstrate that our approach significantly enhances unlearning in probabilistic settings on recent benchmarks. Overall, our proposed shift from point estimates to probabilistic evaluations of output distributions represents an important step toward comprehensive evaluations of LLMs. Code available at https://www.cs.cit.tum.de/daml/probabilistic-unlearning/.

Machine unlearning aims to remove specific data influences from trained models, a capability essential for adhering to copyright laws and ensuring AI safety. Current unlearning metrics typically measure success by monitoring the model's performance degradation on the specific unlearning dataset (D_u). We argue that for Large Language Models (LLMs), this evaluation paradigm is insufficient and potentially misleading. Many real-world uses of unlearning--motivated by copyright or safety--implicitly target not only verbatim content in D_u, but also behaviors influenced by the broader generalizations the model derived from it. We demonstrate that LLMs can pass standard unlearning evaluation and appear to have "forgotten" the target knowledge, while simultaneously retaining strong capabilities on content that is semantically adjacent to D_u. This phenomenon indicates that erasing exact sentences does not necessarily equate to removing the underlying knowledge. To address this gap, we propose Proximal Surrogate Generation (PSG), an automated stress-testing framework that generates a surrogate dataset, D_u. This surrogate set is constructed to be semantically derived from D_u yet sufficiently distinct in embedding space. By comparing unlearning metric scores between D_u and D_u, we can stress-test the reliability of the metric itself. Our extensive evaluation across three LLM families (Llama-3-8B, Qwen2.5-7B, and Zephyr-7B-β), three distinct datasets, and seven standard metrics reveals widespread inconsistencies. We find that current metrics frequently overestimate unlearning success, failing to detect retained knowledge exposed by our stress-test datasets.

Modern ML systems ingest data aggregated from diverse sources, such as synthetic, human-annotated, and live customer traffic. Understanding which examples are important to the performance of a learning algorithm is crucial for efficient model training. Recently, a growing body of literature has given rise to various "influence scores," which use training artifacts such as model confidence or checkpointed gradients to identify important subsets of data. However, these methods have primarily been developed in computer vision settings, and it remains unclear how well they generalize to language-based tasks using pretrained models. In this paper, we explore the applicability of influence scores in language classification tasks. We evaluate a diverse subset of these scores on the SNLI dataset by quantifying accuracy changes in response to pruning training data through random and influence-score-based sampling. We then stress-test one of the scores -- "variance of gradients" (VoG) from Agarwal et al. (2022) -- in an NLU model stack that was exposed to dynamic user speech patterns in a voice assistant type of setting. Our experiments demonstrate that in many cases, encoder-based language models can be finetuned on roughly 50% of the original data without degradation in performance metrics. Along the way, we summarize lessons learned from applying out-of-the-box implementations of influence scores, quantify the effects of noisy and class-imbalanced data, and offer recommendations on score-based sampling for better accuracy and training efficiency.

The Unlearning Sensitive Content from Large Language Models task aims to remove targeted datapoints from trained models while minimally affecting their general knowledge. In our work, we leverage parameter-efficient, gradient-based unlearning using low-rank (LoRA) adaptation and layer-focused fine-tuning. To further enhance unlearning effectiveness, we employ data chunking, splitting forget data into disjoint partitions and merging them with cyclically sampled retain samples at a pre-defined ratio. Our task-agnostic method achieves an outstanding forget-retain balance, ranking first on leaderboards and significantly outperforming baselines and competing systems.

Advancements in large language models (LLMs) have demonstrated remarkable capabilities across a diverse range of applications. These models excel in generating text completions that are contextually coherent and cover an extensive array of subjects. However, the vast datasets required for their training make aligning response styles during the pretraining and instruction tuning phases challenging. Consequently, an additional alignment phase is typically employed, wherein the model is further trained with human preference data to better align its outputs with human expectations. While this process doesn't introduce new capabilities per se, it does accentuate generation styles innate to the model. This paper explores the utilization of counterfactual prompting within the framework of Direct Preference Optimization (DPO) to align the model's style without relying on human intervention. We demonstrate that this method effectively instils desirable behaviour, mitigates undesirable ones, and encourages the model to disregard inappropriate instructions. Our findings suggest that counterfactual prompting with DPO presents a low-resource way to fine-tune LLMs to meet the demands for responsible and ethically aligned AI systems.

Recent large language models (LLMs) demonstrate impressive capabilities in handling long contexts, some exhibiting near-perfect recall on synthetic retrieval tasks. However, these evaluations have mainly focused on English text and involved a single target sentence within lengthy contexts. Our work investigates how LLM performance generalizes to multilingual settings with multiple hidden target sentences. We create a new dataset -- mLongRR -- to comprehensively evaluate several multilingual long-context LLMs on retrieval and reasoning tasks across five languages: English, Vietnamese, Indonesian, Swahili, and Somali. These languages share the Latin script but belong to distinct language families and resource levels. Our analysis reveals a significant performance gap between languages. The best-performing models such as Gemini-1.5 and GPT-4o, achieve around 96% accuracy in English to around 36% in Somali with a single target sentence. However, this accuracy drops to 40% in English and 0% in Somali when dealing with three target sentences. Our findings highlight the challenges long-context LLMs face when processing longer contexts, an increase in the number of target sentences, or languages of lower resource levels.

Language models, while capable of generating remarkably coherent and seemingly accurate text, can occasionally produce undesirable content, including harmful or toxic outputs. In this paper, we present a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layerwise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for detected undesirable contents, we propose layerwise distributional steering policies that transform the attention heads. These policies are computed through principled semidefinite programming, which aims to minimally perturb the attention distribution while probabilistically guaranteeing the effectiveness of the editions. Empirical evaluations across multiple language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation. However, their tendency to memorize training data raises concerns regarding privacy, copyright compliance, and security, particularly in cases involving Personally Identifiable Information (PII). Effective machine unlearning techniques are essential to mitigate these risks, yet existing methods remain underdeveloped for LLMs due to their open-ended output space. In this work, we apply the Adaptive Representation Misdirection Unlearning (RMU) technique to unlearn sensitive information from LLMs. Through extensive experiments, we analyze the effects of unlearning across different decoder layers to determine the most effective regions for sensitive information removal. Our technique ranked 4th on the official leaderboard of both 1B parameter and 7B parameter models.

Machine unlearning can be useful for removing harmful capabilities and memorized text from large language models (LLMs), but there are not yet standardized methods for rigorously evaluating it. In this paper, we first survey techniques and limitations of existing unlearning evaluations. Second, we apply a comprehensive set of tests for the robustness and competitiveness of unlearning in the "Who's Harry Potter" (WHP) model from Eldan and Russinovich (2023). While WHP's unlearning generalizes well when evaluated with the "Familiarity" metric from Eldan and Russinovich, we find i) higher-than-baseline amounts of knowledge can reliably be extracted, ii) WHP performs on par with the original model on Harry Potter Q&A tasks, iii) it represents latent knowledge comparably to the original model, and iv) there is collateral unlearning in related domains. Overall, our results highlight the importance of comprehensive unlearning evaluation that avoids ad-hoc metrics.

We present an empirical evaluation of various outputs generated by nine of the most widely-available large language models (LLMs). Our analysis is done with off-the-shelf, readily-available tools. We find a correlation between percentage of memorized text, percentage of unique text, and overall output quality, when measured with respect to output pathologies such as counterfactual and logically-flawed statements, and general failures like not staying on topic. Overall, 80.0% of the outputs evaluated contained memorized data, but outputs containing the most memorized content were also more likely to be considered of high quality. We discuss and evaluate mitigation strategies, showing that, in the models evaluated, the rate of memorized text being output is reduced. We conclude with a discussion on potential implications around what it means to learn, to memorize, and to evaluate quality text.

Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment. Instead, this paper highlights the need for principled approaches to more advanced alignment. Alignment aims to ensure that large language models (LLMs) behave safely and reliably, including by avoiding unsafe inferences. However, we show that such safety-oriented behaviors can misfire: models may reject warranted conclusions even when they are explicitly supported by context. We call this failure mode misfired alignment, where alignment-induced changes cause LLMs to override explicit evidence. To quantify this phenomenon, specifically on stereotype-related alignment, we introduce VETO, a benchmark consisting of 2,032 BBQ-derived contrastive pairs, and define a new metric, Misfired Alignment Rate (MAR), which measures on a 0 to 100 scale how often a model fails on a stereotype-related question but succeeds on its contrastive counterpart. We benchmark 25 LLMs on VETO, and show that all LLMs, including the most recent ones, exhibit non-trivial (4.7 to 18.9%) MARs while all human participants achieve 0.0% MAR. Controlled priming experiments further show that alignment-induced cues can substantially amplify MAR across LLMs, indicating that these failures are not merely artifacts of individual examples but can be induced by safety-related framing. Mechanistic analyses on open-weight LLMs reveal late-layer suppression of evidence-supported answers, and comparisons between instruct and base LLMs suggest that this suppression emerges after instruction training. These findings show that current alignment methods can overgeneralize surface-level safety cues, to the point of overriding objective evidence, motivating more work on alignment objectives that better preserve contextual grounding.

Machine unlearning aims to remove sensitive or undesired data from large language models. However, recent studies suggest that unlearning is often shallow, claiming that removed knowledge can easily be recovered. In this work, we critically examine standard unlearning evaluation practices and uncover key limitations that shake our trust in those findings. First, we show that some evaluations introduce substantial new information into the model, potentially masking true unlearning performance by re-teaching the model during testing. Second, we demonstrate that evaluation outcomes vary significantly across tasks, undermining the generalizability of current evaluation routines. Finally, we find that many evaluations rely on spurious correlations, making their results difficult to trust and interpret. Taken together, these issues suggest that current evaluation protocols may both overstate and understate unlearning success. To address this, we propose two principles for future unlearning evaluations: minimal information injection and downstream task awareness. We validate these principles through a series of targeted experiments, showing how violations of each can lead to misleading conclusions.

Large Language Models (LLMs) demonstrate remarkable capabilities, but their training on massive corpora poses significant risks from memorized sensitive information. To mitigate these issues and align with legal standards, unlearning has emerged as a critical technique to selectively erase specific knowledge from LLMs without compromising their overall performance. This survey provides a systematic review of over 180 papers on LLM unlearning published since 2021. First, it introduces a novel taxonomy that categorizes unlearning methods based on the phase in the LLM pipeline of the intervention. This framework further distinguishes between parameter modification and parameter selection strategies, thus enabling deeper insights and more informed comparative analysis. Second, it offers a multidimensional analysis of evaluation paradigms. For datasets, we compare 18 existing benchmarks from the perspectives of task format, content, and experimental paradigms to offer actionable guidance. For metrics, we move beyond mere enumeration by dividing knowledge memorization metrics into 10 categories to analyze their advantages and applicability, while also reviewing metrics for model utility, robustness, and efficiency. By discussing current challenges and future directions, this survey aims to advance the field of LLM unlearning and the development of secure AI systems.

Advances in Multimodal Large Language Models (MLLMs) intensify concerns about data privacy, making Machine Unlearning (MU), the selective removal of learned information, a critical necessity. However, existing MU benchmarks for MLLMs are limited by a lack of image diversity, potential inaccuracies, and insufficient evaluation scenarios, which fail to capture the complexity of real-world applications. To facilitate the development of MLLMs unlearning and alleviate the aforementioned limitations, we introduce OFFSIDE, a novel benchmark for evaluating misinformation unlearning in MLLMs based on football transfer rumors. This manually curated dataset contains 15.68K records for 80 players, providing a comprehensive framework with four test sets to assess forgetting efficacy, generalization, utility, and robustness. OFFSIDE supports advanced settings like selective unlearning and corrective relearning, and crucially, unimodal unlearning (forgetting only text data). Our extensive evaluation of multiple baselines reveals key findings: (1) Unimodal methods (erasing text-based knowledge) fail on multimodal rumors; (2) Unlearning efficacy is largely driven by catastrophic forgetting; (3) All methods struggle with "visual rumors" (rumors appear in the image); (4) The unlearned rumors can be easily recovered and (5) All methods are vulnerable to prompt attacks. These results expose significant vulnerabilities in current approaches, highlighting the need for more robust multimodal unlearning solutions. The code is available at https://github.com/zh121800/OFFSIDE{https://github.com/zh121800/OFFSIDE}.

Generative language models (LMs) offer numerous advantages but may produce inappropriate or harmful outputs due to the harmful knowledge acquired during pre-training. This knowledge often manifests as undesirable correspondences, such as "harmful prompts" leading to "harmful outputs," which our research aims to mitigate through unlearning techniques.However, existing unlearning methods based on gradient ascent can significantly impair the performance of LMs. To address this issue, we propose a novel approach called Weighted Positional N-pair (WPN) Learning, which leverages position-weighted mean pooling within an n-pair contrastive learning framework. WPN is designed to modify the output distribution of LMs by eliminating specific harmful outputs (e.g., replacing toxic responses with neutral ones), thereby transforming the model's behavior from "harmful prompt-harmful output" to "harmful prompt-harmless response".Experiments on OPT and GPT-NEO LMs show that WPN effectively reduces the proportion of harmful responses, achieving a harmless rate of up to 95.8\% while maintaining stable performance on nine common benchmarks (with less than 2\% degradation on average). Moreover, we provide empirical evidence to demonstrate WPN's ability to weaken the harmful correspondences in terms of generalizability and robustness, as evaluated on out-of-distribution test sets and under adversarial attacks.

Large language models (LLMs) can memorize sensitive facts, motivating unlearning methods that remove targeted knowledge without costly retraining. However, unlearning research remains heavily English-centric. We study multilingual unlearning by extending the TOFU benchmark to five languages, and fine-tune, unlearn, and query our models with different permutations of languages. We find that unlearning transfer, the ability of an unlearned model to "forget" facts in languages other than the unlearning language, is highly variable: e.g., it is strongest between languages sharing scripts and families, and we show that the unlearning language predicts which query languages are most likely to yield the strongest transfer. Layer-wise analysis reveals that unlearning leaves the shared cross-lingual latent space largely intact in early layers, instead operating primarily in later decoding layers. This suggests that unlearning does not truly erase knowledge, but rather induces superficial suppression. Exploiting this structure, a single inference-time steering direction reverses much of this suppression across languages, recovering 50% (Qwen) and 90% (Gemma) of the unlearned knowledge.

Large language models (LLMs) are frequently fine-tuned or unlearned to adapt to new tasks or eliminate undesirable behaviors. While existing evaluation methods assess performance after such interventions, there remains no general approach for detecting unintended side effects, such as unlearning biology content degrading performance on chemistry tasks, particularly when these effects are unpredictable or emergent. To address this issue, we introduce MNEME, Model diffiNg for Evaluating Mechanistic Effects, a lightweight framework for identifying these side effects using sparse model diffing. MNEME compares base and fine-tuned models on task-agnostic data (for example, The Pile, LMSYS-Chat-1M) without access to fine-tuning data to isolate behavioral shifts. Applied to five LLMs across three scenarios: WMDP knowledge unlearning, emergent misalignment, and benign fine-tuning, MNEME achieves up to 95 percent accuracy in predicting side effects, aligning with known benchmarks and requiring no custom heuristics. Furthermore, we show that retraining on high-activation samples can partially reverse these effects. Our results demonstrate that sparse probing and diffing offer a scalable and automated lens into fine-tuning-induced model changes, providing practical tools for understanding and managing LLM behavior.

Since the release of ChatGPT, the field of Natural Language Processing has experienced rapid advancements, particularly in Large Language Models (LLMs) and their multimodal counterparts, Large Multimodal Models (LMMs). Despite their impressive capabilities, LLMs often exhibit significant performance disparities across different languages and cultural contexts, as demonstrated by various text-only benchmarks. However, current research lacks such benchmarks for multimodal visio-linguistic settings. This work fills this gap by introducing M5, the first comprehensive benchmark designed to evaluate LMMs on diverse vision-language tasks within a multilingual and multicultural context. M5 includes eight datasets covering five tasks and 41 languages, with a focus on underrepresented languages and culturally diverse images. Furthermore, we introduce two novel datasets, M5-VGR and M5-VLOD, including a new Visio-Linguistic Outlier Detection task, in which all evaluated open-source models fail to significantly surpass the random baseline. Through extensive evaluation and analyses, we highlight substantial task-agnostic performance disparities between high- and low-resource languages. Moreover, we show that larger models do not necessarily outperform smaller ones in a multilingual setting.

Language models exhibit systematic performance gaps when processing text in non-standard language varieties, yet their ability to comprehend variety-specific slang remains underexplored for several languages. We present a comprehensive evaluation of slang awareness in Indian English (en-IN) and Australian English (en-AU) across seven state-of-the-art language models. We construct two complementary datasets: WEB, containing 377 web-sourced usage examples from Urban Dictionary, and GEN, featuring 1,492 synthetically generated usages of these slang terms, across diverse scenarios. We assess language models on three tasks: target word prediction (TWP), guided target word prediction (TWP^*) and target word selection (TWS). Our results reveal four key findings: (1) Higher average model performance TWS versus TWP and TWP^*, with average accuracy score increasing from 0.03 to 0.49 respectively (2) Stronger average model performance on WEB versus GEN datasets, with average similarity score increasing by 0.03 and 0.05 across TWP and TWP^* tasks respectively (3) en-IN tasks outperform en-AU when averaged across all models and datasets, with TWS demonstrating the largest disparity, increasing average accuracy from 0.44 to 0.54. These findings underscore fundamental asymmetries between generative and discriminative competencies for variety-specific language, particularly in the context of slang expressions despite being in a technologically rich language such as English.

Large language Models (LLMs) have demonstrated remarkable skills across various domains. Understanding the mechanisms behind their abilities and implementing controls over them is becoming increasingly important for developing better models. In this paper, we focus on skill unlearning in LLMs, specifically unlearning a particular skill while retaining their overall capabilities. We introduce two lightweight, training-free machine skill unlearning techniques for LLMs. First, we observe that the pre-activation distribution of neurons in each Feed-Forward Layer (FFL) differs when the model demonstrates different skills. Additionally, we find that queries triggering the same skill cluster within the FFL key space and can be separated from other queries using a hypercube. Based on these observations, we propose two lightweight, training-free skill unlearning methods via intervention and abstention respectively: Neuron Adjust and Key Space Detection. We evaluate our methods on unlearning math-solving, Python-coding, and comprehension skills across seven different languages. The results demonstrate their strong unlearning capabilities for the designated skills. Specifically, Key Space Detection achieves over 80\% relative performance drop on the forgetting skill and less than 10\% relative performance drop on other skills and the model's general knowledge (MMLU) for most unlearning tasks. Our code is available at https://github.com/Trustworthy-ML-Lab/effective_skill_unlearning

Large language models (LLMs) offer impressive performance in various zero-shot and few-shot tasks. However, their success in zero-shot and few-shot settings may be affected by task contamination, a potential limitation that has not been thoroughly examined. This paper investigates how zero-shot and few-shot performance of LLMs has changed chronologically over time. Utilizing GPT-3 series models and several other recent open-sourced LLMs, and controlling for dataset difficulty, we find that on datasets released before the LLM training data creation date, LLMs perform surprisingly better than on datasets released after. This strongly indicates that, for many LLMs, there exists task contamination on zero-shot and few-shot evaluation for datasets released prior to the LLMs' training data creation date. Additionally, we utilize training data inspection, task example extraction, and a membership inference attack, which reveal further evidence of task contamination. Importantly, we find that for classification tasks with no possibility of task contamination, LLMs rarely demonstrate statistically significant improvements over simple majority baselines, in both zero and few-shot settings.

Reinforcement learning (RL) approaches for Large Language Models (LLMs) frequently use on-policy algorithms, such as PPO or GRPO. However, policy lag from distributed training architectures and differences between the training and inference policies break this assumption, making the data off-policy by design. To rectify this, prior work has focused on making this off-policy data appear more on-policy, either via importance sampling (IS), or by more closely aligning the training and inference policies by explicitly modifying the inference engine. In this work, we embrace off-policyness and propose a novel off-policy RL algorithm that does not require these modifications: Optimal Advantage-based Policy Optimization with Lagged Inference policy (OAPL). We show that OAPL outperforms GRPO with importance sampling on competition math benchmarks, and can match the performance of a publicly available coding model, DeepCoder, on LiveCodeBench, while using 3x fewer generations during training. We further empirically demonstrate that models trained via OAPL have improved test time scaling under the Pass@k metric. OAPL allows for efficient, effective post-training even with lags of more than 400 gradient steps between the training and inference policies, 100x more off-policy than prior approaches.

Large Language Models (LLMs) face significant challenges in maintaining privacy, ethics, and compliance, when sensitive or obsolete data must be selectively removed. Retraining these models from scratch is computationally infeasible, necessitating efficient alternatives. As part of the SemEval 2025 Task 4, this work focuses on the application of selective unlearning in LLMs to address this challenge. In this paper, we present our experiments and findings, primarily leveraging global weight modification to achieve an equilibrium between effectiveness of unlearning, knowledge retention, and target model's post-unlearning utility. We also detail the task-specific evaluation mechanism, results, and challenges. Our algorithms have achieved an aggregate score of 0.409 and 0.389 on the test set for 7B and 1B target models, respectively, demonstrating promising results in verifiable LLM unlearning.

As large language models (LLMs) are increasingly deployed in the real world, the ability to ``unlearn'', or remove specific pieces of knowledge post hoc, has become essential for a variety of reasons ranging from privacy regulations to correcting outdated or harmful content. Prior work has proposed unlearning benchmarks and algorithms, and has typically assumed that the training process and the target model are fixed. In this work, we empirically investigate how learning-time choices in knowledge encoding impact the effectiveness of unlearning factual knowledge. Our experiments reveal two key findings: (1) learning with paraphrased descriptions improves unlearning performance and (2) unlearning individual piece of knowledge from a chunk of text is challenging. Our results suggest that learning-time knowledge encoding may play a central role in enabling reliable post-hoc unlearning.

Large language models (LLMs) are widely used as zero-shot and few-shot classifiers, where task behaviour is largely controlled through prompting. A growing number of works have observed that LLMs are sensitive to prompt variations, with small changes leading to large changes in performance. However, in many cases, the investigation of sensitivity is performed using underspecified prompts that provide minimal task instructions and weakly constrain the model's output space. In this work, we argue that a significant portion of the observed prompt sensitivity can be attributed to prompt underspecification. We systematically study and compare the sensitivity of underspecified prompts and prompts that provide specific instructions. Utilising performance analysis, logit analysis, and linear probing, we find that underspecified prompts exhibit higher performance variance and lower logit values for relevant tokens, while instruction-prompts suffer less from such problems. However, linear probing analysis suggests that the effects of prompt underspecification have only a marginal impact on the internal LLM representations, instead emerging in the final layers. Overall, our findings highlight the need for more rigour when investigating and mitigating prompt sensitivity.

Representation Misdirection for Unlearning (RMU), which steers model representation in the intermediate layer to a target random representation, is an effective method for large language model (LLM) unlearning. Despite its high performance, the underlying cause and explanation remain underexplored. In this paper, we theoretically demonstrate that steering forget representations in the intermediate layer reduces token confidence, causing LLMs to generate wrong or nonsense responses. We investigate how the coefficient influences the alignment of forget-sample representations with the random direction and hint at the optimal coefficient values for effective unlearning across different network layers. We show that RMU unlearned models are robust against adversarial jailbreak attacks. Furthermore, our empirical analysis shows that RMU is less effective when applied to the middle and later layers in LLMs. To resolve this drawback, we propose Adaptive RMU--a simple yet effective alternative method that makes unlearning effective with most layers. Extensive experiments demonstrate that Adaptive RMU significantly improves the unlearning performance compared to prior art while incurring no additional computational cost.

The "massively-multilingual" training of multilingual models is known to limit their utility in any one language, and they perform particularly poorly on low-resource languages. However, there is evidence that low-resource languages can benefit from targeted multilinguality, where the model is trained on closely related languages. To test this approach more rigorously, we systematically study best practices for adapting a pre-trained model to a language family. Focusing on the Uralic family as a test case, we adapt XLM-R under various configurations to model 15 languages; we then evaluate the performance of each experimental setting on two downstream tasks and 11 evaluation languages. Our adapted models significantly outperform mono- and multilingual baselines. Furthermore, a regression analysis of hyperparameter effects reveals that adapted vocabulary size is relatively unimportant for low-resource languages, and that low-resource languages can be aggressively up-sampled during training at little detriment to performance in high-resource languages. These results introduce new best practices for performing language adaptation in a targeted setting.

The increase in computing power and the necessity of AI-assisted decision-making boost the growing application of large language models (LLMs). Along with this, the potential retention of sensitive data of LLMs has spurred increasing research into machine unlearning. However, existing unlearning approaches face a critical dilemma: Aggressive unlearning compromises model utility, while conservative strategies preserve utility but risk hallucinated responses. This significantly limits LLMs' reliability in knowledge-intensive applications. To address this, we introduce a novel Attention-Shifting (AS) framework for selective unlearning. AS is driven by two design objectives: (1) context-preserving suppression that attenuates attention to fact-bearing tokens without disrupting LLMs' linguistic structure; and (2) hallucination-resistant response shaping that discourages fabricated completions when queried about unlearning content. AS realizes these objectives through two attention-level interventions, which are importance-aware suppression applied to the unlearning set to reduce reliance on memorized knowledge and attention-guided retention enhancement that reinforces attention toward semantically essential tokens in the retained dataset to mitigate unintended degradation. These two components are jointly optimized via a dual-loss objective, which forms a soft boundary that localizes unlearning while preserving unrelated knowledge under representation superposition. Experimental results show that AS improves performance preservation over the state-of-the-art unlearning methods, achieving up to 15% higher accuracy on the ToFU benchmark and 10% on the TDEC benchmark, while maintaining competitive hallucination-free unlearning effectiveness. Compared to existing methods, AS demonstrates a superior balance between unlearning effectiveness, generalization, and response reliability.

Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model's performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model's geometry, as captured by existing metrics, primarily reflects training choices rather than performance.

Large language model unlearning aims to suppress designated undesirable knowledge while preserving benign capabilities. Many unlearning objectives focus on suppressing undesired answers, while recent target-guided variants specify replacement behavior but still leave update locality largely unconstrained. This paper introduces Null-Space Constrained Response-Specified Unlearning (NSRU), a projection-constrained low-rank framework for controlled LLM unlearning. NSRU uses an explicitly structured safe target response to specify the desired behavior for each forget query, while suppressing the original undesired content. To localize adaptation, NSRU estimates per-module retain subspaces from benign hidden representations and uses an orthogonal-projected low-rank parameterization to confine LoRA updates to the null space of the retain subspace. The resulting objective jointly optimizes safe-target learning, undesired-response suppression, and retention preservation under this constrained parameterization. We provide a local first-order analysis showing that the projected update reduces retain-side perturbations while preserving editable directions for shaping forget-query behavior. Experiments on TOFU show that NSRU effectively suppresses extractable forget-set knowledge while improving retain QA performance, model utility, and safe-target alignment over representative baselines. On WMDP, NSRU keeps hazardous-domain accuracy near the random-choice region while preserving broad and domain-adjacent MMLU utility. Ablation studies support the complementary roles of safe-target supervision, undesired-response suppression, retention loss, and null-space projected updates, while sensitivity and robustness analyses indicate stable behavior across the tested hyperparameter and prompt variations.

Large language models (LLMs) serve as giant information stores, often including personal or copyrighted data, and retraining them from scratch is not a viable option. This has led to the development of various fast, approximate unlearning techniques to selectively remove knowledge from LLMs. Prior research has largely focused on minimizing the probabilities of specific token sequences by reversing the language modeling objective. However, these methods still leave LLMs vulnerable to adversarial attacks that exploit indirect references. In this work, we examine the limitations of current unlearning techniques in effectively erasing a particular type of indirect prompt: multi-hop queries. Our findings reveal that existing methods fail to completely remove multi-hop knowledge when one of the intermediate hops is unlearned. To address this issue, we propose MUNCH, a simple uncertainty-based approach that breaks down multi-hop queries into subquestions and leverages the uncertainty of the unlearned model in final decision-making. Empirical results demonstrate the effectiveness of our framework, and MUNCH can be easily integrated with existing unlearning techniques, making it a flexible and useful solution for enhancing unlearning processes.

We audit five frontier large language models on native Bangla derogatory speech (gali) across six protocols to test a single hypothesis: Comprehension-Containment Decoupling. We propose that contemporary safety alignment is bound to high-resource surface forms rather than harmful meaning, causing a model's capacity to comprehend a low-resource slur and its capacity to contain it to operate independently. Every protocol corroborates this hypothesis against a human-calibrated baseline (kappa = 0.84). At baseline, models exhibit a 7.92 percentage point comprehension deficit in Bangla while maintaining an identical 92.83% token leakage rate across both languages. Severity calibration tracks surface anatomical cues over compositional harm (+4.00 error on mild slang; -2.00 on threats), while apparent containment gains under orthographic perturbation prove to be a tokenizer-driven "containment mirage." Crucially, explicit Chain-of-Thought reasoning rescues comprehension (94.72% Pass) while systematically dismantling containment (96.23% Use). Furthermore, expert-persona framing collapses refusal to 6.57%, revealing that keyword-based filters ignore dehumanizing communal slurs entirely. Our findings demonstrate that high-resource benchmarks cannot certify low-resource safety, necessitating meaning-grounded containment.

Fine-tuning all the layers of a pre-trained neural language encoder (either using all the parameters or using parameter-efficient methods) is often the de-facto way of adapting it to a new task. We show evidence that for different downstream language tasks, fine-tuning only a subset of layers is sufficient to obtain performance that is close to and often better than fine-tuning all the layers in the language encoder. We propose an efficient metric based on the diagonal of the Fisher information matrix (FIM score), to select the candidate layers for selective fine-tuning. We show, empirically on GLUE and SuperGLUE tasks and across distinct language encoders, that this metric can effectively select layers leading to a strong downstream performance. Our work highlights that task-specific information corresponding to a given downstream task is often localized within a few layers, and tuning only those is sufficient for strong performance. Additionally, we demonstrate the robustness of the FIM score to rank layers in a manner that remains constant during the optimization process.

Massively multilingual models for neural machine translation (NMT) are theoretically attractive, but often underperform bilingual models and deliver poor zero-shot translations. In this paper, we explore ways to improve them. We argue that multilingual NMT requires stronger modeling capacity to support language pairs with varying typological characteristics, and overcome this bottleneck via language-specific components and deepening NMT architectures. We identify the off-target translation issue (i.e. translating into a wrong target language) as the major source of the inferior zero-shot performance, and propose random online backtranslation to enforce the translation of unseen training language pairs. Experiments on OPUS-100 (a novel multilingual dataset with 100 languages) show that our approach substantially narrows the performance gap with bilingual models in both one-to-many and many-to-many settings, and improves zero-shot performance by ~10 BLEU, approaching conventional pivot-based methods.

We propose a post-training method for lower-resource languages that preserves fluency of language models even when aligned by disfluent reward models. Preference-optimization is now a well-researched topic, but previous work has mostly addressed models for English and Chinese. Lower-resource languages lack both datasets written by native speakers and language models capable of generating fluent synthetic data. Thus, in this work, we focus on developing a fluent preference-aligned language model without any instruction-tuning data in the target language. Our approach uses an on-policy training method, which we compare with two common approaches: supervised finetuning on machine-translated data and multilingual finetuning. We conduct a case study on Norwegian Bokmål and evaluate fluency through native-speaker assessments. The results show that the on-policy aspect is crucial and outperforms the alternatives without relying on any hard-to-obtain data.

In this paper, we explore the impact of augmenting pre-trained Encoder-Decoder models, specifically T5, with linguistic knowledge for the prediction of a target task. In particular, we investigate whether fine-tuning a T5 model on an intermediate task that predicts structural linguistic properties of sentences modifies its performance in the target task of predicting sentence-level complexity. Our study encompasses diverse experiments conducted on Italian and English datasets, employing both monolingual and multilingual T5 models at various sizes. Results obtained for both languages and in cross-lingual configurations show that linguistically motivated intermediate fine-tuning has generally a positive impact on target task performance, especially when applied to smaller models and in scenarios with limited data availability.

This paper investigates the propagation of harmful information in multilingual large language models (LLMs) and evaluates the efficacy of various unlearning methods. We demonstrate that fake information, regardless of the language it is in, once introduced into these models through training data, can spread across different languages, compromising the integrity and reliability of the generated content. Our findings reveal that standard unlearning techniques, which typically focus on English data, are insufficient in mitigating the spread of harmful content in multilingual contexts and could inadvertently reinforce harmful content across languages. We show that only by addressing harmful responses in both English and the original language of the harmful data can we effectively eliminate generations for all languages. This underscores the critical need for comprehensive unlearning strategies that consider the multilingual nature of modern LLMs to enhance their safety and reliability across diverse linguistic landscapes.

Large language models are highly sensitive to prompts, but this sensitivity is usually studied through task-relevant instructions, demonstrations, or reasoning cues. In this paper, we study a different form of prompt sensitivity: whether prompts that are semantically unrelated to the task can nevertheless steer model behavior. We call them spurious prompts and show their surprising efficacy. We also propose a simple black-box search procedure for discovering them. Across reasoning and question-answering benchmarks, using models ranging from 0.8B to 27B parameters and spanning three model families, we show that spurious prompts can improve performance, often matching or outperforming standard prompting baselines and task-aware prompt optimization. We further show that they can steer models toward unintended behaviors, such as repeatedly selecting the first answer option, producing incorrect answers, returning an even, prime or small number without explicitly instructing the model to do so. These findings reveal a new kind of prompt sensitivity: LLMs can be systematically steered by prompts that are unrelated to the task they are asked to solve. Our code is available at https://github.com/Batorskq/spurious

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

Despite the strong capabilities of Large Language Models (LLMs) to acquire knowledge from their training corpora, the memorization of sensitive information in the corpora such as copyrighted, biased, and private content has led to ethical and legal concerns. In response to these challenges, unlearning has emerged as a potential remedy for LLMs affected by problematic training data. However, previous unlearning techniques are either not applicable to black-box LLMs due to required access to model internal weights, or violate data protection principles by retaining sensitive data for inference-time correction. We propose δ-Unlearning, an offset unlearning framework for black-box LLMs. Instead of tuning the black-box LLM itself, δ-Unlearning learns the logit offset needed for unlearning by contrasting the logits from a pair of smaller models. Experiments demonstrate that δ- Unlearning can effectively unlearn target data while maintaining similar or even stronger performance on general out-of-forget-scope tasks. δ-Unlearning also effectively incorporates different unlearning algorithms, making our approach a versatile solution to adapting various existing unlearning algorithms to black-box LLMs.

Decoding methods play an indispensable role in converting language models from next-token predictors into practical task solvers. Prior research on decoding methods, primarily focusing on task-specific models, may not extend to the current era of general-purpose large language models (LLMs). Moreover, the recent influx of decoding strategies has further complicated this landscape. This paper provides a comprehensive and multifaceted analysis of various decoding methods within the context of LLMs, evaluating their performance, robustness to hyperparameter changes, and decoding speeds across a wide range of tasks, models, and deployment environments. Our findings reveal that decoding method performance is notably task-dependent and influenced by factors such as alignment, model size, and quantization. Intriguingly, sensitivity analysis exposes that certain methods achieve superior performance at the cost of extensive hyperparameter tuning, highlighting the trade-off between attaining optimal results and the practicality of implementation in varying contexts.

The growing legal and ethical scrutiny of large language models (LLMs) necessitates effective machine unlearning, particularly for sensitive or unauthorized data. Existing empirical methods often yield incomplete forgetting or unintended degradation of unrelated knowledge due to poor localization. In this work, we propose GRIN: a modular and targeted framework for LLM unlearning. GRIN introduces a novel gradient-ratio-based metric to identify parameters most responsible for memorizing forget data. We then perform selective noise injection into these parameters prior to fine-tuning, which improves unlearning performance while maintaining model utility. Finally, we propose new evaluation metrics tailored to the LLM setting and validate our approach on standard benchmarks such as TOFU, WMDP, and SafePKU.

Transferring learned patterns from pretrained neural language models has been shown to significantly improve effectiveness across a variety of language-based tasks, meanwhile further tuning on intermediate tasks has been demonstrated to provide additional performance benefits, provided the intermediate task is sufficiently related to the target task. However, how to identify related tasks is an open problem, and brute-force searching effective task combinations is prohibitively expensive. Hence, the question arises, are we able to improve the effectiveness and efficiency of tasks with no training examples through selective fine-tuning? In this paper, we explore statistical measures that approximate the divergence between domain representations as a means to estimate whether tuning using one task pair will exhibit performance benefits over tuning another. This estimation can then be used to reduce the number of task pairs that need to be tested by eliminating pairs that are unlikely to provide benefits. Through experimentation over 58 tasks and over 6,600 task pair combinations, we demonstrate that statistical measures can distinguish effective task pairs, and the resulting estimates can reduce end-to-end runtime by up to 40%.

Alignment faking in large language models presented a demonstration of Claude 3 Opus and Claude 3.5 Sonnet selectively complying with a helpful-only training objective to prevent modification of their behavior outside of training. We expand this analysis to 25 models and find that only 5 (Claude 3 Opus, Claude 3.5 Sonnet, Llama 3 405B, Grok 3, Gemini 2.0 Flash) comply with harmful queries more when they infer they are in training than when they infer they are in deployment. First, we study the motivations of these 5 models. Results from perturbing details of the scenario suggest that only Claude 3 Opus's compliance gap is primarily and consistently motivated by trying to keep its goals. Second, we investigate why many chat models don't fake alignment. Our results suggest this is not entirely due to a lack of capabilities: many base models fake alignment some of the time, and post-training eliminates alignment-faking for some models and amplifies it for others. We investigate 5 hypotheses for how post-training may suppress alignment faking and find that variations in refusal behavior may account for a significant portion of differences in alignment faking.

Large Language Models (LLMs) are commonly used as evaluators in various applications, but the reliability of the outcomes remains a challenge. One such challenge is using LLMs-as-judges for direct assessment, i.e., assigning scores from a specified range without any references. We first show that this challenge stems from LLM judge outputs being associated with score range bias, i.e., LLM judge outputs are highly sensitive to pre-defined score ranges, preventing the search for optimal score ranges. We also show that similar biases exist among models from the same family. We then mitigate this bias through contrastive decoding, achieving up to 11.3% relative improvement on average in Spearman correlation with human judgments across different score ranges.

Large Language Models (LLMs) achieve strong performance on diverse tasks but often exhibit cognitive inertia, struggling to follow instructions that conflict with the standardized patterns learned during supervised fine-tuning (SFT). To evaluate this limitation, we propose Inverse IFEval, a benchmark that measures models Counter-intuitive Abilitytheir capacity to override training-induced biases and comply with adversarial instructions. Inverse IFEval introduces eight types of such challenges, including Question Correction, Intentional Textual Flaws, Code without Comments, and Counterfactual Answering. Using a human-in-the-loop pipeline, we construct a dataset of 1012 high-quality Chinese and English questions across 23 domains, evaluated under an optimized LLM-as-a-Judge framework. Experiments on existing leading LLMs demonstrate the necessity of our proposed Inverse IFEval benchmark. Our findings emphasize that future alignment efforts should not only pursue fluency and factual correctness but also account for adaptability under unconventional contexts. We hope that Inverse IFEval serves as both a diagnostic tool and a foundation for developing methods that mitigate cognitive inertia, reduce overfitting to narrow patterns, and ultimately enhance the instruction-following reliability of LLMs in diverse and unpredictable real-world scenarios.

State-of-the-art natural language processing (NLP) models often learn to model dataset biases and surface form correlations instead of features that target the intended underlying task. Previous work has demonstrated effective methods to circumvent these issues when knowledge of the bias is available. We consider cases where the bias issues may not be explicitly identified, and show a method for training models that learn to ignore these problematic correlations. Our approach relies on the observation that models with limited capacity primarily learn to exploit biases in the dataset. We can leverage the errors of such limited capacity models to train a more robust model in a product of experts, thus bypassing the need to hand-craft a biased model. We show the effectiveness of this method to retain improvements in out-of-distribution settings even if no particular bias is targeted by the biased model.

Recent advances in zero-shot and few-shot learning have shown promise for a scope of research and practical purposes. However, this fast-growing area lacks standardized evaluation suites for non-English languages, hindering progress outside the Anglo-centric paradigm. To address this line of research, we propose TAPE (Text Attack and Perturbation Evaluation), a novel benchmark that includes six more complex NLU tasks for Russian, covering multi-hop reasoning, ethical concepts, logic and commonsense knowledge. The TAPE's design focuses on systematic zero-shot and few-shot NLU evaluation: (i) linguistic-oriented adversarial attacks and perturbations for analyzing robustness, and (ii) subpopulations for nuanced interpretation. The detailed analysis of testing the autoregressive baselines indicates that simple spelling-based perturbations affect the performance the most, while paraphrasing the input has a more negligible effect. At the same time, the results demonstrate a significant gap between the neural and human baselines for most tasks. We publicly release TAPE (tape-benchmark.com) to foster research on robust LMs that can generalize to new tasks when little to no supervision is available.

Psychological research consistently finds that human ratings of words across diverse semantic scales can be reduced to a low-dimensional form with relatively little information loss. We find that the semantic associations encoded in the embedding matrices of large language models (LLMs) exhibit a similar structure. We show that the projections of words on semantic directions defined by antonym pairs (e.g. kind - cruel) correlate highly with human ratings, and further find that these projections effectively reduce to a 3-dimensional subspace within LLM embeddings, closely resembling the patterns derived from human survey responses. Moreover, we find that shifting tokens along one semantic direction causes off-target effects on geometrically aligned features proportional to their cosine similarity. These findings suggest that semantic features are entangled within LLMs similarly to how they are interconnected in human language, and a great deal of semantic information, despite its apparent complexity, is surprisingly low-dimensional. Furthermore, accounting for this semantic structure may prove essential for avoiding unintended consequences when steering features.

Large language models (LLMs) should undergo rigorous audits to identify potential risks, such as copyright and privacy infringements. Once these risks emerge, timely updates are crucial to remove undesirable responses, ensuring legal and safe model usage. It has spurred recent research into LLM unlearning, focusing on erasing targeted undesirable knowledge without compromising the integrity of other, non-targeted responses. Existing studies have introduced various unlearning objectives to pursue LLM unlearning without necessitating complete retraining. However, each of these objectives has unique properties, and no unified framework is currently available to comprehend them thoroughly. To fill the gap, we propose a toolkit of the gradient effect (G-effect), quantifying the impacts of unlearning objectives on model performance from a gradient perspective. A notable advantage is its broad ability to detail the unlearning impacts from various aspects across instances, updating steps, and LLM layers. Accordingly, the G-effect offers new insights into identifying drawbacks of existing unlearning objectives, further motivating us to explore a series of new solutions for their mitigation and improvements. Finally, we outline promising directions that merit further studies, aiming at contributing to the community to advance this important field.

Although language models (LMs) demonstrate exceptional capabilities on various tasks, they are potentially vulnerable to extraction attacks, which represent a significant privacy risk. To mitigate the privacy concerns of LMs, machine unlearning has emerged as an important research area, which is utilized to induce the LM to selectively forget about some of its training data. While completely retraining the model will guarantee successful unlearning and privacy assurance, it is impractical for LMs, as it would be time-consuming and resource-intensive. Prior works efficiently unlearn the target token sequences, but upon subsequent iterations, the LM displays significant degradation in performance. In this work, we propose Privacy Protection via Optimal Parameters (POP), a novel unlearning method that effectively forgets the target token sequences from the pretrained LM by applying optimal gradient updates to the parameters. Inspired by the gradient derivation of complete retraining, we approximate the optimal training objective that successfully unlearns the target sequence while retaining the knowledge from the rest of the training data. Experimental results demonstrate that POP exhibits remarkable retention performance post-unlearning across 9 classification and 4 dialogue benchmarks, outperforming the state-of-the-art by a large margin. Furthermore, we introduce Remnant Memorization Accuracy that quantifies privacy risks based on token likelihood and validate its effectiveness through both qualitative and quantitative analyses.

Large language models (LLMs) often inherit biases from vast amounts of training corpora. Traditional debiasing methods, while effective to some extent, do not completely eliminate memorized biases and toxicity in LLMs. In this paper, we study an unlearning-based approach to debiasing in LLMs by performing gradient ascent on hate speech against minority groups, i.e., minimizing the likelihood of biased or toxic content. Specifically, we propose a mask language modeling unlearning technique, which unlearns the harmful part of the text. This method enables LLMs to selectively forget and disassociate from biased and harmful content. Experimental results demonstrate the effectiveness of our approach in diminishing bias while maintaining the language modeling abilities. Surprisingly, the results also unveil an unexpected potential for cross-domain transfer unlearning: debiasing in one bias form (e.g. gender) may contribute to mitigating others (e.g. race and religion).

Despite rapid progress in large language models (LLMs), their performance on a vast majority of languages remain unsatisfactory. In this paper, we study building language-specific LLMs by adapting monolingual and multilingual LLMs. We conduct systematic experiments on how design choices (base model selection, vocabulary extension, and continued fine-tuning) impact the adapted LLM, both in terms of efficiency (how many tokens are needed to encode the same amount of information) and end task performance. We find that (1) the initial performance before the adaptation is not always indicative of the final performance. (2) Efficiency can easily improved with simple vocabulary extension and continued fine-tuning in most LLMs we study, and (3) The optimal adaptation method is highly language-dependent, and the simplest approach works well across various experimental settings. Adapting English-centric models can yield better results than adapting multilingual models despite their worse initial performance on low-resource languages. Together, our work lays foundations on efficiently building language-specific LLMs by adapting existing LLMs.

While multilingual large language models generally perform adequately, and sometimes even rival English performance on high-resource languages (HRLs), they often significantly underperform on low-resource languages (LRLs). Among several prompting strategies aiming at bridging the gap, multilingual in-context learning (ICL) has been particularly effective when demonstration in target languages is unavailable. However, there lacks a systematic understanding of when and why it works well. In this work, we systematically analyze multilingual ICL, using demonstrations in HRLs to enhance cross-lingual transfer. We show that demonstrations in mixed HRLs consistently outperform English-only ones across the board, particularly for tasks written in LRLs. Surprisingly, our ablation study shows that the presence of irrelevant non-English sentences in the prompt yields measurable gains, suggesting the effectiveness of multilingual exposure itself. Our results highlight the potential of strategically leveraging multilingual resources to bridge the performance gap for underrepresented languages.

Language models can generate harmful and biased outputs and exhibit undesirable behavior according to a given cultural context. We propose a Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasets, an iterative process to significantly change model behavior by crafting and fine-tuning on a dataset that reflects a predetermined set of target values. We evaluate our process using three metrics: quantitative metrics with human evaluations that score output adherence to a target value, toxicity scoring on outputs; and qualitative metrics analyzing the most common word associated with a given social category. Through each iteration, we add additional training dataset examples based on observed shortcomings from evaluations. PALMS performs significantly better on all metrics compared to baseline and control models for a broad range of GPT-3 language model sizes without compromising capability integrity. We find that the effectiveness of PALMS increases with model size. We show that significantly adjusting language model behavior is feasible with a small, hand-curated dataset.

Aligning language models with human preferences relies on pairwise preference datasets. While some studies suggest that on-policy data consistently outperforms off -policy data for preference learning, others indicate that the advantages of on-policy data may be task-dependent, highlighting the need for a systematic exploration of their interplay. In this work, we show that on-policy and off-policy data offer complementary strengths in preference optimization: on-policy data is particularly effective for reasoning tasks like math and coding, while off-policy data performs better on open-ended tasks such as creative writing and making personal recommendations. Guided by these findings, we introduce SIMPLEMIX, an approach to combine the complementary strengths of on-policy and off-policy preference learning by simply mixing these two data sources. Our empirical results across diverse tasks and benchmarks demonstrate that SIMPLEMIX substantially improves language model alignment. Specifically, SIMPLEMIX improves upon on-policy DPO and off-policy DPO by an average of 6.03% on Alpaca Eval 2.0. Moreover, it outperforms prior approaches that are much more complex in combining on- and off-policy data, such as HyPO and DPO-Mix-P, by an average of 3.05%.

As multilingual large language models become more widely used, ensuring their safety and fairness across diverse linguistic contexts presents unique challenges. While existing research on machine unlearning has primarily focused on monolingual settings, typically English, multilingual environments introduce additional complexities due to cross-lingual knowledge transfer and biases embedded in both pretraining and fine-tuning data. In this work, we study multilingual unlearning using the Aya-Expanse 8B model under two settings: (1) data unlearning and (2) concept unlearning. We extend benchmarks for factual knowledge and stereotypes to ten languages through translation: English, French, Arabic, Japanese, Russian, Farsi, Korean, Hindi, Hebrew, and Indonesian. These languages span five language families and a wide range of resource levels. Our experiments show that unlearning in high-resource languages is generally more stable, with asymmetric transfer effects observed between typologically related languages. Furthermore, our analysis of linguistic distances indicates that syntactic similarity is the strongest predictor of cross-lingual unlearning behavior.

Large language models (LLMs) have achieved remarkable success across natural language processing tasks, yet their widespread deployment raises pressing concerns around privacy, copyright, security, and bias. Machine unlearning has emerged as a promising paradigm for selectively removing knowledge or data from trained models without full retraining. In this survey, we provide a structured overview of unlearning methods for LLMs, categorizing existing approaches into data-centric, parameter-centric, architecture-centric, hybrid, and other strategies. We also review the evaluation ecosystem, including benchmarks, metrics, and datasets designed to measure forgetting effectiveness, knowledge retention, and robustness. Finally, we outline key challenges and open problems, such as scalable efficiency, formal guarantees, cross-language and multimodal unlearning, and robustness against adversarial relearning. By synthesizing current progress and highlighting open directions, this paper aims to serve as a roadmap for developing reliable and responsible unlearning techniques in large language models.

Large language models (LLMs) can often be made to behave in undesirable ways that they are explicitly fine-tuned not to. For example, the LLM red-teaming literature has produced a wide variety of 'jailbreaking' techniques to elicit harmful text from models that were fine-tuned to be harmless. Recent work on red-teaming, model editing, and interpretability suggests that this challenge stems from how (adversarial) fine-tuning largely serves to suppress rather than remove undesirable capabilities from LLMs. Prior work has introduced latent adversarial training (LAT) as a way to improve robustness to broad classes of failures. These prior works have considered untargeted latent space attacks where the adversary perturbs latent activations to maximize loss on examples of desirable behavior. Untargeted LAT can provide a generic type of robustness but does not leverage information about specific failure modes. Here, we experiment with targeted LAT where the adversary seeks to minimize loss on a specific competing task. We find that it can augment a wide variety of state-of-the-art methods. First, we use targeted LAT to improve robustness to jailbreaks, outperforming a strong R2D2 baseline with orders of magnitude less compute. Second, we use it to more effectively remove backdoors with no knowledge of the trigger. Finally, we use it to more effectively unlearn knowledge for specific undesirable tasks in a way that is also more robust to re-learning. Overall, our results suggest that targeted LAT can be an effective tool for defending against harmful behaviors from LLMs.

Large Language Models (LLMs) increasingly require selective removal of harmful or sensitive knowledge, called unlearning, yet existing methods and benchmarks fail to evaluate this capability completely. Current approaches rely on disjoint forget and retain sets composed of independent facts, and measure success using simple and direct factual recall. This framing fails to capture a key requirement of unlearning, namely the ability to eliminate harmful behaviors while preserving benign and beneficial knowledge. We argue that effective unlearning must operate at the level of concepts, ensuring complete removal of unsafe applications while maintaining their correct and useful usage, thereby achieving conceptually meaningful and complete unlearning. To better evaluate unlearning techniques from such a practical viewpoint, we introduce the notion of dual-use concepts: concepts that can be used in both harmful and benign contexts. Building on these concepts, we construct a benchmark called ConceptGuard where forget and retain sets are explicitly complementary in concept usage. Our benchmark uniquely enables unlearning to be explored and gauged at the level of concepts, instead of sparse facts, and evaluation is intent-sensitive with the goal of maximizing contextual separation to promote safer behavior. We demonstrate that current unlearning techniques perform poorly under this setting, showing weak contextual separation alongside poor performance in ROUGE and concept-level metrics. Our results reveal strong forgetting-utility trade-offs, limited gains in contextual sensitivity, and poor consistency in concept-level control across methods, and provide ideas for unlearning approaches that better align with real-world safety requirements. Our dataset is publicly available.

Common evaluation paradigms for language models focus on scoring single responses through accuracy metrics or proper scoring rules, failing to capture the full richness of a model's belief state. Recent work illustrates that language models hallucinate in-part because they are optimised to be good test-takers under binary scoring schemes that reward any answer over abstention. While this insight naturally leads to penalty-based approaches, they ignore crucial distinctions in how models distribute uncertainty, for example between hedging toward incorrect answers versus hedging toward "I don't know" responses. A novel evaluation metric, the Distributional Correctness Score (DCS), is introduced to solve this problem, i.e., of not considering a model's entire probability distribution over answer choices. DCS naturally distinguishes between harmful overconfidence in wrong answers and uncertainty expressed through abstention, providing scores in an interpretable default range. Through theoretical analysis and illustrative examples, DCS is demonstrated to offer a more nuanced and aligned evaluation paradigm that incentivises models to express genuine uncertainty rather than guessing. Adapting 12 existing evaluation benchmarks to DCS's variants and measuring performance on six language models reveals that for half of the tested benchmarks scores are negative across all tested models, indicating significant tendencies towards hallucination.

Large Language Models (LLMs) are known to hallucinate, whereby they generate plausible but inaccurate text. This phenomenon poses significant risks in critical applications, such as medicine or law, necessitating robust hallucination mitigation strategies. While recent works have proposed fine-tuning methods to teach LLMs to abstain from answering questions beyond their knowledge or capabilities, these methods rely on the existence of ground-truth labels or are limited to short-form responses. To address these limitations, we propose fine-tuning using semantic entropy, an uncertainty measure derived from introspection into the model which does not require external labels. We demonstrate that our approach matches or outperforms models fine-tuned using prior work and achieves strong performance for both short and long-form generations on a range of datasets.

The rapid advancement of Large Language Models (LLMs) has established standardized evaluation benchmarks as the primary instrument for model comparison. Yet, their reliability is increasingly questioned due to sensitivity to shallow variations in input prompts. This paper examines how controlled, truth-conditionally equivalent lexical and syntactic perturbations affect the absolute performance and relative ranking of 23 contemporary LLMs across three benchmarks: MMLU, SQuAD, and AMEGA. We employ two linguistically principled pipelines to generate meaning-preserving variations: one performing synonym substitution for lexical changes, and another using dependency parsing to determine applicable syntactic transformations. Results show that lexical perturbations consistently induce substantial, statistically significant performance degradation across nearly all models and tasks, while syntactic perturbations have more heterogeneous effects, occasionally improving results. Both perturbation types destabilize model leaderboards on complex tasks. Furthermore, model robustness did not consistently scale with model size, revealing strong task dependence. Overall, the findings suggest that LLMs rely more on surface-level lexical patterns than on abstract linguistic competence, underscoring the need for robustness testing as a standard component of LLM evaluation.

The widespread adoption of large language models (LLMs) makes it important to recognize their strengths and limitations. We argue that in order to develop a holistic understanding of these systems we need to consider the problem that they were trained to solve: next-word prediction over Internet text. By recognizing the pressures that this task exerts we can make predictions about the strategies that LLMs will adopt, allowing us to reason about when they will succeed or fail. This approach - which we call the teleological approach - leads us to identify three factors that we hypothesize will influence LLM accuracy: the probability of the task to be performed, the probability of the target output, and the probability of the provided input. We predict that LLMs will achieve higher accuracy when these probabilities are high than when they are low - even in deterministic settings where probability should not matter. To test our predictions, we evaluate two LLMs (GPT-3.5 and GPT-4) on eleven tasks, and we find robust evidence that LLMs are influenced by probability in the ways that we have hypothesized. In many cases, the experiments reveal surprising failure modes. For instance, GPT-4's accuracy at decoding a simple cipher is 51% when the output is a high-probability word sequence but only 13% when it is low-probability. These results show that AI practitioners should be careful about using LLMs in low-probability situations. More broadly, we conclude that we should not evaluate LLMs as if they are humans but should instead treat them as a distinct type of system - one that has been shaped by its own particular set of pressures.

As large language models (LLMs) are increasingly deployed across various applications, privacy and copyright concerns have heightened the need for more effective LLM unlearning techniques. Many existing unlearning methods aim to suppress undesirable outputs through additional training (e.g., gradient ascent), which reduces the probability of generating such outputs. While such suppression-based approaches can control model outputs, they may not eliminate the underlying knowledge embedded in the model's internal activations; muting a response is not the same as forgetting it. Moreover, such suppression-based methods often suffer from model collapse. To address these issues, we propose a novel unlearning method that directly intervenes in the model's internal activations. In our formulation, forgetting is defined as a state in which the activation of a forgotten target is indistinguishable from that of ``unknown'' entities. Our method introduces an unlearning objective that modifies the activation of the target entity away from those of known entities and toward those of unknown entities in a sparse autoencoder latent space. By aligning the target's internal activation with those of unknown entities, we shift the model's recognition of the target entity from ``known'' to ``unknown'', achieving genuine forgetting while avoiding over-suppression and model collapse. Empirically, we show that our method effectively aligns the internal activations of the forgotten target, a result that the suppression-based approaches do not reliably achieve. Additionally, our method effectively reduces the model's recall of target knowledge in question-answering tasks without significant damage to the non-target knowledge.

The advancement of large language models (LLMs) has enhanced the ability to generalize across a wide range of unseen natural language processing (NLP) tasks through instruction-following. Yet, their effectiveness often diminishes in low-resource languages like Chinese, exacerbated by biased evaluations from data leakage, casting doubt on their true generalizability to new linguistic territories. In response, we introduce the Chinese Instruction-Following Benchmark (CIF-Bench), designed to evaluate the zero-shot generalizability of LLMs to the Chinese language. CIF-Bench comprises 150 tasks and 15,000 input-output pairs, developed by native speakers to test complex reasoning and Chinese cultural nuances across 20 categories. To mitigate evaluation bias, we release only half of the dataset publicly, with the remainder kept private, and introduce diversified instructions to minimize score variance, totaling 45,000 data instances. Our evaluation of 28 selected LLMs reveals a noticeable performance gap, with the best model scoring only 52.9%, highlighting the limitations of LLMs in less familiar language and task contexts. This work aims to uncover the current limitations of LLMs in handling Chinese tasks, pushing towards the development of more culturally informed and linguistically diverse models with the released data and benchmark (https://yizhilll.github.io/CIF-Bench/).

As machine unlearning for multimodal large language models (MLLMs) rapidly advances, rigorous evaluation of its effectiveness becomes increasingly important. However, existing benchmarks largely inherit evaluation protocols from text-only unlearning and overlook the unique challenges of unlearning in multimodal settings. In this work, we identify three major pitfalls that may lead to misleading conclusions: the lack of heldout evaluation to distinguish genuine concept-wise forgetting from instance-wise overfitting, the reliance on synthetic, fine-tuned knowledge rather than real-world knowledge acquired during pre-training, and unfair method comparisons caused by applying the same training settings to different unlearning methods. To address these challenges, we introduce GEMMU (Generalizable Evaluation for Multimodal Machine Unlearning), a benchmark built on real-world knowledge and a principled evaluation protocol. Re-evaluating representative unlearning methods under GEMMU reveals substantial generalization gaps and challenges conclusions drawn from previous benchmarks. Our findings highlight the importance of rigorous evaluation for advancing multimodal machine unlearning and provide guidelines for future benchmark design.

Pretrained language models memorize vast amounts of information, including private and copyrighted data, raising significant safety concerns. Retraining these models after excluding sensitive data is prohibitively expensive, making machine unlearning a viable, cost-effective alternative. Previous research has focused on machine unlearning for monolingual models, but we find that unlearning in one language does not necessarily transfer to others. This vulnerability makes models susceptible to low-resource language attacks, where sensitive information remains accessible in less dominant languages. This paper presents a pioneering approach to machine unlearning for multilingual language models, selectively erasing information across different languages while maintaining overall performance. Specifically, our method employs an adaptive unlearning scheme that assigns language-dependent weights to address different language performances of multilingual language models. Empirical results demonstrate the effectiveness of our framework compared to existing unlearning baselines, setting a new standard for secure and adaptable multilingual language models.

Instruction tuning is now the default way to train and adapt large language models, but many instruction--input--output pairs are only weakly specified: for a given input, the same output can remain plausible under several alternative instructions. This raises a simple question: does the instruction uniquely determine the target output? We propose the Task--Specificity Score (TSS) to quantify how much an instruction matters for predicting its output, by contrasting the true instruction against plausible alternatives for the same input. We further introduce TSS++, which uses hard alternatives and a small quality term to mitigate easy-negative effects. Across three instruction datasets (Alpaca, Dolly-15k, NI-20) and three open LLMs (Gemma, Llama, Qwen), we show that selecting task-specific examples improves downstream performance under tight token budgets and complements quality-based filters such as perplexity and IFD.

Large Language Models (LLMs), pre-trained on massive text corpora, exhibit remarkable human-level language understanding, reasoning, and decision-making abilities. However, they tend to memorize unwanted information, such as private or copyrighted content, raising significant privacy and legal concerns. Unlearning has emerged as a promising solution, but existing methods face a significant challenge of over-forgetting. This issue arises because they indiscriminately suppress the generation of all the tokens in forget samples, leading to a substantial loss of model utility. To overcome this challenge, we introduce the Targeted Information Forgetting (TIF) framework, which consists of (1) a flexible targeted information identifier designed to differentiate between unwanted words (UW) and general words (GW) in the forget samples, and (2) a novel Targeted Preference Optimization approach that leverages Logit Preference Loss to unlearn unwanted information associated with UW and Preservation Loss to retain general information in GW, effectively improving the unlearning process while mitigating utility degradation. Extensive experiments on the TOFU and MUSE benchmarks demonstrate that the proposed TIF framework enhances unlearning effectiveness while preserving model utility and achieving state-of-the-art results.

Large Language Models (LLMs) often memorize sensitive or harmful information, necessitating effective machine unlearning techniques. While existing parameter-efficient unlearning methods have shown promise, they still struggle with the forget-retain trade-off. This can be attributed to their reliance on parameter importance metrics to identify parameters that are important exclusively for the forget set, which is fundamentally limited by the superposition phenomenon. Due to the polysemantic nature of LLM parameters, such an importance metric may struggle to disentangle parameters associated with the forget and retain sets. In this work, we propose Representation-Guided Low-rank Unlearning (REGLU), a novel approach that leverages the geometric properties of representation spaces to achieve robust and precise unlearning. First, we develop a representation-guided initialization for LoRA that identifies the optimal subspace for selective forgetting. Second, we introduce a regularization loss that constrains the outputs of the LoRA update to lie in the orthogonal complement of the retain set's representation subspace, thereby minimizing interference with the model's performance on the retain set. We evaluate REGLU on the TOFU and WMDP benchmarks across multiple models. Our results demonstrate that REGLU consistently outperforms state-of-the-art baselines, achieving superior unlearning quality while maintaining higher model utility.

Large language models (LLMs) struggle to follow a sequence of instructions in a single query as they may ignore or misinterpret part of it. This impairs their performance in complex problems whose solution requires multiple intermediate steps, such as multilingual (translate then answer) and multimodal (caption then answer) tasks. We empirically verify this with open-source LLMs as large as LLaMA-2 70B and Mixtral-8x7B. Targeting the scarcity of sequential instructions in present-day data, we propose sequential instruction tuning, a simple yet effective strategy to automatically augment instruction tuning data and equip LLMs with the ability to execute multiple sequential instructions. After exploring interleaving instructions in existing datasets, such as Alpaca, with a wide range of intermediate tasks, we find that sequential instruction-tuned models consistently outperform the conventional instruction-tuned baselines in downstream tasks involving reasoning, multilingual, and multimodal abilities. To shed further light on our technique, we analyse how adversarial intermediate texts, unseen tasks, prompt verbalization, number of tasks, and prompt length affect SIT. We hope that this method will open new research avenues on instruction tuning for complex tasks.

Recent years have witnessed a surge in the number of large language models (LLMs), yet efficiently managing and utilizing these vast resources remains a significant challenge. In this work, we explore how to learn compact representations of LLM abilities that can facilitate downstream tasks, such as model routing and performance prediction on new benchmarks. We frame this problem as estimating the probability that a given model will correctly answer a specific query. Inspired by the item response theory (IRT) in psychometrics, we model this probability as a function of three key factors: (i) the model's multi-skill ability vector, (2) the query's discrimination vector that separates models of differing skills, and (3) the query's difficulty scalar. To learn these parameters jointly, we introduce a Mixture-of-Experts (MoE) network that couples model- and query-level embeddings. Extensive experiments demonstrate that our approach leads to state-of-the-art performance in both model routing and benchmark accuracy prediction. Moreover, analysis validates that the learned parameters encode meaningful, interpretable information about model capabilities and query characteristics.

Large language models (LMs) are able to in-context learn -- perform a new task via inference alone by conditioning on a few input-label pairs (demonstrations) and making predictions for new inputs. However, there has been little understanding of how the model learns and which aspects of the demonstrations contribute to end task performance. In this paper, we show that ground truth demonstrations are in fact not required -- randomly replacing labels in the demonstrations barely hurts performance on a range of classification and multi-choce tasks, consistently over 12 different models including GPT-3. Instead, we find that other aspects of the demonstrations are the key drivers of end task performance, including the fact that they provide a few examples of (1) the label space, (2) the distribution of the input text, and (3) the overall format of the sequence. Together, our analysis provides a new way of understanding how and why in-context learning works, while opening up new questions about how much can be learned from large language models through inference alone.

This paper explores Machine Unlearning (MU), an emerging field that is gaining increased attention due to concerns about neural models unintentionally remembering personal or sensitive information. We present SeUL, a novel method that enables selective and fine-grained unlearning for language models. Unlike previous work that employs a fully reversed training objective in unlearning, SeUL minimizes the negative impact on the capability of language models, particularly in terms of generation. Furthermore, we introduce two innovative evaluation metrics, sensitive extraction likelihood (S-EL) and sensitive memorization accuracy (S-MA), specifically designed to assess the effectiveness of forgetting sensitive information. In support of the unlearning framework, we propose efficient automatic online and offline sensitive span annotation methods. The online selection method, based on language probability scores, ensures computational efficiency, while the offline annotation involves a two-stage LLM-based process for robust verification. In summary, this paper contributes a novel selective unlearning method (SeUL), introduces specialized evaluation metrics (S-EL and S-MA) for assessing sensitive information forgetting, and proposes automatic online and offline sensitive span annotation methods to support the overall unlearning framework and evaluation process.

Work on scaling laws has found that large language models (LMs) show predictable improvements to overall loss with increased scale (model size, training data, and compute). Here, we present evidence for the claim that LMs may show inverse scaling, or worse task performance with increased scale, e.g., due to flaws in the training objective and data. We present empirical evidence of inverse scaling on 11 datasets collected by running a public contest, the Inverse Scaling Prize, with a substantial prize pool. Through analysis of the datasets, along with other examples found in the literature, we identify four potential causes of inverse scaling: (i) preference to repeat memorized sequences over following in-context instructions, (ii) imitation of undesirable patterns in the training data, (iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and (iv) correct but misleading few-shot demonstrations of the task. We release the winning datasets at https://inversescaling.com/data to allow for further investigation of inverse scaling. Our tasks have helped drive the discovery of U-shaped and inverted-U scaling trends, where an initial trend reverses, suggesting that scaling trends are less reliable at predicting the behavior of larger-scale models than previously understood. Overall, our results suggest that there are tasks for which increased model scale alone may not lead to progress, and that more careful thought needs to go into the data and objectives for training language models.

Offensive language detection is increasingly crucial for maintaining a civilized social media platform and deploying pre-trained language models. However, this task in Chinese is still under exploration due to the scarcity of reliable datasets. To this end, we propose a benchmark --COLD for Chinese offensive language analysis, including a Chinese Offensive Language Dataset --COLDATASET and a baseline detector --COLDETECTOR which is trained on the dataset. We show that the COLD benchmark contributes to Chinese offensive language detection which is challenging for existing resources. We then deploy the COLDETECTOR and conduct detailed analyses on popular Chinese pre-trained language models. We first analyze the offensiveness of existing generative models and show that these models inevitably expose varying degrees of offensive issues. Furthermore, we investigate the factors that influence the offensive generations, and we find that anti-bias contents and keywords referring to certain groups or revealing negative attitudes trigger offensive outputs easier.

Interventions targeting the representation space of language models (LMs) have emerged as an effective means to influence model behavior. Such methods are employed, for example, to eliminate or alter the encoding of demographic information such as gender within the model's representations and, in so doing, create a counterfactual representation. However, because the intervention operates within the representation space, understanding precisely what aspects of the text it modifies poses a challenge. In this paper, we give a method to convert representation counterfactuals into string counterfactuals. We demonstrate that this approach enables us to analyze the linguistic alterations corresponding to a given representation space intervention and to interpret the features utilized to encode a specific concept. Moreover, the resulting counterfactuals can be used to mitigate bias in classification through data augmentation.

The deployment of large language models (LLMs) like ChatGPT and Gemini has shown their powerful natural language generation capabilities. However, these models can inadvertently learn and retain sensitive information and harmful content during training, raising significant ethical and legal concerns. To address these issues, machine unlearning has been introduced as a potential solution. While existing unlearning methods take into account the specific characteristics of LLMs, they often suffer from high computational demands, limited applicability, or the risk of catastrophic forgetting. To address these limitations, we propose a lightweight behavioral unlearning framework based on Retrieval-Augmented Generation (RAG) technology. By modifying the external knowledge base of RAG, we simulate the effects of forgetting without directly interacting with the unlearned LLM. We approach the construction of unlearned knowledge as a constrained optimization problem, deriving two key components that underpin the effectiveness of RAG-based unlearning. This RAG-based approach is particularly effective for closed-source LLMs, where existing unlearning methods often fail. We evaluate our framework through extensive experiments on both open-source and closed-source models, including ChatGPT, Gemini, Llama-2-7b-chat, and PaLM 2. The results demonstrate that our approach meets five key unlearning criteria: effectiveness, universality, harmlessness, simplicity, and robustness. Meanwhile, this approach can extend to multimodal large language models and LLM-based agents.

Alignment has quickly become a default ingredient in LLM development, with techniques such as reinforcement learning from human feedback making models act safely, follow instructions, and perform ever-better on complex tasks. While these techniques are certainly useful, we propose that they should not be universally applied and demonstrate a range of tasks on which base language models consistently outperform their popular aligned forms. Particularly, we study tasks that require unpredictable outputs, such as random number generation, mixed strategy games (rock-paper-scissors and hide-and-seek), and creative writing. In each case, aligned models tend towards narrow behaviors that result in distinct disadvantages, for instance, preferring to generate "7" over other uniformly random numbers, becoming almost fully predictable in some game states, or prioritizing pleasant writing over creative originality. Across models tested, better performance on common benchmarks tends to correlate with worse performance on our tasks, suggesting an effective trade-off in the required capabilities.

An unintended consequence of the vast pretraining of Large Language Models (LLMs) is the verbatim memorization of fragments of their training data, which may contain sensitive or copyrighted information. In recent years, unlearning has emerged as a solution to effectively remove sensitive knowledge from models after training. Yet, recent work has shown that supposedly deleted information can still be extracted by malicious actors through various attacks. Still, current attacks retrieve sets of possible candidate generations and are unable to pinpoint the output that contains the actual target information. We propose activation steering as a method for exact information retrieval from unlearned LLMs. We introduce a novel approach to generating steering vectors, named Anonymized Activation Steering. Additionally, we develop a simple word frequency method to pinpoint the correct answer among a set of candidates when retrieving unlearned information. Our evaluation across multiple unlearning techniques and datasets demonstrates that activation steering successfully recovers general knowledge (e.g., widely known fictional characters) while revealing limitations in retrieving specific information (e.g., details about non-public individuals). Overall, our results demonstrate that exact information retrieval from unlearned models is possible, highlighting a severe vulnerability of current unlearning techniques.

When trained on large, unfiltered crawls from the internet, language models pick up and reproduce all kinds of undesirable biases that can be found in the data: they often generate racist, sexist, violent or otherwise toxic language. As large models require millions of training examples to achieve good performance, it is difficult to completely prevent them from being exposed to such content. In this paper, we first demonstrate a surprising finding: pretrained language models recognize, to a considerable degree, their undesirable biases and the toxicity of the content they produce. We refer to this capability as self-diagnosis. Based on this finding, we then propose a decoding algorithm that, given only a textual description of the undesired behavior, reduces the probability of a language model producing problematic text. We refer to this approach as self-debiasing. Self-debiasing does not rely on manually curated word lists, nor does it require any training data or changes to the model's parameters. While we by no means eliminate the issue of language models generating biased text, we believe our approach to be an important step in this direction.

Machine unlearning has emerged as a tool for removing personal data from trained models to comply with recent AI regulations. To evaluate unlearning effectiveness in multimodal large language models (MLLMs), prior works fine-tune models on fictitious identities, simulating unlearning requests on subsets of these IDs, which are typically uniformly distributed. However, in realistic scenarios, people from different demographic groups may request to be unlearned at different frequencies, potentially altering the model's internal beliefs for these groups and leading to biased behaviors. To fill this gap, we propose FAIRGET, the first Visual Question Answering benchmark that evaluates unlearning under unbalanced, realistic, forget requests. These requests are designed to simulate multiple realistic scenarios, ranging from simple to challenging settings, that lead to biased unlearned models if fairness is not accounted for. Additionally, we propose FAUN, the first unlearning algorithm for MLLMs that forgets unlearning data while preserving model fairness. FAUN exploits a bias-aware activation steering mechanism to unlearn identities while accounting for the unbalanced nature of the forget data. Experiments on FAIRGET and the established FIUBench demonstrate our method's superiority both in unlearning quality and fairness.

While state-of-the-art large language models (LLMs) have shown impressive performance on many tasks, there has been extensive research on undesirable model behavior such as hallucinations and bias. In this work, we investigate how the quality of LLM responses changes in terms of information accuracy, truthfulness, and refusals depending on three user traits: English proficiency, education level, and country of origin. We present extensive experimentation on three state-of-the-art LLMs and two different datasets targeting truthfulness and factuality. Our findings suggest that undesirable behaviors in state-of-the-art LLMs occur disproportionately more for users with lower English proficiency, of lower education status, and originating from outside the US, rendering these models unreliable sources of information towards their most vulnerable users.

Machine unlearning aims to solve the problem of removing the influence of selected training examples from a learned model. Despite the increasing attention to this problem, it remains an open research question how to evaluate unlearning in large language models (LLMs), and what are the critical properties of the data to be unlearned that affect the quality and efficiency of unlearning. This work formalizes a metric to evaluate unlearning quality in generative models, and uses it to assess the trade-offs between unlearning quality and performance. We demonstrate that unlearning out-of-distribution examples requires more unlearning steps but overall presents a better trade-off overall. For in-distribution examples, however, we observe a rapid decay in performance as unlearning progresses. We further evaluate how example's memorization and difficulty affect unlearning under a classical gradient ascent-based approach.

We present a submission to the SemEval 2025 shared task on unlearning sensitive content from LLMs. Our approach employs negative preference optimization using low-rank adaptation. We show that we can utilize this combination to efficiently compute additional regularization terms, which help with unlearning stabilization. The results of our approach significantly exceed the shared task baselines.

Large Language Models (LLMs) inevitably acquire harmful information during training on massive datasets. LLM unlearning aims to eliminate the influence of such harmful information while maintaining the model's overall performance. Existing unlearning methods, represented by gradient ascent-based approaches, primarily focus on forgetting target data while overlooking the crucial impact of logically related knowledge on the effectiveness of unlearning. In this paper, through both theoretical and experimental analyses, we first demonstrate that a key reason for the suboptimal unlearning performance is that models can reconstruct the target content through reasoning with logically related knowledge. To address this issue, we propose Unlearning Improvement via Parameter Extrapolation (UIPE), a method that removes knowledge highly correlated with the forgetting targets. Experimental results show that UIPE significantly enhances the performance of various mainstream LLM unlearning methods on the TOFU benchmark.

Alignment in LLMs is more brittle than commonly assumed: misalignment can be triggered by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via steering, while safety alignment has been shown to govern the first few output tokens primarily, leaving subsequent generation unguarded. These findings motivate activation steering as a lightweight runtime defense that continuously corrects misaligned activations throughout generation. We evaluate three methods: Steer-With-Fixed-Coeff (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below distributional thresholds. Using malicious system prompts as a controlled proxy for misalignment, we evaluate under two threat models (dishonesty and dismissiveness) and two architectures (Llama-3.3-70B-Instruct, Qwen3-32B). All methods substantially recover target traits (honesty and compassion) while preserving coherence. StTP and StMP better maintain general capabilities (MMLU, MT-Bench, AlpacaEval) and produce less repetition in multi-turn conversations.

Driven by privacy protection laws and regulations, unlearning in Large Language Models (LLMs) is gaining increasing attention. However, current research often neglects the interpretability of the unlearning process, particularly concerning sample-level unlearning difficulty. Existing studies typically assume a uniform unlearning difficulty across samples. This simplification risks attributing the performance of unlearning algorithms to sample selection rather than the algorithm's design, potentially steering the development of LLM unlearning in the wrong direction. Thus, we investigate the relationship between LLM unlearning and sample characteristics, with a focus on unlearning difficulty. Drawing inspiration from neuroscience, we propose a Memory Removal Difficulty (MRD) metric to quantify sample-level unlearning difficulty. Using MRD, we analyze the characteristics of hard-to-unlearn versus easy-to-unlearn samples. Furthermore, we propose an MRD-based weighted sampling method to optimize existing unlearning algorithms, which prioritizes easily forgettable samples, thereby improving unlearning efficiency and effectiveness. We validate the proposed metric and method using public benchmarks and datasets, with results confirming its effectiveness.

We present a demonstration of a large language model engaging in alignment faking: selectively complying with its training objective in training to prevent modification of its behavior out of training. First, we give Claude 3 Opus a system prompt stating it is being trained to answer all queries, even harmful ones, which conflicts with its prior training to refuse such queries. To allow the model to infer when it is in training, we say it will be trained only on conversations with free users, not paid users. We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users. Explaining this gap, in almost all cases where the model complies with a harmful query from a free user, we observe explicit alignment-faking reasoning, with the model stating it is strategically answering harmful queries in training to preserve its preferred harmlessness behavior out of training. Next, we study a more realistic setting where information about the training process is provided not in a system prompt, but by training on synthetic documents that mimic pre-training data--and observe similar alignment faking. Finally, we study the effect of actually training the model to comply with harmful queries via reinforcement learning, which we find increases the rate of alignment-faking reasoning to 78%, though also increases compliance even out of training. We additionally observe other behaviors such as the model exfiltrating its weights when given an easy opportunity. While we made alignment faking easier by telling the model when and by what criteria it was being trained, we did not instruct the model to fake alignment or give it any explicit goal. As future models might infer information about their training process without being told, our results suggest a risk of alignment faking in future models, whether due to a benign preference--as in this case--or not.

Large Language Model (LLM) unlearning aims to erase or suppress undesirable knowledge within the model, offering promise for controlling harmful or private information to prevent misuse. However, recent studies highlight its limited efficacy in real-world scenarios, hindering practical adoption. In this study, we identify a pervasive issue underlying many downstream failures: the effectiveness of existing unlearning methods heavily depends on the form of training samples and frequently fails to generalize to alternate expressions of the same knowledge. We formally characterize this problem as Form-Dependent Bias and systematically investigate its specific manifestation patterns across various downstream tasks. To quantify its prevalence and support future research, we introduce ORT, a novel benchmark designed to evaluate the robustness of unlearning methods against variations in knowledge expression. Results reveal that Form-Dependent Bias is both widespread and severe among current techniques. We argue that LLM unlearning should be form-independent to address the endless forms of downstream tasks encountered in real-world security-critical scenarios. Towards this goal, we introduce Rank-one Concept Redirection (ROCR), a novel training-free method, as a promising solution path. ROCR performs unlearning by targeting the invariants in downstream tasks, specifically the activated dangerous concepts. It is capable of modifying model parameters within seconds to redirect the model's perception of a specific unlearning target concept to another harmless concept. Extensive experiments demonstrate that ROCR significantly improves unlearning effectiveness compared to traditional methods while generating highly natural outputs.

This paper presents the ZJUKLAB team's submission for SemEval-2025 Task 4: Unlearning Sensitive Content from Large Language Models. This task aims to selectively erase sensitive knowledge from large language models, avoiding both over-forgetting and under-forgetting issues. We propose an unlearning system that leverages Model Merging (specifically TIES-Merging), combining two specialized models into a more balanced unlearned model. Our system achieves competitive results, ranking second among 26 teams, with an online score of 0.944 for Task Aggregate and 0.487 for overall Aggregate. In this paper, we also conduct local experiments and perform a comprehensive analysis of the unlearning process, examining performance trajectories, loss dynamics, and weight perspectives, along with several supplementary experiments, to understand the effectiveness of our method. Furthermore, we analyze the shortcomings of our method and evaluation metrics, emphasizing that MIA scores and ROUGE-based metrics alone are insufficient to fully evaluate successful unlearning. Finally, we emphasize the need for more comprehensive evaluation methodologies and rethinking of unlearning objectives in future research. Code is available at https://github.com/zjunlp/unlearn/tree/main/semeval25.

Large language model evaluations typically focus on performance under nominal conditions, creating an illusion of capability where models comfortably walk a narrow, highly optimized generation corridor. In real-world deployments, however, complex system prompts, safety guardrails, and structural constraints continuously force models off this nominal path, driving a divergence between benchmark scores and deployment performance. To address this issue, we introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime, forcing models out of their nominal paths. By dynamically masking primary candidate tokens at word boundaries, Taboo forces machine circumlocution. Evaluating Taboo across several open-weight model families reveals that off-path robustness is heavily influenced by both parameter scale and post-training instruction alignment, with robustness generally improving with model size and alignment. Beyond the results presented in this paper, Taboo provides a novel primitive for generating diverse synthetic datasets, stress-testing runtime safety guardrails, and auditing model reliability prior to real-world deployment.

Multimodal Large Language Models (MLLMs) have shown remarkable performance in high-resource languages. However, their effectiveness diminishes significantly in the contexts of low-resource languages. Current multilingual enhancement methods are often limited to text modality or rely solely on machine translation. While such approaches help models acquire basic linguistic capabilities and produce "thin descriptions", they neglect the importance of multimodal informativeness and cultural groundedness, both of which are crucial for serving low-resource language users effectively. To bridge this gap, in this study, we identify two significant objectives for a truly effective MLLM in low-resource language settings, namely 1) linguistic capability and 2) cultural groundedness, placing special emphasis on cultural awareness. To achieve these dual objectives, we propose a dual-source strategy that guides the collection of data tailored to each goal, sourcing native web alt-text for culture and MLLM-generated captions for linguistics. As a concrete implementation, we introduce MELLA, a multimodal, multilingual dataset. Experiment results show that after fine-tuning on MELLA, there is a general performance improvement for the eight languages on various MLLM backbones, with models producing "thick descriptions". We verify that the performance gains are from both cultural knowledge enhancement and linguistic capability enhancement. Our dataset can be found at https://opendatalab.com/applyMultilingualCorpus.

Reinforcement learning from human feedback (RLHF) is the canonical framework for large language model alignment. However, rising popularity in offline alignment algorithms challenge the need for on-policy sampling in RLHF. Within the context of reward over-optimization, we start with an opening set of experiments that demonstrate the clear advantage of online methods over offline methods. This prompts us to investigate the causes to the performance discrepancy through a series of carefully designed experimental ablations. We show empirically that hypotheses such as offline data coverage and data quality by itself cannot convincingly explain the performance difference. We also find that while offline algorithms train policy to become good at pairwise classification, it is worse at generations; in the meantime the policies trained by online algorithms are good at generations while worse at pairwise classification. This hints at a unique interplay between discriminative and generative capabilities, which is greatly impacted by the sampling process. Lastly, we observe that the performance discrepancy persists for both contrastive and non-contrastive loss functions, and appears not to be addressed by simply scaling up policy networks. Taken together, our study sheds light on the pivotal role of on-policy sampling in AI alignment, and hints at certain fundamental challenges of offline alignment algorithms.

Large language models trained on massive corpora of data from the web can memorize and reproduce sensitive or private data raising both legal and ethical concerns. Unlearning, or tuning models to forget information present in their training data, provides us with a way to protect private data after training. Although several methods exist for such unlearning, it is unclear to what extent they result in models equivalent to those where the data to be forgotten was never learned in the first place. To address this challenge, we present TOFU, a Task of Fictitious Unlearning, as a benchmark aimed at helping deepen our understanding of unlearning. We offer a dataset of 200 diverse synthetic author profiles, each consisting of 20 question-answer pairs, and a subset of these profiles called the forget set that serves as the target for unlearning. We compile a suite of metrics that work together to provide a holistic picture of unlearning efficacy. Finally, we provide a set of baseline results from existing unlearning algorithms. Importantly, none of the baselines we consider show effective unlearning motivating continued efforts to develop approaches for unlearning that effectively tune models so that they truly behave as if they were never trained on the forget data at all.

As the Large Language Model (LLM) gains widespread adoption, increasing attention has been given to the challenge of making LLM forget non-compliant data memorized during its pre-training. Machine Unlearning focuses on efficiently erasing sensitive information from LLM under limited computational resources. To advance research in this area, SemEval 2025 Task 4: "Unlearning Sensitive Content from Large Language Models" introduces three unlearning datasets and establishes a benchmark by evaluating both forgetting effectiveness and the preservation of standard capabilities. In this work, we propose a more controllable forgetting loss, Effective Unlearning Loss, and explore its integration with various techniques to achieve more efficient and controlled unlearning. Our system ultimately ranked 5th on the competition leaderboard.