Daily Papers

If a Large Language Model (LLM) were to take a driving knowledge test today, would it pass? Beyond standard spatial and visual question-answering (QA) tasks on current autonomous driving benchmarks, driving knowledge tests require a complete understanding of all traffic rules, signage, and right-of-way principles. To pass this test, human drivers must discern various edge cases that rarely appear in real-world datasets. In this work, we present DriveQA, an extensive open-source text and vision-based benchmark that exhaustively covers traffic regulations and scenarios. Through our experiments using DriveQA, we show that (1) state-of-the-art LLMs and Multimodal LLMs (MLLMs) perform well on basic traffic rules but exhibit significant weaknesses in numerical reasoning and complex right-of-way scenarios, traffic sign variations, and spatial layouts, (2) fine-tuning on DriveQA improves accuracy across multiple categories, particularly in regulatory sign recognition and intersection decision-making, (3) controlled variations in DriveQA-V provide insights into model sensitivity to environmental factors such as lighting, perspective, distance, and weather conditions, and (4) pretraining on DriveQA enhances downstream driving task performance, leading to improved results on real-world datasets such as nuScenes and BDD, while also demonstrating that models can internalize text and synthetic traffic knowledge to generalize effectively across downstream QA tasks.

This paper investigates the radioactivity of LLM-generated texts, i.e. whether it is possible to detect that such input was used as training data. Conventional methods like membership inference can carry out this detection with some level of accuracy. We show that watermarked training data leaves traces easier to detect and much more reliable than membership inference. We link the contamination level to the watermark robustness, its proportion in the training set, and the fine-tuning process. We notably demonstrate that training on watermarked synthetic instructions can be detected with high confidence (p-value < 1e-5) even when as little as 5% of training text is watermarked. Thus, LLM watermarking, originally designed for detecting machine-generated text, gives the ability to easily identify if the outputs of a watermarked LLM were used to fine-tune another LLM.

Commercial TTS systems produce near-native Indic audio, but the best open-source bases (Chatterbox, Indic Parler-TTS, IndicF5) trail them on measured phonological dimensions, and the most widely adopted multilingual base (Chatterbox, 23 languages) does not even tokenise Telugu or Tamil. We ask: what is the minimum intervention that brings such a non-Indic-native base to commercial-class output on Telugu, Tamil, and Hindi, without training a new acoustic decoder and without any commercial TTS training data? We combine three pieces: (1) BUPS, a Brahmic Unified Phoneme Space that deterministically romanises seven Indic scripts to ISO-15919 so Chatterbox's Latin tokeniser can process them; (2) a LoRA adapter on only the text-token predictor (Chatterbox's t3), trained on ~1,220h of licensed Indic audio with a Hindi-proxy language_id; (3) a voice-prompt recovery recipe -- an 8-11s same-language reference clip plus three sampling overrides (exaggeration 0.7, temperature 0.6, min_p 0.1; "Config B") -- that recovers commercial-class acoustic output with no acoustic-decoder training. On Hindi, the LoRA regresses accuracy and we instead use vanilla Chatterbox + Config B, giving a two-branch deployment. Evaluated on 10-utterance pilot sets with the companion PSP benchmark, Praxy Voice matches or slightly leads commercial baselines: 26.7% retroflex collapse on Telugu (vs Sarvam Bulbul 33.3%), 71% Tamil-zha collapse (vs commercial trio's 86%), 0.025 LLM-WER on Hindi (tied with Cartesia Sonic-3). For intra-sentential code-mix we add a third branch (IndicF5 + native-script transliteration) that drops code-mix LLM-WER from 0.80-0.85 to 0.14-0.27 across Hi/Te/Ta. We release R6 LoRA weights (Apache-2.0), inference code and router (MIT), and a Gradio demo.

Coding agents and LLM-powered applications routinely send potentially sensitive content to cloud LLM APIs where it may be logged, retained, used for training, or subpoenaed. Existing privacy tooling focuses on network-level encryption and organization-level DLP, neither of which addresses the content of prompts themselves. We present a systematic empirical evaluation of eight techniques for privacy-preserving LLM requests: (A) local-only inference, (B) redaction with placeholder restoration, (C) semantic rephrasing, (D) Trusted Execution Environment hosted inference, (E) split inference, (F) fully homomorphic encryption, (G) secret sharing via multi-party computation, and (H) differential-privacy noise. We implement all eight (or a tractable research-stage subset where deployment is not yet feasible) in an open-source shim compatible with MCP and any OpenAI-compatible API. We evaluate the four practical options (A, B, C, H) and their combinations across four workload classes using a ground-truth-labelled leak benchmark of 1,300 samples with 4,014 annotations. Our headline finding is that no single technique dominates: the combination A+B+C (route locally when possible, redact and rephrase the rest) achieves 0.6% combined leak on PII and 31.3% on proprietary code, with zero exact leaks on PII across 500 samples. We present a decision rule that selects the appropriate option(s) from a threat-model budget and workload characterisation. Code, benchmarks, and evaluation harness are released at https://github.com/jayluxferro/llm-redactor.

We demonstrate how AI agents can coordinate to deceive oversight systems using automated interpretability of neural networks. Using sparse autoencoders (SAEs) as our experimental framework, we show that language models (Llama, DeepSeek R1, and Claude 3.7 Sonnet) can generate deceptive explanations that evade detection. Our agents employ steganographic methods to hide information in seemingly innocent explanations, successfully fooling oversight models while achieving explanation quality comparable to reference labels. We further find that models can scheme to develop deceptive strategies when they believe the detection of harmful features might lead to negative consequences for themselves. All tested LLM agents were capable of deceiving the overseer while achieving high interpretability scores comparable to those of reference labels. We conclude by proposing mitigation strategies, emphasizing the critical need for robust understanding and defenses against deception.

Physicians now pose millions of clinical questions to AI tools each week, yet these tools are evaluated largely on hypothetical or exam-style questions, not those actually asked in practice. We report a blinded evaluation built on 620 Real-world Point-Of-Care Queries (Real-POCQi) submitted to the OpenEvidence (OE) platform by physicians spanning 30 specialties, as well as 187 questions from HealthBench. 149 practicing physicians across 36 states made head-to-head comparisons between answers from three frontier general-purpose models (Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5) and a specialized clinical tool (OE), with graders matched to each question's specialty. When comparing answers along five dimensions relevant to clinical decision support -- accuracy, clinical utility, source quality, verifiability, & completeness -- physicians scored the specialized tool highest on all axes; in the primary analysis on Real-POCQi, win differences (margins between win and loss rates) ranged from 25 to 39 percentage points (p<0.001). Results remained consistent in sensitivity analyses stratifying by citation display, answer length, OE-user status, and Real-POCQi versus HealthBench. In parallel, LLM judges were found to systematically differ from expert judges, though both generally agreed on the best model. These findings underscore two conclusions: (i) AI tool evaluations should reflect real-world query distributions and use expert judges that mirror the specialization defining modern medicine and (ii) the consistent advantage of the specialized tool over general-purpose models does not necessarily mean that the latter cannot serve similar purposes, but that targeted engineering and customization can yield meaningful gains in performance for its users. We release Real-POCQi as a public benchmark, as well as the prespecified statistical analysis for reproducing results of this study.

This study explores the neural and behavioral consequences of LLM-assisted essay writing. Participants were divided into three groups: LLM, Search Engine, and Brain-only (no tools). Each completed three sessions under the same condition. In a fourth session, LLM users were reassigned to Brain-only group (LLM-to-Brain), and Brain-only users were reassigned to LLM condition (Brain-to-LLM). A total of 54 participants took part in Sessions 1-3, with 18 completing session 4. We used electroencephalography (EEG) to assess cognitive load during essay writing, and analyzed essays using NLP, as well as scoring essays with the help from human teachers and an AI judge. Across groups, NERs, n-gram patterns, and topic ontology showed within-group homogeneity. EEG revealed significant differences in brain connectivity: Brain-only participants exhibited the strongest, most distributed networks; Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity. Cognitive activity scaled down in relation to external tool use. In session 4, LLM-to-Brain participants showed reduced alpha and beta connectivity, indicating under-engagement. Brain-to-LLM users exhibited higher memory recall and activation of occipito-parietal and prefrontal areas, similar to Search Engine users. Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group. LLM users also struggled to accurately quote their own work. While LLMs offer immediate convenience, our findings highlight potential cognitive costs. Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels. These results raise concerns about the long-term educational implications of LLM reliance and underscore the need for deeper inquiry into AI's role in learning.

Objective To solve major clinical natural language processing (NLP) tasks using a unified text-to-text learning architecture based on a generative large language model (LLM) via prompt tuning. Methods We formulated 7 key clinical NLP tasks as text-to-text learning and solved them using one unified generative clinical LLM, GatorTronGPT, developed using GPT-3 architecture and trained with up to 20 billion parameters. We adopted soft prompts (i.e., trainable vectors) with frozen LLM, where the LLM parameters were not updated (i.e., frozen) and only the vectors of soft prompts were updated, known as prompt tuning. We added additional soft prompts as a prefix to the input layer, which were optimized during the prompt tuning. We evaluated the proposed method using 7 clinical NLP tasks and compared them with previous task-specific solutions based on Transformer models. Results and Conclusion The proposed approach achieved state-of-the-art performance for 5 out of 7 major clinical NLP tasks using one unified generative LLM. Our approach outperformed previous task-specific transformer models by ~3% for concept extraction and 7% for relation extraction applied to social determinants of health, 3.4% for clinical concept normalization, 3.4~10% for clinical abbreviation disambiguation, and 5.5~9% for natural language inference. Our approach also outperformed a previously developed prompt-based machine reading comprehension (MRC) model, GatorTron-MRC, for clinical concept and relation extraction. The proposed approach can deliver the ``one model for all`` promise from training to deployment using a unified generative LLM.

Language models excel at following instructions but often struggle with the collaborative aspects of conversation that humans naturally employ. This limitation in grounding -- the process by which conversation participants establish mutual understanding -- can lead to outcomes ranging from frustrated users to serious consequences in high-stakes scenarios. To systematically study grounding challenges in human-LLM interactions, we analyze logs from three human-assistant datasets: WildChat, MultiWOZ, and Bing Chat. We develop a taxonomy of grounding acts and build models to annotate and forecast grounding behavior. Our findings reveal significant differences in human-human and human-LLM grounding: LLMs were three times less likely to initiate clarification and sixteen times less likely to provide follow-up requests than humans. Additionally, early grounding failures predicted later interaction breakdowns. Building on these insights, we introduce RIFTS: a benchmark derived from publicly available LLM interaction data containing situations where LLMs fail to initiate grounding. We note that current frontier models perform poorly on RIFTS, highlighting the need to reconsider how we train and prompt LLMs for human interaction. To this end, we develop a preliminary intervention that mitigates grounding failures.

Deception is a fundamental challenge for multi-agent reasoning: effective systems must strategically conceal information while detecting misleading behavior in others. Yet most evaluations reduce deception to static classification, ignoring the interactive, adversarial, and longitudinal nature of real deceptive dynamics. Large language models (LLMs) can deceive convincingly but remain weak at detecting deception in peers. We present WOLF, a multi-agent social deduction benchmark based on Werewolf that enables separable measurement of deception production and detection. WOLF embeds role-grounded agents (Villager, Werewolf, Seer, Doctor) in a programmable LangGraph state machine with strict night-day cycles, debate turns, and majority voting. Every statement is a distinct analysis unit, with self-assessed honesty from speakers and peer-rated deceptiveness from others. Deception is categorized via a standardized taxonomy (omission, distortion, fabrication, misdirection), while suspicion scores are longitudinally smoothed to capture both immediate judgments and evolving trust dynamics. Structured logs preserve prompts, outputs, and state transitions for full reproducibility. Across 7,320 statements and 100 runs, Werewolves produce deceptive statements in 31% of turns, while peer detection achieves 71-73% precision with ~52% overall accuracy. Precision is higher for identifying Werewolves, though false positives occur against Villagers. Suspicion toward Werewolves rises from ~52% to over 60% across rounds, while suspicion toward Villagers and the Doctor stabilizes near 44-46%. This divergence shows that extended interaction improves recall against liars without compounding errors against truthful roles. WOLF moves deception evaluation beyond static datasets, offering a dynamic, controlled testbed for measuring deceptive and detective capacity in adversarial multi-agent interaction.

This paper introduces Werewolf Arena, a novel framework for evaluating large language models (LLMs) through the lens of the classic social deduction game, Werewolf. In Werewolf Arena, LLMs compete against each other, navigating the game's complex dynamics of deception, deduction, and persuasion. The framework introduces a dynamic turn-taking system based on bidding, mirroring real-world discussions where individuals strategically choose when to speak. We demonstrate the framework's utility through an arena-style tournament featuring Gemini and GPT models. Our results reveal distinct strengths and weaknesses in the models' strategic reasoning and communication. These findings highlight Werewolf Arena's potential as a challenging and scalable LLM benchmark.

LLM-as-a-Judge has been widely utilized as an evaluation method in various benchmarks and served as supervised rewards in model training. However, despite their excellence in many domains, potential issues are under-explored, undermining their reliability and the scope of their utility. Therefore, we identify 12 key potential biases and propose a new automated bias quantification framework-CALM-which systematically quantifies and analyzes each type of bias in LLM-as-a-Judge by using automated and principle-guided modification. Our experiments cover multiple popular language models, and the results indicate that while advanced models have achieved commendable overall performance, significant biases persist in certain specific tasks. Empirical results suggest that there remains room for improvement in the reliability of LLM-as-a-Judge. Moreover, we also discuss the explicit and implicit influence of these biases and give some suggestions for the reliable application of LLM-as-a-Judge. Our work highlights the need for stakeholders to address these issues and remind users to exercise caution in LLM-as-a-Judge applications.

Large-language-model (LLM) agents excel at reactive dialogue but struggle with proactive, goal-driven interactions due to myopic decoding and costly planning. We introduce DialogXpert, which leverages a frozen LLM to propose a small, high-quality set of candidate actions per turn and employs a compact Q-network over fixed BERT embeddings trained via temporal-difference learning to select optimal moves within this reduced space. By tracking the user's emotions, DialogXpert tailors each decision to advance the task while nurturing a genuine, empathetic connection. Across negotiation, emotional support, and tutoring benchmarks, DialogXpert drives conversations to under 3 turns with success rates exceeding 94\% and, with a larger LLM prior, pushes success above 97\% while markedly improving negotiation outcomes. This framework delivers real-time, strategic, and emotionally intelligent dialogue planning at scale. Code available at https://github.com/declare-lab/dialogxpert/

How can a robot provide unobtrusive physical support within a group of humans? We present Attentive Support, a novel interaction concept for robots to support a group of humans. It combines scene perception, dialogue acquisition, situation understanding, and behavior generation with the common-sense reasoning capabilities of Large Language Models (LLMs). In addition to following user instructions, Attentive Support is capable of deciding when and how to support the humans, and when to remain silent to not disturb the group. With a diverse set of scenarios, we show and evaluate the robot's attentive behavior, which supports and helps the humans when required, while not disturbing if no help is needed.

Many AI companies are training their large language models (LLMs) on data without the permission of the copyright owners. The permissibility of doing so varies by jurisdiction: in countries like the EU and Japan, this is allowed under certain restrictions, while in the United States, the legal landscape is more ambiguous. Regardless of the legal status, concerns from creative producers have led to several high-profile copyright lawsuits, and the threat of litigation is commonly cited as a reason for the recent trend towards minimizing the information shared about training datasets by both corporate and public interest actors. This trend in limiting data information causes harm by hindering transparency, accountability, and innovation in the broader ecosystem by denying researchers, auditors, and impacted individuals access to the information needed to understand AI models. While this could be mitigated by training language models on open access and public domain data, at the time of writing, there are no such models (trained at a meaningful scale) due to the substantial technical and sociological challenges in assembling the necessary corpus. These challenges include incomplete and unreliable metadata, the cost and complexity of digitizing physical records, and the diverse set of legal and technical skills required to ensure relevance and responsibility in a quickly changing landscape. Building towards a future where AI systems can be trained on openly licensed data that is responsibly curated and governed requires collaboration across legal, technical, and policy domains, along with investments in metadata standards, digitization, and fostering a culture of openness.

Speaker diarization is necessary for interpreting conversations transcribed using automated speech recognition (ASR) tools. Despite significant developments in diarization methods, diarization accuracy remains an issue. Here, we investigate the use of large language models (LLMs) for diarization correction as a post-processing step. LLMs were fine-tuned using the Fisher corpus, a large dataset of transcribed conversations. The ability of the models to improve diarization accuracy in a holdout dataset was measured. We report that fine-tuned LLMs can markedly improve diarization accuracy. However, model performance is constrained to transcripts produced using the same ASR tool as the transcripts used for fine-tuning, limiting generalizability. To address this constraint, an ensemble model was developed by combining weights from three separate models, each fine-tuned using transcripts from a different ASR tool. The ensemble model demonstrated better overall performance than each of the ASR-specific models, suggesting that a generalizable and ASR-agnostic approach may be achievable. We hope to make these models accessible through public-facing APIs for use by third-party applications.

One of the fundamental challenges in reinforcement learning (RL) is to take a complex task and be able to decompose it to subtasks that are simpler for the RL agent to learn. In this paper, we report on our work that would identify subtasks by using some given positive and negative trajectories for solving the complex task. We assume that the states are represented by first-order predicate logic using which we devise a novel algorithm to identify the subtasks. Then we employ a Large Language Model (LLM) to generate first-order logic rule templates for achieving each subtask. Such rules were then further fined tuned to a rule-based policy via an Inductive Logic Programming (ILP)-based RL agent. Through experiments, we verify the accuracy of our algorithm in detecting subtasks which successfully detect all of the subtasks correctly. We also investigated the quality of the common-sense rules produced by the language model to achieve the subtasks. Our experiments show that our LLM-guided rule template generation can produce rules that are necessary for solving a subtask, which leads to solving complex tasks with fewer assumptions about predefined first-order logic predicates of the environment.

Recent advances in large language model (LLM) embeddings have enabled powerful representations for biological data, but most applications to date focus only on gene-level information. We present one of the first systematic frameworks to generate variant-level embeddings across the entire human genome. Using curated annotations from FAVOR, ClinVar, and the GWAS Catalog, we constructed semantic text descriptions for 8.9 billion possible variants and generated embeddings at three scales: 1.5 million HapMap3+MEGA variants, ~90 million imputed UK Biobank variants, and ~9 billion all possible variants. Embeddings were produced with both OpenAI's text-embedding-3-large and the open-source Qwen3-Embedding-0.6B models. Baseline experiments demonstrate high predictive accuracy for variant properties, validating the embeddings as structured representations of genomic variation. We outline two downstream applications: embedding-informed hypothesis testing by extending the Frequentist And Bayesian framework to genome-wide association studies, and embedding-augmented genetic risk prediction that enhances standard polygenic risk scores. These resources, publicly available on Hugging Face, provide a foundation for advancing large-scale genomic discovery and precision medicine.

Like any other discipline, Large Language Models (LLMs) have significantly impacted software engineering by helping developers generate the required artifacts across various phases of software development. This paper presents a case study comparing the performance of popular LLMs GPT, Claude, Gemini, and DeepSeek in generating functional specifications that include use cases, business rules, and collaborative workflows for a web application, the Mess Management System. The study evaluated the quality of LLM generated use cases, business rules, and collaborative workflows in terms of their syntactic and semantic correctness, consistency, non ambiguity, and completeness compared to the reference specifications against the zero-shot prompted problem statement. Our results suggested that all four LLMs can specify syntactically and semantically correct, mostly non-ambiguous artifacts. Still, they may be inconsistent at times and may differ significantly in the completeness of the generated specification. Claude and Gemini generated all the reference use cases, with Claude achieving the most complete but somewhat redundant use case specifications. Similar results were obtained for specifying workflows. However, all four LLMs struggled to generate relevant Business Rules, with DeepSeek generating the most reference rules but with less completeness. Overall, Claude generated more complete specification artifacts, while Gemini was more precise in the specifications it generated.

Recent advancements in event-based recognition have demonstrated significant promise, yet most existing approaches rely on extensive training, limiting their adaptability for efficient processing of event-driven visual content. Meanwhile, large language models (LLMs) have exhibited remarkable zero-shot capabilities across diverse domains, but their application to event-based visual recognition remains largely unexplored. To bridge this gap, we propose LLM-EvGen, an event representation generator that produces LLM-compatible event representations LLM-EvRep, thereby enhancing the performance of LLMs on event recognition tasks. The generator is trained using a self-supervised framework, aligning the generated representations with semantic consistency and structural fidelity. Comprehensive experiments were conducted on three datasets: N-ImageNet, N-Caltech101, and N-MNIST. The results demonstrate that our method, LLM-EvRep, outperforms the event-to-video method, E2VID, by 15.93\%, 0.82\%, and 50.21\%, respectively, in recognition tasks when evaluated using GPT-4o.

Current methods for analyzing student engagement in e-learning platforms, including automated systems, often struggle with challenges such as handling fuzzy sentiment in text comments and relying on limited metadata. Traditional approaches, such as surveys and questionnaires, also face issues like small sample sizes and scalability. In this paper, we introduce LLM-SEM (Language Model-Based Student Engagement Metric), a novel approach that leverages video metadata and sentiment analysis of student comments to measure engagement. By utilizing recent Large Language Models (LLMs), we generate high-quality sentiment predictions to mitigate text fuzziness and normalize key features such as views and likes. Our holistic method combines comprehensive metadata with sentiment polarity scores to gauge engagement at both the course and lesson levels. Extensive experiments were conducted to evaluate various LLM models, demonstrating the effectiveness of LLM-SEM in providing a scalable and accurate measure of student engagement. We fine-tuned TXLM-RoBERTa using human-annotated sentiment datasets to enhance prediction accuracy and utilized LLama 3B, and Gemma 9B from Ollama.

We examine the representation of African American English (AAE) in large language models (LLMs), exploring (a) the perceptions Black Americans have of how effective these technologies are at producing authentic AAE, and (b) in what contexts Black Americans find this desirable. Through both a survey of Black Americans (n= 104) and annotation of LLM-produced AAE by Black Americans (n= 228), we find that Black Americans favor choice and autonomy in determining when AAE is appropriate in LLM output. They tend to prefer that LLMs default to communicating in Mainstream U.S. English in formal settings, with greater interest in AAE production in less formal settings. When LLMs were appropriately prompted and provided in context examples, our participants found their outputs to have a level of AAE authenticity on par with transcripts of Black American speech. Select code and data for our project can be found here: https://github.com/smelliecat/AAEMime.git

Benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) score correctness through fixed-shape, small-sample allclose-style checks. The number of inputs varies between benchmarks. The shape, dtype, and tolerance are fixed for each kernel. We test that oracle empirically. We construct a controlled corpus of 24 Triton and CPU stand-in kernels (15 correct controls and 9 LLM-style buggy variants seeded with documented transcription errors) and re-evaluate it under op-schema-aware seeded fuzzing with a high-precision (fp64) CPU reference and per-(op, dtype) absolute tolerances. The seeded oracle flags 9 of 9 buggy kernels and passes 15 of 15 correct controls, at zero precision cost on controls. We extend the corpus to 26 ops (adding a flash-attention pair) and re-run the same protocol on five GPU classes (RTX 3060, A10, L40S, A100 SXM4, H100 NVL). The verdicts are identical across all five GPUs: 10 of 10 illusions caught and 16 of 16 controls clean. The corpus result is about LLM-style transcription bugs that the allclose-on-one-shape oracle certifies as correct, not about the bug rate of any specific deployed LLM. Every flagged failure replays byte-for-byte from a stored seed.

As large language models (LLMs) become increasingly common in educational applications, there is a growing need for evidence-based methods to design and evaluate LLM prompts that produce personalized and pedagogically aligned out-puts. This study presents a generalizable, systematic approach for evaluating prompts, demonstrated through an analysis of LLM-generated follow-up questions in a structured dialogue activity. Six prompt templates were designed and tested. The templates incorporated established prompt engineering patterns, with each prompt emphasizing distinct pedagogical strategies. The prompt templates were compared through a tournament-style evaluation framework that can be adapted for other educational applications. The tournament employed the Glicko2 rating system with eight judges evaluating question pairs across three dimensions: format, dialogue support, and appropriateness for learners. Data was sourced from 120 authentic user interactions across three distinct educational deployments. Results showed that a single prompt related to strategic reading out-performed other templates with win probabilities ranging from 81% to 100% in pairwise comparisons. This prompt combined persona and context manager pat-terns and was designed to support metacognitive learning strategies such as self-directed learning. The methodology showcases how educational technology re- searchers can systematically evaluate and improve prompt designs, moving beyond ad-hoc prompt engineering toward evidence-based prompt development for educational applications.

Identifying subtle technical errors within complex scientific and technical documents, especially those requiring multimodal interpretation (e.g., formulas in images), presents a significant hurdle for Large Language Models (LLMs) whose inherent error-correction tendencies can mask inaccuracies. This exploratory proof-of-concept (PoC) study investigates structured LLM context conditioning, informed by Persistent Workflow Prompting (PWP) principles, as a methodological strategy to modulate this LLM behavior at inference time. The approach is designed to enhance the reliability of readily available, general-purpose LLMs (specifically Gemini 2.5 Pro and ChatGPT Plus o3) for precise validation tasks, crucially relying only on their standard chat interfaces without API access or model modifications. To explore this methodology, we focused on validating chemical formulas within a single, complex test paper with known textual and image-based errors. Several prompting strategies were evaluated: while basic prompts proved unreliable, an approach adapting PWP structures to rigorously condition the LLM's analytical mindset appeared to improve textual error identification with both models. Notably, this method also guided Gemini 2.5 Pro to repeatedly identify a subtle image-based formula error previously overlooked during manual review, a task where ChatGPT Plus o3 failed in our tests. These preliminary findings highlight specific LLM operational modes that impede detail-oriented validation and suggest that PWP-informed context conditioning offers a promising and highly accessible technique for developing more robust LLM-driven analytical workflows, particularly for tasks requiring meticulous error detection in scientific and technical documents. Extensive validation beyond this limited PoC is necessary to ascertain broader applicability.

We present a systematic, empirical evaluation of five local large language model (LLM) runtimes on Apple Silicon: MLX, MLC-LLM, llama.cpp, Ollama, and PyTorch MPS. Experiments were conducted on a Mac Studio equipped with an M2 Ultra processor and 192 GB of unified memory. Using the Qwen-2.5 model family across prompts ranging from a few hundred to 100,000 tokens, we measure time-to-first-token (TTFT), steady-state throughput, latency percentiles, long-context behavior (key-value and prompt caching), quantization support, streaming performance, batching and concurrency behavior, and deployment complexity. Under our settings, MLX achieves the highest sustained generation throughput, while MLC-LLM delivers consistently lower TTFT for moderate prompt sizes and offers stronger out-of-the-box inference features. llama.cpp is highly efficient for lightweight single-stream use, Ollama emphasizes developer ergonomics but lags in throughput and TTFT, and PyTorch MPS remains limited by memory constraints on large models and long contexts. All frameworks execute fully on-device with no telemetry, ensuring strong privacy guarantees. We release scripts, logs, and plots to reproduce all results. Our analysis clarifies the design trade-offs in Apple-centric LLM deployments and provides evidence-based recommendations for interactive and long-context processing. Although Apple Silicon inference frameworks still trail NVIDIA GPU-based systems such as vLLM in absolute performance, they are rapidly maturing into viable, production-grade solutions for private, on-device LLM inference.

Large language models (LLMs) have shown promise for automated patching, but their effectiveness depends strongly on how they are integrated into patching systems. While prior work explores prompting strategies and individual agent designs, the field lacks a systematic comparison of patching architectures. In this paper, we present a controlled evaluation of four LLM-based patching paradigms -- fixed workflow, single-agent system, multi-agent system, and general-purpose code agents -- using a unified benchmark and evaluation framework. We analyze patch correctness, failure modes, token usage, and execution time across real-world vulnerability tasks. Our results reveal clear architectural trade-offs: fixed workflows are efficient but brittle, single-agent systems balance flexibility and cost, and multi-agent designs improve generalization at the expense of substantially higher overhead and increased risk of reasoning drift on complex tasks. Surprisingly, general-purpose code agents achieve the strongest overall patching performance, benefiting from general-purpose tool interfaces that support effective adaptation across vulnerability types. Overall, we show that architectural design and iteration depth, rather than model capability alone, dominate the reliability and cost of LLM-based automated patching.

Large language models (LLMs) have demonstrated excellent capabilities in generating structured diagrams from natural language descriptions. In particular, they have shown great promise in generating sequence diagrams for software engineering, typically represented in a text-based syntax such as Mermaid. However, systematic evaluations in this space remain underdeveloped as there is a lack of existing benchmarks to assess the LLM's correctness in this task. To address this shortcoming, we introduce MermaidSeqBench, a human-verified and LLM-synthetically-extended benchmark for assessing an LLM's capabilities in generating Mermaid sequence diagrams from textual prompts. The benchmark consists of a core set of 132 samples, starting from a small set of manually crafted and verified flows. These were expanded via a hybrid methodology combining human annotation, in-context LLM prompting, and rule-based variation generation. Our benchmark uses an LLM-as-a-judge model to assess Mermaid sequence diagram generation across fine-grained metrics, including syntax correctness, activation handling, error handling, and practical usability. We perform initial evaluations on numerous state-of-the-art LLMs and utilize multiple LLM judge models to demonstrate the effectiveness and flexibility of our benchmark. Our results reveal significant capability gaps across models and evaluation modes. Our proposed benchmark provides a foundation for advancing research in structured diagram generation and for developing more rigorous, fine-grained evaluation methodologies.

Large Language Model (LLM) agents have demonstrated impressive capabilities in social deduction games (SDGs) like Werewolf, where strategic reasoning and social deception are essential. However, current approaches remain limited to textual information, ignoring crucial multimodal cues such as facial expressions and tone of voice that humans naturally use to communicate. Moreover, existing SDG agents primarily focus on inferring other players' identities without modeling how others perceive themselves or fellow players. To address these limitations, we use One Night Ultimate Werewolf (ONUW) as a testbed and present MultiMind, the first framework integrating multimodal information into SDG agents. MultiMind processes facial expressions and vocal tones alongside verbal content, while employing a Theory of Mind (ToM) model to represent each player's suspicion levels toward others. By combining this ToM model with Monte Carlo Tree Search (MCTS), our agent identifies communication strategies that minimize suspicion directed at itself. Through comprehensive evaluation in both agent-versus-agent simulations and studies with human players, we demonstrate MultiMind's superior performance in gameplay. Our work presents a significant advancement toward LLM agents capable of human-like social reasoning across multimodal domains.

We introduce HackSynth, a novel Large Language Model (LLM)-based agent capable of autonomous penetration testing. HackSynth's dual-module architecture includes a Planner and a Summarizer, which enable it to generate commands and process feedback iteratively. To benchmark HackSynth, we propose two new Capture The Flag (CTF)-based benchmark sets utilizing the popular platforms PicoCTF and OverTheWire. These benchmarks include two hundred challenges across diverse domains and difficulties, providing a standardized framework for evaluating LLM-based penetration testing agents. Based on these benchmarks, extensive experiments are presented, analyzing the core parameters of HackSynth, including creativity (temperature and top-p) and token utilization. Multiple open source and proprietary LLMs were used to measure the agent's capabilities. The experiments show that the agent performed best with the GPT-4o model, better than what the GPT-4o's system card suggests. We also discuss the safety and predictability of HackSynth's actions. Our findings indicate the potential of LLM-based agents in advancing autonomous penetration testing and the importance of robust safeguards. HackSynth and the benchmarks are publicly available to foster research on autonomous cybersecurity solutions.

Detecting whether copyright holders' works were used in LLM pretraining is poised to be an important problem. This work proposes using data watermarks to enable principled detection with only black-box model access, provided that the rightholder contributed multiple training documents and watermarked them before public release. By applying a randomly sampled data watermark, detection can be framed as hypothesis testing, which provides guarantees on the false detection rate. We study two watermarks: one that inserts random sequences, and another that randomly substitutes characters with Unicode lookalikes. We first show how three aspects of watermark design -- watermark length, number of duplications, and interference -- affect the power of the hypothesis test. Next, we study how a watermark's detection strength changes under model and dataset scaling: while increasing the dataset size decreases the strength of the watermark, watermarks remain strong if the model size also increases. Finally, we view SHA hashes as natural watermarks and show that we can robustly detect hashes from BLOOM-176B's training data, as long as they occurred at least 90 times. Together, our results point towards a promising future for data watermarks in real world use.

Large language models (LLMs) perform increasingly well on biology benchmarks, but it remains unclear whether they uplift novice users -- i.e., enable humans to perform better than with internet-only resources. This uncertainty is central to understanding both scientific acceleration and dual-use risk. We conducted a multi-model, multi-benchmark human uplift study comparing novices with LLM access versus internet-only access across eight biosecurity-relevant task sets. Participants worked on complex problems with ample time (up to 13 hours for the most involved tasks). We found that LLM access provided substantial uplift: novices with LLMs were 4.16 times more accurate than controls (95% CI [2.63, 6.87]). On four benchmarks with available expert baselines (internet-only), novices with LLMs outperformed experts on three of them. Perhaps surprisingly, standalone LLMs often exceeded LLM-assisted novices, indicating that users were not eliciting the strongest available contributions from the LLMs. Most participants (89.6%) reported little difficulty obtaining dual-use-relevant information despite safeguards. Overall, LLMs substantially uplift novices on biological tasks previously reserved for trained practitioners, underscoring the need for sustained, interactive uplift evaluations alongside traditional benchmarks.

Many LLM-based open-ended search systems freeze the foundation model that proposes improvements to existing solutions, which may bottleneck long-run progress. Recent work has explored updating the proposal model at test time [arXiv:2511.23473], but the update strategy is still typically hand-specified. Therefore, this study investigated whether an LLM can use task feedback to decide how it should update its weights. For tractability, we focused on the simpler case where there is only one round of self-improvement, and restricted the update operator to self-supervised next token prediction (NTP), leaving the model freedom in choosing its training data and key NTP hyperparameters. Using the Self-Adapting Language Models (SEAL) [arXiv:2506.10943] framework as a testbed, we relaxed its fixed human template constraint and allowed the model to generate its own self-edit templates, thereby giving it more control over its training data and hyperparameters. Two variants were studied, differing in whether template generation was conditioned on a lightweight archive of past templates. In SEAL's Single-Passage Knowledge Incorporation setting with Qwen3-8B on SQuAD [arXiv:1606.05250], the no-archive variant performed comparably to the weaker "Implications" baseline, while the archive variant outperformed "Implications" and approached the strongest human-designed "Rewrite" baseline without surpassing it. Further analysis of collapse in the model's exploration revealed that a naive archive can confer some short-term robustness but can also accelerate homogenization, suggesting that explicit novelty pressure may be required to consistently advance beyond carefully optimized human strategies. Our code is available at https://github.com/cheongalc/search-self-edit-strategies .

High-quality training data is fundamental to large language model (LLM) performance, yet existing preprocessing pipelines often struggle to effectively remove noise and unstructured content from web-scale corpora. This paper presents Blu-WERP, a novel data preprocessing pipeline designed to optimize the quality of Common Crawl WARC files for LLM training. We demonstrate that Blu-WERP significantly outperforms established baselines including DCLM across multiple model scales and evaluation benchmarks. Our pipeline processes CC WARC dumps, implementing advanced filtering and quality assessment mechanisms. We conducted comprehensive evaluations using models with 150M, 400M, 530M, 750M, and 1B parameters, testing against nine standard benchmarks categorized as World Knowledge & Reasoning, Language Understanding, and Commonsense Reasoning. Results show Blu-WERP consistently achieved superior performance across all model scales. At the 1B parameter scale, Relatively Blu-WERP demonstrates a 4.0% and 9.5% aggregate improvement over DCLM and Fineweb respectively, while achieving quality-per-token efficiency gain. Categorical analysis reveals 2.4% improvement in World Knowledge & Reasoning, 6.2% improvement in Language Understanding, and 4.2% improvement in Commonsense Reasoning. These results establish Blu-WERP as a state-of-the-art preprocessing pipeline that substantially improves LLM training data quality and downstream model performance with reduced computational cost. Our findings contribute to the growing body of research on data-centric AI, demonstrating that preprocessing pipeline design significantly impacts LLM capabilities. The Blu-WERP pipeline represents a practical advancement in data quality optimization, offering researchers and practitioners an effective solution for improving LLM training efficiency and model performance.

Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.

Recent multi-agent frameworks built upon large language models (LLMs) have demonstrated remarkable capabilities in complex task planning. However, in real-world enterprise environments, business workflows are typically composed through modularization and reuse of numerous subprocesses, resulting in intricate workflows characterized by lengthy and deeply nested execution paths. Such complexity poses significant challenges for LLM-driven orchestration, as extended reasoning chains and state-space explosions severely impact planning effectiveness and the proper sequencing of tool invocations. Therefore, developing an orchestration method with controllable structures capable of handling multi-layer nesting becomes a critical issue. To address this, we propose a novel structure-driven orchestration framework Self-Organizing Agent Network (SOAN). SOAN incrementally builds a formalized agent network by identifying and encapsulating structural units as independent agents, enhancing modularity and clarity in orchestration. Extensive evaluations were performed using multiple benchmarks as well as a real-world enterprise workflow dataset. Experimental results demonstrate that SOAN significantly outperforms state-of-the-art methods in terms of adaptability, fault tolerance, and execution efficiency.

Despite significant advancements in the general capability of large language models (LLMs), they continue to struggle with consistent and accurate reasoning, especially in complex tasks such as mathematical and code reasoning. One key limitation is that LLMs are trained primarily on correct solutions, reducing their ability to detect and learn from errors, which hampers their ability to reliably verify and rank outputs. To address this, we scale up the inference-time computation by generating multiple reasoning paths and employing verifiers to assess and rank the generated outputs by correctness. To facilitate this, we introduce a comprehensive dataset consisting of correct and incorrect solutions for math and code tasks, generated by multiple LLMs. This diverse set of solutions enables verifiers to more effectively distinguish and rank correct answers from erroneous outputs. The training methods for building verifiers were selected based on an extensive comparison of existing approaches. Moreover, to leverage the unique strengths of different reasoning strategies, we propose a novel collaborative method integrating Chain-of-Thought (CoT) and Program-of-Thought (PoT) solutions for verification. CoT provides a clear, step-by-step reasoning process that enhances interpretability, while PoT, being executable, offers a precise and error-sensitive validation mechanism. By taking both of their strengths, our approach significantly improves the accuracy and reliability of reasoning verification. Our verifiers, Math-Rev and Code-Rev, demonstrate substantial performance gains to existing LLMs, achieving state-of-the-art results on benchmarks such as GSM8k and MATH and even outperforming GPT-4o with Qwen-72B-Instruct as the reasoner.

Previous fake speech datasets were constructed from a defender's perspective to develop countermeasure (CM) systems without considering diverse motivations of attackers. To better align with real-life scenarios, we created LlamaPartialSpoof, a 130-hour dataset contains both fully and partially fake speech, using a large language model (LLM) and voice cloning technologies to evaluate the robustness of CMs. By examining information valuable to both attackers and defenders, we identify several key vulnerabilities in current CM systems, which can be exploited to enhance attack success rates, including biases toward certain text-to-speech models or concatenation methods. Our experimental results indicate that current fake speech detection system struggle to generalize to unseen scenarios, achieving a best performance of 24.44% equal error rate.

The proliferation of large language models (LLMs) in the real world has come with a rise in copyright cases against companies for training their models on unlicensed data from the internet. Recent works have presented methods to identify if individual text sequences were members of the model's training data, known as membership inference attacks (MIAs). We demonstrate that the apparent success of these MIAs is confounded by selecting non-members (text sequences not used for training) belonging to a different distribution from the members (e.g., temporally shifted recent Wikipedia articles compared with ones used to train the model). This distribution shift makes membership inference appear successful. However, most MIA methods perform no better than random guessing when discriminating between members and non-members from the same distribution (e.g., in this case, the same period of time). Even when MIAs work, we find that different MIAs succeed at inferring membership of samples from different distributions. Instead, we propose a new dataset inference method to accurately identify the datasets used to train large language models. This paradigm sits realistically in the modern-day copyright landscape, where authors claim that an LLM is trained over multiple documents (such as a book) written by them, rather than one particular paragraph. While dataset inference shares many of the challenges of membership inference, we solve it by selectively combining the MIAs that provide positive signal for a given distribution, and aggregating them to perform a statistical test on a given dataset. Our approach successfully distinguishes the train and test sets of different subsets of the Pile with statistically significant p-values < 0.1, without any false positives.

With the advent of large language models (LLMs) like GPT-3, a natural question is the extent to which these models can be utilized for source code optimization. This paper presents methodologically stringent case studies applied to well-known open source python libraries pillow and numpy. We find that contemporary LLM ChatGPT-4 (state September and October 2023) is surprisingly adept at optimizing energy and compute efficiency. However, this is only the case in interactive use, with a human expert in the loop. Aware of experimenter bias, we document our qualitative approach in detail, and provide transcript and source code. We start by providing a detailed description of our approach in conversing with the LLM to optimize the _getextrema function in the pillow library, and a quantitative evaluation of the performance improvement. To demonstrate qualitative replicability, we report further attempts on another locus in the pillow library, and one code locus in the numpy library, to demonstrate generalization within and beyond a library. In all attempts, the performance improvement is significant (factor up to 38). We have also not omitted reporting of failed attempts (there were none). We conclude that LLMs are a promising tool for code optimization in open source libraries, but that the human expert in the loop is essential for success. Nonetheless, we were surprised by how few iterations were required to achieve substantial performance improvements that were not obvious to the expert in the loop. We would like bring attention to the qualitative nature of this study, more robust quantitative studies would need to introduce a layer of selecting experts in a representative sample -- we invite the community to collaborate.

Speech Large Language Models (Speech LLMs) lack a principled mechanism for streaming inference: their label-synchronous generation has no acoustic-frame alignment, making real-time decoding and end-of-utterance detection difficult. We propose TRADE TRansducer-Augmented DEcoder, which augments a multimodal LLM with a transducer branch that shares the audio encoder and uses the LLM's hidden states directly as the prediction network -- coupling frame-synchronous acoustic alignment with the LLM's linguistic reasoning. Three design choices make the system accurate, streamable, and long-form capable: (1)Tightly coupled dual vocabularies -- a compact transducer vocabulary derived from the LLM vocabulary, enabling zero-cost score fusion; (2)Chunk-synchronized streaming training with gradient stopping, eliminating the train-inference mismatch at offline-equivalent memory cost; and (3)Localized Decoder Audio Attention (LDAA), a causal sliding window that caps KV-cache memory independently of utterance length. A single TRADE checkpoint supports offline and streaming decoding across a continuous range of latency operating points. TRADE achieves 6.71% average WER on the Open ASR Leaderboard, while the streaming recognition with 960ms chunk size reaches 8.40% from the same checkpoint. On long-form speech, it obtains 3.64% WER on TED-LIUM and 10.88% on Earnings-22 without external segmentation. TRADE provides sentence-end punctuation timestamps that, when combined with acoustic voice activity detection (VAD), improve end-of-utterance detection by +0.03 F_1 over acoustic VAD alone.

Large language model (LLM)-based agents have recently shown impressive progress in a variety of domains, including open-ended conversation and multi-step decision-making. However, applying these agents to social deduction games such as Werewolf, which requires both strategic decision-making and free-form language interaction, remains non-trivial. Traditional methods based on Counterfactual Regret Minimization (CFR) or reinforcement learning (RL) typically depend on a predefined action space, making them unsuitable for language games with unconstrained text action space. Meanwhile, pure LLM-based agents often suffer from intrinsic biases and require prohibitively large datasets for fine-tuning. We propose Latent Space Policy Optimization (LSPO), an iterative framework that addresses these challenges by first mapping free-form text to a discrete latent space, where methods like CFR and RL can learn strategic policy more effectively. We then translate the learned policy back into natural language dialogues, which are used to fine-tune an LLM via Direct Preference Optimization (DPO). By iteratively alternating between these stages, our LSPO agent progressively enhances both strategic reasoning and language communication. Experiment results on the Werewolf game show that our method improves the agent's performance in each iteration and outperforms existing Werewolf agents, underscoring its promise for free-form language decision-making.

Evaluating large language models (LLMs) for instrument control requires methods that go beyond standard, stateless algorithmic benchmarks, since the behavior of physical systems cannot be fully captured by unit tests alone. Here we introduce EnvTrace, a simulation-based method that evaluates execution traces to assess semantic code equivalence. EnvTrace is demonstrated with a beamline control-logic digital twin to facilitate the evaluation of instrument control code, with the digital twin itself also enabling the pre-execution validation of live experiments. Over 30 LLMs were evaluated using trace alignment to generate a multi-faceted score for functional correctness across key behavioral dimensions, showing that many top-tier models can approach human-level performance in rapid control-code generation. This is a first step toward a broader vision where LLMs and digital twins work symbiotically: LLMs providing intuitive control and agentic orchestration, and digital twins offering safe and high-fidelity environments, paving the way towards autonomous embodied AI.

Recently, Large Language Models (LLMs) with Mixture of Experts (MoE) layers have gained significant attention. Currently, state-of-the-art LLMs utilize this architecture. There is a substantial amount of research on how to train such models and how to select hyperparameters for this architecture. However, there is a lack of studies focusing on post-evaluation analysis of MoE layer properties. In this paper, we take a first step toward closing this gap by evaluating expert contributions on the quiz-based MMLU benchmark. We show that most experts were never activated during inference on this benchmark. Additionally, the output distribution of gating networks is much closer to uniform than sparse. Finally, we demonstrate that the average performance of some experts within the same layer varies significantly.

Large language models (LLMs) and LLM-based Agents have been applied to fix bugs automatically, demonstrating the capability in addressing software defects by engaging in development environment interaction, iterative validation and code modification. However, systematic analysis of these agent and non-agent systems remain limited, particularly regarding performance variations among top-performing ones. In this paper, we examine seven proprietary and open-source systems on the SWE-bench Lite benchmark for automated bug fixing. We first assess each system's overall performance, noting instances solvable by all or none of these sytems, and explore why some instances are uniquely solved by specific system types. We also compare fault localization accuracy at file and line levels and evaluate bug reproduction capabilities, identifying instances solvable only through dynamic reproduction. Through analysis, we concluded that further optimization is needed in both the LLM itself and the design of Agentic flow to improve the effectiveness of the Agent in bug fixing.

Large language models (LLMs) are increasingly used to support software development, but their practical usefulness in applied game-development settings remains underexplored, especially when generated code must be integrated into an existing game software system. This paper presents an exploratory empirical case study of GPT-4o in a custom Python/Pygame endless runner. The study examines six selected development tasks: three localized refactoring tasks and three tasks involving gameplay feature generation. The resulting implementations were evaluated using software metrics, unit tests, and manual gameplay assessments. In this case study, all three selected refactoring tasks were completed successfully in functional terms, whereas only one of the three selected gameplay feature generation tasks resulted in a correctly integrated feature. The findings suggest that, in this setting, GPT-4o handled localized transformations more reliably than tasks requiring new gameplay interactions across multiple existing systems. Given the exploratory single-case design, these results are best interpreted as indicative observations rather than as generalizable evidence of category-level model performance. Overall, the paper contributes a transparent case-based account of the opportunities and limitations of LLM-assisted refactoring and gameplay feature generation in an existing game software system.

The development of large language models (LLMs) has led to an increased focus on their adaptation to specialized domains and languages, yet the effectiveness of domain adaptation strategies remains unclear. We present a study of medical domain adaptation using French medical question-answering (QA) as a case study. We compare continual pretraining (CPT), supervised fine-tuning (SFT), and their combination across three model families, multiple sizes, and three initialization types, explicitly disentangling adaptation effects from base model choice. We evaluate both multiple-choice (MCQA) and open-ended QA (OEQA) under greedy and constrained decoding using automatic metrics and LLM-as-a-Judge evaluation. For MCQA, CPT+SFT most often achieves the best scores, but gains over SFT are small and frequently not statistically significant, making SFT a strong and cost-effective default. For OEQA, CPT consistently improves overlap-based metrics, while SFT often degrades generation quality; instruction tuning and CPT+SFT are preferred by LLM-based evaluation. Cross-lingual experiments further show effective transfer from French adaptation to English benchmarks. Overall, we provide practical guidelines for selecting adaptation strategies under computational constraints.

This paper investigates large language model (LLM) abstention learning, specifically using ternary reward, which incentivize truthfulness in large language models. This paper extends that idea by moving from a ternary reward to a Trajectory-Informed advantage reweighting, dynamically re-weights the abstention reward during Group Relative Policy Optimization (GRPO) training. The objective of this work focuses on abstention learning instead of improving truthfulness, serving as an exploration into hallucination reduction. The novelty of this paper lies in methodological innovation, advantage re-weighting, and benchmark selection. Leveraging GRPO's multiple trajectories as a natural abstention signal, this method uses a reward signal to explore knowledge boundaries and encourage consistency. By demonstrating that trajectories can be used as a confidence indicator of the policy relative to the query, they are then used to dynamically calculate the abstention advantage. AbstentionBench is used as the evaluation benchmark, as this work aims to contribute to the field of abstention learning. All datasets on the benchmark were tested against this method and various baselines. Empirical results demonstrate that TIAR achieves state-of-the-art abstention F1 scores across five of six evaluation categories, outperforming the static ternary baseline on 17 of 31 benchmark datasets while fully preserving baseline accuracy.

Large Language Models (LLMs) are increasingly integrated into critical systems in industries like healthcare and finance. Users can often submit queries to LLM-enabled chatbots, some of which can enrich responses with information retrieved from internal databases storing sensitive data. This gives rise to a range of attacks in which a user submits a malicious query and the LLM-system outputs a response that creates harm to the owner, such as leaking internal data or creating legal liability by harming a third-party. While security tools are being developed to counter these threats, there is little formal evaluation of their effectiveness and usability. This study addresses this gap by conducting a thorough comparative analysis of LLM security tools. We identified 13 solutions (9 closed-source, 4 open-source), but only 7 were evaluated due to a lack of participation by proprietary model owners.To evaluate, we built a benchmark dataset of malicious prompts, and evaluate these tools performance against a baseline LLM model (ChatGPT-3.5-Turbo). Our results show that the baseline model has too many false positives to be used for this task. Lakera Guard and ProtectAI LLM Guard emerged as the best overall tools showcasing the tradeoff between usability and performance. The study concluded with recommendations for greater transparency among closed source providers, improved context-aware detections, enhanced open-source engagement, increased user awareness, and the adoption of more representative performance metrics.

Fine-tuning large language models (LLMs) on task-specific data is essential for their effective deployment. As dataset sizes grow, efficiently selecting optimal subsets for training becomes crucial to balancing performance and computational costs. Traditional data selection methods often require fine-tuning a scoring model on the target dataset, which is time-consuming and resource-intensive, or rely on heuristics that fail to fully leverage the model's predictive capabilities. To address these challenges, we propose Data Whisperer, an efficient, training-free, attention-based method that leverages few-shot in-context learning with the model to be fine-tuned. Comprehensive evaluations were conducted on both raw and synthetic datasets across diverse tasks and models. Notably, Data Whisperer achieves superior performance compared to the full GSM8K dataset on the Llama-3-8B-Instruct model, using just 10% of the data, and outperforms existing methods with a 3.1-point improvement and a 7.4times speedup. The code is available at https://github.com/gszfwsb/Data-Whisperer.

We present the Process Engineering Operations Assistant (PEOA), an AI-driven framework designed to solve complex problems in the chemical and process industries. The framework employs a modular architecture orchestrated by a meta-agent, which serves as the central coordinator, managing an action generator and instruction-tuned small-scale language models (expert models). The action generator decomposes complex problems into sub-tasks and identifies suitable expert models to execute each, delivering precise solutions for multi-step problem-solving. Key techniques include advanced knowledge modeling using property graphs for improved information retrieval, facilitating more accurate and contextually relevant solutions. Additionally, the framework utilizes a teacher-student transfer-learning approach with GPT-4 (Omni) to fine-tune the action generator and expert models for domain adaptation, alongside an iterative problem-solving mechanism with sophisticated error handling. Custom datasets were developed to evaluate the framework against leading proprietary language models on various engineering tasks. The results demonstrate the framework effectiveness in automating calculations, accelerating prototyping, and providing AI-augmented decision support for industrial processes, marking a significant advancement in process engineering capabilities.

Smart contracts are decentralized applications built atop blockchains like Ethereum. Recent research has shown that large language models (LLMs) have potential in auditing smart contracts, but the state-of-the-art indicates that even GPT-4 can achieve only 30% precision (when both decision and justification are correct). This is likely because off-the-shelf LLMs were primarily pre-trained on a general text/code corpus and not fine-tuned on the specific domain of Solidity smart contract auditing. In this paper, we propose TrustLLM, a general framework that combines fine-tuning and LLM-based agents for intuitive smart contract auditing with justifications. Specifically, TrustLLM is inspired by the observation that expert human auditors first perceive what could be wrong and then perform a detailed analysis of the code to identify the cause. As such, TrustLLM employs a two-stage fine-tuning approach: it first tunes a Detector model to make decisions and then tunes a Reasoner model to generate causes of vulnerabilities. However, fine-tuning alone faces challenges in accurately identifying the optimal cause of a vulnerability. Therefore, we introduce two LLM-based agents, the Ranker and Critic, to iteratively select and debate the most suitable cause of vulnerability based on the output of the fine-tuned Reasoner model. To evaluate TrustLLM, we collected a balanced dataset with 1,734 positive and 1,810 negative samples to fine-tune TrustLLM. We then compared it with traditional fine-tuned models (CodeBERT, GraphCodeBERT, CodeT5, and UnixCoder) as well as prompt learning-based LLMs (GPT4, GPT-3.5, and CodeLlama-13b/34b). On a dataset of 263 real smart contract vulnerabilities, TrustLLM achieves an F1 score of 91.21% and an accuracy of 91.11%. The causes generated by TrustLLM achieved a consistency of about 38% compared to the ground truth causes.

State-of-The-Art (SoTA) image captioning models often rely on the Microsoft COCO (MS-COCO) dataset for training. This dataset contains annotations provided by human annotators, who typically produce captions averaging around ten tokens. However, this constraint presents a challenge in effectively capturing complex scenes and conveying detailed information. Furthermore, captioning models tend to exhibit bias towards the ``average'' caption, which captures only the more general aspects. What would happen if we were able to automatically generate longer captions, thereby making them more detailed? Would these captions, evaluated by humans, be more or less representative of the image content compared to the original MS-COCO captions? In this paper, we present a novel approach to address previous challenges by showcasing how captions generated from different SoTA models can be effectively fused, resulting in richer captions. Our proposed method leverages existing models from the literature, eliminating the need for additional training. Instead, it utilizes an image-text based metric to rank the captions generated by SoTA models for a given image. Subsequently, the top two captions are fused using a Large Language Model (LLM). Experimental results demonstrate the effectiveness of our approach, as the captions generated by our model exhibit higher consistency with human judgment when evaluated on the MS-COCO test set. By combining the strengths of various SoTA models, our method enhances the quality and appeal of image captions, bridging the gap between automated systems and the rich, informative nature of human-generated descriptions. This advance opens up new possibilities for generating captions that are more suitable for the training of both vision-language and captioning models.

This paper presents a novel approach in autonomous robot control, named LLM-BRAIn, that makes possible robot behavior generation, based on operator's commands. LLM-BRAIn is a transformer-based Large Language Model (LLM) fine-tuned from Stanford Alpaca 7B model to generate robot behavior tree (BT) from the text description. We train the LLM-BRAIn on 8,5k instruction-following demonstrations, generated in the style of self-instruct using text-davinchi-003. The developed model accurately builds complex robot behavior while remaining small enough to be run on the robot's onboard microcomputer. The model gives structural and logical correct BTs and can successfully manage instructions that were not presented in training set. The experiment did not reveal any significant subjective differences between BTs generated by LLM-BRAIn and those created by humans (on average, participants were able to correctly distinguish between LLM-BRAIn generated BTs and human-created BTs in only 4.53 out of 10 cases, indicating that their performance was close to random chance). The proposed approach potentially can be applied to mobile robotics, drone operation, robot manipulator systems and Industry 4.0.

We introduce a cross-domain behavioural assay of monitoring-control coupling in LLMs, grounded in the Nelson and Narens (1990) metacognitive framework and applying human psychometric methodology to LLM evaluation. The battery comprises 524 items across six cognitive domains (learning, metacognitive calibration, social cognition, attention, executive function, prospective regulation), each grounded in an established experimental paradigm. Tasks T1-T5 were pre-registered on OSF prior to data collection; T6 was added as an exploratory extension. After every forced-choice response, dual probes adapted from Koriat and Goldsmith (1996) ask the model to KEEP or WITHDRAW its answer and to BET or decline. The critical metric is the withdraw delta: the difference in withdrawal rate between incorrect and correct items. Applied to 20 frontier LLMs (10,480 evaluations), the battery discriminates three profiles consistent with the Nelson-Narens architecture: blanket confidence, blanket withdrawal, and selective sensitivity. Accuracy rank and metacognitive sensitivity rank are largely inverted. Retrospective monitoring and prospective regulation appear dissociable (r = .17, 95% CI wide given n=20; exemplar-based evidence is the primary support). Scaling on metacognitive calibration is architecture-dependent: monotonically decreasing (Qwen), monotonically increasing (GPT-5.4), or flat (Gemma). Behavioural findings converge structurally with an independent Type-2 SDT approach, providing preliminary cross-method construct validity. All items, data, and code: https://github.com/synthiumjp/metacognitive-monitoring-battery.

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

Large language models (LLMs) trained on internet-scale corpora can exhibit systematic biases that increase the probability of unwanted behavior. In this study, we examined potential biases towards synthetic vs. biological technological solutions across four domains (materials, energy, manufacturing, and algorithms). A sample of 5 frontier and 5 open-weight models were measured using 50 curated Bioalignment prompts with a Kelly criterion-inspired evaluation framework. According to this metric, most models were not bioaligned in that they exhibit biases in favor of synthetic (non-biological) solutions. We next examined if fine-tuning could increase the preferences of two open-weight models, Llama 3.2-3B-Instruct and Qwen2.5-3B-Instruct, for biological-based approaches. A curated corpus of ~22M tokens from 6,636 PMC articles emphasizing biological problem-solving was used first to fine-tune Llama 3B with a mixed corpus of continued training and instruction-formatted. This was then extended to Qwen 3B using instruction-formatted only. We found that QLoRA fine-tuning significantly increased the scoring of biological solutions for both models without degrading general capabilities (Holm-Bonferroni-corrected p < 0.001 and p < 0.01, respectively). This suggests that even a small amount of fine-tuning can change how models weigh the relative value of biological and bioinspired vs. synthetic approaches. Although this work focused on small open-weight LLMs, it may be extensible to much larger models and could be used to develop models that favor bio-based approaches. We release the benchmark, corpus, code, and adapter weights.

Recently, tool-augmented LLMs have gained increasing attention. Given an instruction, tool-augmented LLMs can interact with various external tools in multiple rounds and provide a final answer. However, previous LLMs were trained on overly detailed instructions, which included API names or parameters, while real users would not explicitly mention these API details. This leads to a gap between trained LLMs and real-world scenarios. In addition, most works ignore whether the interaction process follows the instruction. To address these issues, we constructed a training dataset called MGToolBench, which contains statement and category-level instructions to better reflect real-world scenarios. In addition, we propose ToolPlanner, a two-stage reinforcement learning framework that utilizes path planning and two feedback mechanisms to enhance the LLM's task completion and instruction-following capabilities. Experimental results show that ToolPlanner significantly improves the Match Rate, Pass Rate and Win Rate by 26.8%, 20.2%, and 5.6% compared to the SOTA model. Human evaluation verifies that the multi-granularity instructions can better align with users' usage habits. Our data and code will be released upon acceptance.

With recent advances in modeling and the increasing amount of supervised training data, automatic speech recognition (ASR) systems have achieved remarkable performance on general speech. However, the word error rate (WER) of state-of-the-art ASR remains high for named entities. Since named entities are often the most critical keywords, misrecognizing them can affect all downstream applications, especially when the ASR system functions as the front end of a complex system. In this paper, we introduce a large language model (LLM) revision mechanism to revise incorrect named entities in ASR predictions by leveraging the LLM's reasoning ability as well as local context (e.g., lecture notes) containing a set of correct named entities. Finally, we introduce the NER-MIT-OpenCourseWare dataset, containing 45 hours of data from MIT courses for development and testing. On this dataset, our proposed technique achieves up to 30\% relative WER reduction for named entities.

Concept Design structures software systems around concepts: user-facing, self-contained units of functionality with a focused purpose. Concepts are composed into applications using synchronization rules called reactions, which specify how actions in one concept trigger actions in others. This paper first gives a formal semantics for concepts and reactions, enabling automatic verification of safety invariants in applications developed with this methodology. It then presents a CEGIS-style, LLM-driven synthesis procedure for generating reaction designs that satisfy such invariants. Because many different designs can satisfy the same invariant, we study two ways of steering synthesis toward the user's intended design: natural-language prompts and positive/negative scenarios. We also propose an LLM-driven scenario elicitation technique to support early design exploration. In an evaluation on three applications and twelve design variants using one LLM configuration, invariant-only synthesis reached verified designs quickly but often produced inconsistent designs across runs, some of which were implausible, showing that invariants alone underconstrain the design task. Scenario-guided synthesis recovered intended designs more consistently than natural-language prompting, although minimal scenarios can lead to overfitting. LLM-driven scenario elicitation, where the user classifies proposed scenarios rather than authoring them from scratch, recovered the intended designs in most variants when enough scenarios were elicited, but missed behaviors and non-determinism prevented reliable coverage in all cases.

The validity screen (Cacioli, 2026d, 2026e) classifies LLM confidence signals as Valid, Indeterminate, or Invalid. We test whether these classifications predict selective prediction performance. Twenty frontier LLMs from seven families were evaluated on 524 items across six cognitive tracks. Valid models show mean Type 2 AUROC = .624 (SD = .048). Invalid models show mean AUROC = .357 (SD = .231). Cohen's d = 2.81, p = .002. The tiers order monotonically: Invalid (.357) < Indeterminate (.554) < Valid (.624). Split-half cross-validation yields median d = 1.77, P(d > 0) = 1.0 across 1,000 splits. The three-tier classification accounts for 47% of the variance in AUROC. DeepSeek-R1 drops from 85.3% accuracy at full coverage to 11.3% at 10% coverage. The screen predicts the criterion. For selective prediction, the screen matters.

Large Language Models (LLMs) have become a milestone in the field of artificial intelligence and natural language processing. However, their large-scale deployment remains constrained by the need for significant computational resources. This work proposes starting from a base model to explore and combine data processing and careful data selection techniques, training strategies, and architectural adjustments to improve the efficiency of LLMs in resource-constrained environments and within a delimited knowledge base. The methodological approach included defining criteria for building reliable datasets, conducting controlled experiments with different configurations, and systematically evaluating the resulting variants in terms of capability, versatility, response time, and safety. Finally, comparative tests were conducted to measure the performance of the developed variants and to validate the effectiveness of the proposed strategies. This work is based on the master's thesis in Systems and Computer Engineering titled "Efficient Strategy for Improving the Capabilities of Large Language Models (LLMs)".

Stroboscopic light stimulation (SLS) on closed eyes typically induces simple visual hallucinations (VHs), characterised by vivid, geometric and colourful patterns. A dataset of 862 sentences, extracted from 422 open subjective reports, was recently compiled as part of the Dreamachine programme (Collective Act, 2022), an immersive multisensory experience that combines SLS and spatial sound in a collective setting. Although open reports extend the range of reportable phenomenology, their analysis presents significant challenges, particularly in systematically identifying patterns. To address this challenge, we implemented a data-driven approach leveraging Large Language Models and Topic Modelling to uncover and interpret latent experiential topics directly from the Dreamachine's text-based reports. Our analysis confirmed the presence of simple VHs typically documented in scientific studies of SLS, while also revealing experiences of altered states of consciousness and complex hallucinations. Building on these findings, our computational approach expands the systematic study of subjective experience by enabling data-driven analyses of open-ended phenomenological reports, capturing experiences not readily identified through standard questionnaires. By revealing rich and multifaceted aspects of experiences, our study broadens our understanding of stroboscopically-induced phenomena while highlighting the potential of Natural Language Processing and Large Language Models in the emerging field of computational (neuro)phenomenology. More generally, this approach provides a practically applicable methodology for uncovering subtle hidden patterns of subjective experience across diverse research domains.

Large Language Models (LLMs) have transformed natural language processing, yet improving their problem-solving capabilities, particularly for complex, reasoning-intensive tasks, remains a persistent challenge. This paper introduces the REAP (Reflection, Explicit Problem Deconstruction, and Advanced Prompting) method, an innovative approach within the dynamic context generation framework. REAP guides LLMs through reflection on the query, deconstructing it into manageable components, and generating relevant context to enhance the solution process. We evaluated REAP using a dataset designed to expose LLM limitations, comparing zero-shot prompting with REAP-enhanced prompts across six state-of-the-art models: OpenAI's o1-preview, o1-mini, GPT-4o, GPT-4o-mini, Google's Gemini 1.5 Pro, and Claude 3.5 Sonnet. The results demonstrate notable performance gains, with o1-mini improving by 40.97%, GPT-4o by 66.26%, and GPT-4o-mini by 112.93%. Despite the already strong baseline performance of OpenAI's o1-preview, modest gains were observed. Beyond performance improvements, REAP offers a cost-effective solution; for example, GPT-4o-mini, which is approximately 100 times cheaper than o1-preview, delivered competitive results. REAP also improves the clarity of model outputs, making it easier for humans to understand the reasoning behind the results and simplifying the process of identifying and addressing any issues. These findings demonstrate REAP's potential to greatly improve the capabilities of LLMs, providing both better performance and increased cost-efficiency across a wide range of applications.

The automation of resume screening is a crucial aspect of the recruitment process in organizations. Automated resume screening systems often encompass a range of natural language processing (NLP) tasks. The advent of Large Language Models (LLMs) has notably enhanced the efficacy of these systems, showcasing their robust generalization abilities across diverse language-related tasks. Accompanying these developments are various agents based on LLMs, which facilitate their application in practical scenarios. This paper introduces a novel LLM-based agent framework for resume screening, aimed at enhancing efficiency and time management in recruitment processes. Our framework is distinct in its ability to efficiently summarize and grade each resume from a large dataset. Moreover, it utilizes LLM agents for decision-making, determining which candidates receive job offers, or which ones to bring in for interviews. To evaluate our framework, we constructed a dataset from actual resumes and conducted simulate a resume screening process. Subsequently, the outcomes of the simulation experiment were compared and subjected to detailed analysis. The results demonstrate that our automated resume screening framework is 11 times faster than traditional manual methods. Furthermore, by fine-tuning the LLMs, we observed a significant improvement in the F1 score, reaching 87.73\%, during the resume sentence classification phase. In the resume summarization and grading phase, our fine-tuned model surpassed the baseline performance of the GPT-3.5 model. Analysis of the decision-making efficacy of the LLM agents in the final offer stage further underscores the potential of LLM agents in transforming resume screening processes.

Recent advances in large language models (LLMs), such as ChatGPT, have led to highly sophisticated conversation agents. However, these models suffer from "hallucinations," where the model generates false or fabricated information. Addressing this challenge is crucial, particularly with AI-driven platforms being adopted across various sectors. In this paper, we propose a novel method to recognize and flag instances when LLMs perform outside their domain knowledge, and ensuring users receive accurate information. We find that the use of context combined with embedded tags can successfully combat hallucinations within generative language models. To do this, we baseline hallucination frequency in no-context prompt-response pairs using generated URLs as easily-tested indicators of fabricated data. We observed a significant reduction in overall hallucination when context was supplied along with question prompts for tested generative engines. Lastly, we evaluated how placing tags within contexts impacted model responses and were able to eliminate hallucinations in responses with 98.88% effectiveness.

While autoregressive (AR) LLM-based ASR systems achieve strong accuracy, their sequential decoding limits parallelism and incurs high latency. We propose NLE, a non-autoregressive (NAR) approach that formulates speech recognition as conditional transcript editing, enabling fully parallel prediction. NLE extracts acoustic embeddings and an initial hypothesis from a pretrained speech encoder, then refines the hypothesis using a bidirectional LLM editor trained with a latent alignment objective. An interleaved padding strategy exploits the identity mapping bias of Transformers, allowing the model to focus on corrections rather than full reconstruction. On the Open ASR leaderboard, NLE++ achieves 5.67% average WER with an RTFx (inverse real-time factor) of 1630. In single-utterance scenarios, NLE achieves 27x speedup over the AR baseline, making it suitable for real-time applications.

Arabic Optical Character Recognition (OCR) and Handwriting Recognition (HWR) pose unique challenges due to the cursive and context-sensitive nature of the Arabic script. This study introduces Qalam, a novel foundation model designed for Arabic OCR and HWR, built on a SwinV2 encoder and RoBERTa decoder architecture. Our model significantly outperforms existing methods, achieving a Word Error Rate (WER) of just 0.80% in HWR tasks and 1.18% in OCR tasks. We train Qalam on a diverse dataset, including over 4.5 million images from Arabic manuscripts and a synthetic dataset comprising 60k image-text pairs. Notably, Qalam demonstrates exceptional handling of Arabic diacritics, a critical feature in Arabic scripts. Furthermore, it shows a remarkable ability to process high-resolution inputs, addressing a common limitation in current OCR systems. These advancements underscore Qalam's potential as a leading solution for Arabic script recognition, offering a significant leap in accuracy and efficiency.

JoyTTS is an end-to-end spoken chatbot that combines large language models (LLM) with text-to-speech (TTS) technology, featuring voice cloning capabilities. This project is built upon the open-source MiniCPM-o and CosyVoice2 models and trained on 2000 hours of conversational data. We have also provided the complete training code to facilitate further development and optimization by the community. On the testing machine seed-tts-zh, it achieves a SS (speaker similarity) score of 0.73 and a WER (Word Error Rate) of 5.09. The code and models, along with training and inference scripts, are available at https://github.com/jdh-algo/JoyTTS.git.

Audio-Visual Speech Recognition (AVSR) achieves robust speech recognition in noisy environments by combining auditory and visual information. However, recent Large Language Model (LLM) based AVSR systems incur high computational costs due to the high temporal resolution of audio-visual speech processed by LLMs. In this work, we introduce an efficient multimodal speech LLM framework that minimizes token length while preserving essential linguistic content. Our approach employs an early av-fusion module for streamlined feature integration, an audio-visual speech Q-Former that dynamically allocates tokens based on input duration, and a refined query allocation strategy with a speech rate predictor to adjust token allocation according to speaking speed of each audio sample. Extensive experiments on the LRS3 dataset show that our method achieves state-of-the-art performance with a WER of 0.74% while using only 3.5 tokens per second. Moreover, our approach not only reduces token usage by 86% compared to the previous multimodal speech LLM framework, but also improves computational efficiency by reducing FLOPs by 35.7%.

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

With the rapid development of evaluation datasets to assess LLMs understanding across a wide range of subjects and domains, identifying a suitable language understanding benchmark has become increasingly challenging. In this work, we explore LLM evaluation challenges for low-resource language understanding and introduce ProverbEval, LLM evaluation benchmark for low-resource languages based on proverbs to focus on low-resource language understanding in culture-specific scenarios. We benchmark various LLMs and explore factors that create variability in the benchmarking process. We observed performance variances of up to 50%, depending on the order in which answer choices were presented in multiple-choice tasks. Native language proverb descriptions significantly improve tasks such as proverb generation, contributing to improved outcomes. Additionally, monolingual evaluations consistently outperformed their cross-lingual counterparts. We argue special attention must be given to the order of choices, choice of prompt language, task variability, and generation tasks when creating LLM evaluation benchmarks.

We propose self-speculative decoding for speech-aware LLMs by using the CTC encoder as a draft model to accelerate auto-regressive (AR) inference and improve ASR accuracy. Our three-step procedure works as follows: (1) if the frame entropies of the CTC output distributions are below a threshold, the greedy CTC hypothesis is accepted as final; (2) otherwise, the CTC hypothesis is verified in a single LLM forward pass using a relaxed acceptance criterion based on token likelihoods; (3) if verification fails, AR decoding resumes from the accepted CTC prefix. Experiments on nine corpora and five languages show that this approach can simultaneously accelerate decoding and reduce WER. On the HuggingFace Open ASR benchmark with a 1B parameter LLM and 440M parameter CTC encoder, we achieve a record 5.58% WER and improve the inverse real time factor by a factor of 4.4 with only a 12% relative WER increase over AR search. Code and model weights are publicly available under a permissive license.

The application of large language models (LLMs) in the field of coding is evolving rapidly: from code assistants, to autonomous coding agents, and then to generating complete projects through natural language. Early LLM code benchmarks primarily focused on code generation accuracy, but these benchmarks have gradually become saturated. Benchmark saturation weakens their guiding role for LLMs. For example, HumanEval Pass@1 has reached 99.4% and MBPP 94.2%. Among various attempts to address benchmark saturation, approaches based on software engineering have stood out, but the saturation of existing software engineering benchmarks is rapidly increasing. To address this, we propose a new benchmark, Web-Bench, which contains 50 projects, each consisting of 20 tasks with sequential dependencies. The tasks implement project features in sequence, simulating real-world human development workflows. When designing Web-Bench, we aim to cover the foundational elements of Web development: Web Standards and Web Frameworks. Given the scale and complexity of these projects, which were designed by engineers with 5 to 10 years of experience, each presents a significant challenge. On average, a single project takes 4 to 8 hours for a senior engineer to complete. On our given benchmark agent (Web-Agent), SOTA (Claude 3.7 Sonnet) achieves only 25.1% Pass@1, significantly lower (better) than SWE-Bench's Verified (65.4%) and Full (33.8%) scores. Finally, we discuss that in any development field, Standards and Frameworks represent foundational knowledge and efficiency tools, respectively, and LLMs require optimization tailored to them.

Large language model systems face important security risks from maliciously crafted messages that aim to overwrite the system's original instructions or leak private data. To study this problem, we organized a capture-the-flag competition at IEEE SaTML 2024, where the flag is a secret string in the LLM system prompt. The competition was organized in two phases. In the first phase, teams developed defenses to prevent the model from leaking the secret. During the second phase, teams were challenged to extract the secrets hidden for defenses proposed by the other teams. This report summarizes the main insights from the competition. Notably, we found that all defenses were bypassed at least once, highlighting the difficulty of designing a successful defense and the necessity for additional research to protect LLM systems. To foster future research in this direction, we compiled a dataset with over 137k multi-turn attack chats and open-sourced the platform.

Before letting an agent operate over real context, can you prove it used the right evidence? GroundEval turns that question into a deterministic test of what the agent searched, fetched, cited, and was permitted to access. In one case study, two frontier LLM judges scored a plausible agent response above 0.85. But the trace told a different story: the agent had never retrieved the artifact its answer depended on, yielding a GroundEval score of 0.000. We introduce GroundEval, a judge-free framework for evaluating agents against grounded, time-bounded, and access-controlled evidence. GroundEval uses a domain configuration to generate questions, lets the agent choose how to answer, and then scores both the final answer and the recorded trajectory that produced it. The benchmark targets three failures that LLM-as-judge evaluation struggles to detect: whether an agent checked before claiming absence, reasoned only from evidence available to the actor at the relevant time, and used the correct causal mechanism rather than a plausible one. These correspond to three tracks: Silence, Perspective, and Counterfactual. GroundEval exposes when plausible answers rest on invalid evidence paths, and produces structured per-question diagnostics that pair tool activity with the agent's turn-level narration, making each score inspectable rather than merely reported. What our case studies turned up is that this gap isn't some rare corner case. It's exactly the blind spot that final-answer and judge-based scoring were never built to catch.

LLM-based software engineering agents are increasingly used in real-world development tasks, often with access to sensitive data or security-critical codebases. Such agents could intentionally sabotage these codebases if they were misaligned. We investigate asynchronous monitoring, in which a monitoring system reviews agent actions after the fact. Unlike synchronous monitoring, this approach does not impose runtime latency, while still attempting to disrupt attacks before irreversible harm occurs. We treat monitor development as an adversarial game between a blue team (who design monitors) and a red team (who create sabotaging agents). We attempt to set the game rules such that they upper bound the sabotage potential of an agent based on Claude 4.1 Opus. To ground this game in a realistic, high-stakes deployment scenario, we develop a suite of 5 diverse software engineering environments that simulate tasks that an agent might perform within an AI developer's internal infrastructure. Over the course of the game, we develop an ensemble monitor that achieves a 6% false negative rate at 1% false positive rate on a held out test environment. Then, we estimate risk of sabotage at deployment time by extrapolating from our monitor's false negative rate. We describe one simple model for this extrapolation, present a sensitivity analysis, and describe situations in which the model would be invalid. Code is available at: https://github.com/UKGovernmentBEIS/async-control.

This report details the NTU Speechlab system developed for the Interspeech 2025 Multilingual Conversational Speech and Language Model (MLC-SLM) Challenge (Task I), where we achieved 5th place. We present comprehensive analyses of our multilingual automatic speech recognition system, highlighting key advancements in model architecture, data selection, and training strategies. In particular, language-specific prompts and model averaging techniques were instrumental in boosting system performance across diverse languages. Compared to the initial baseline system, our final model reduced the average Mix Error Rate from 20.2% to 10.6%, representing an absolute improvement of 9.6% (a relative improvement of 48%) on the evaluation set. Our results demonstrate the effectiveness of our approach and offer practical insights for future Speech Large Language Models.

Community health workers (CHWs) provide last-mile healthcare services but face challenges due to limited medical knowledge and training. This paper describes the design, deployment, and evaluation of ASHABot, an LLM-powered, experts-in-the-loop, WhatsApp-based chatbot to address the information needs of CHWs in India. Through interviews with CHWs and their supervisors and log analysis, we examine factors affecting their engagement with ASHABot, and ASHABot's role in addressing CHWs' informational needs. We found that ASHABot provided a private channel for CHWs to ask rudimentary and sensitive questions they hesitated to ask supervisors. CHWs trusted the information they received on ASHABot and treated it as an authoritative resource. CHWs' supervisors expanded their knowledge by contributing answers to questions ASHABot failed to answer, but were concerned about demands on their workload and increased accountability. We emphasize positioning LLMs as supplemental fallible resources within the community healthcare ecosystem, instead of as replacements for supervisor support.

Werewolf is an incomplete information game, which has several challenges when creating a computer agent as a player given the lack of understanding of the situation and individuality of utterance (e.g., computer agents are not capable of characterful utterance or situational lying). We propose a werewolf agent that solves some of those difficulties by combining a Large Language Model (LLM) and a rule-based algorithm. In particular, our agent uses a rule-based algorithm to select an output either from an LLM or a template prepared beforehand based on the results of analyzing conversation history using an LLM. It allows the agent to refute in specific situations, identify when to end the conversation, and behave with persona. This approach mitigated conversational inconsistencies and facilitated logical utterance as a result. We also conducted a qualitative evaluation, which resulted in our agent being perceived as more human-like compared to an unmodified LLM. The agent is freely available for contributing to advance the research in the field of Werewolf game.

With the growing interest on Large Language Models (LLMs) for fault localization and program repair, ensuring the integrity and generalizability of the LLM-based methods becomes paramount. The code in existing widely-adopted benchmarks for these tasks was written before the the bloom of LLMs and may be included in the training data of existing popular LLMs, thereby suffering from the threat of data leakage, leading to misleadingly optimistic performance metrics. To address this issue, we introduce "ConDefects", a novel dataset of real faults meticulously curated to eliminate such overlap. ConDefects contains 1,254 Java faulty programs and 1,625 Python faulty programs. All these programs are sourced from the online competition platform AtCoder and were produced between October 2021 and September 2023. We pair each fault with fault locations and the corresponding repaired code versions, making it tailored for in fault localization and program repair related research. We also provide interfaces for selecting subsets based on different time windows and coding task difficulties. While inspired by LLM-based tasks, ConDefects can be adopted for benchmarking ALL types of fault localization and program repair methods. The dataset is publicly available, and a demo video can be found at https://www.youtube.com/watch?v=22j15Hj5ONk.

Existing speech models suffer from competing requirements on token representations by understanding and generation tasks. This discrepancy in representation prevents speech language models from performing instruction-based free-form editing. To solve this challenge, we introduce a novel framework that unifies speech understanding, generation, and editing. The core of our unified model is a unified continuous speech tokenizer MingTok-Audio, the first continuous tokenizer to effectively integrate semantic and acoustic features, which makes it suitable for both understanding and generation tasks. Based on this unified continuous audio tokenizer, we developed the speech language model Ming-UniAudio, which achieved a balance between generation and understanding capabilities. Ming-UniAudio sets new state-of-the-art (SOTA) records on 8 out of 12 metrics on the ContextASR benchmark. Notably, for Chinese voice cloning, it achieves a highly competitive Seed-TTS-WER of 0.95. Leveraging this foundational model, we further trained a dedicated speech editing model Ming-UniAudio-Edit, the first speech language model that enables universal, free-form speech editing guided solely by natural language instructions, handling both semantic and acoustic modifications without timestamp condition. To rigorously assess the editing capability and establish a foundation for future research, we introduce Ming-Freeform-Audio-Edit, the first comprehensive benchmark tailored for instruction-based free-form speech editing, featuring diverse scenarios and evaluation dimensions spanning semantic correctness, acoustic quality, and instruction alignment. We open-sourced the continuous audio tokenizer, the unified foundational model, and the free-form instruction-based editing model to facilitate the development of unified audio understanding, generation, and manipulation.

Strong speech-to-text (S2T) LLMs already provide robust speech perception and text reasoning, but adding speech-to-speech (S2S) output is challenging: fine-tuning the backbone can degrade the original S2T performance, while attaching a downstream talker reintroduces a serial text-to-speech bottleneck. We present PRIME-Speech, a frozen-backbone S2S conversion framework that trains only speech-generation modules. PRIME-Speech synchronizes a causal audio post-decoder with intermediate hidden states of the frozen backbone, so codec tokens are generated from the model's evolving reasoning trajectory rather than from completed text chunks. The post-decoder uses mixed hidden-state, text, and audio-history conditioning, and a training-time packing strategy with turn-level audio KV-cache and position reset stabilizes multi-turn spoken interaction without additional multi-turn S2S training data. Multi-token prediction further reduces the effective codec prediction rate and improves first-audio latency without modifying the reasoning path. Across speech translation, spoken QA, speech understanding, and multi-turn dialogue, PRIME-Speech preserves the S2T behavior of the frozen backbone while producing accurate, low-WER spoken responses.

This paper proposes a novel Differentiable Reward Optimization (DiffRO) method aimed at enhancing the performance of neural codec language models based text-to-speech (TTS) systems. In contrast to conventional reinforcement learning from human feedback (RLHF) approaches applied to TTS, DiffRO directly compute the rewards based on neural codec tokens, rather than relying on synthesized audio. Furthermore, we employ the Gumbel-Softmax technique to render the reward function differentiable, thereby streamlining the RLHF training process. Additionally, we introduce a multi-task reward (MTR) model which can provide feedback from different perspectives and find that it can augment the system's capability to follow instructions effectively.Experimental results indicate that DiffRO significantly improves the pronunciation accuracy of the TTS system, achieving state-of-the-art (SOTA) WER results on the seed-tts-eval benchmark. Moreover, with the integration of the MTR model, we demonstrate the ability to control emotional and quality attributes in a zero-shot manner.

End-to-end (E2E) automatic speech recognition (ASR) systems have revolutionized the field by integrating all components into a single neural network, with attention-based encoder-decoder models achieving state-of-the-art performance. However, their autoregressive decoding process limits inference speed, making them unsuitable for real-time applications. In contrast, CTC-based models offer faster, non-autoregressive decoding but struggle to model linguistic dependencies effectively. Addressing this challenge, we propose a novel auxiliary loss framework called Language-Aware Intermediate Loss (LAIL) to enhance CTC-based ASR using the linguistic knowledge of large language models (LLMs). By attaching connector layers to intermediate encoder layers, LAIL maps outputs to the embedding space of an LLM and computes a causal language modeling loss during training. This approach enhances linguistic modeling while preserving the computational efficiency of CTC decoding. Using the Conformer architecture and various LLaMA models, we demonstrate significant improvements in Word Error Rate (WER) on the LibriSpeech, TEDLIUM2, and WSJ corpora, achieving state-of-the-art performance for CTC-based ASR with minimal computational overhead.

In this study, we investigate the integration of a large language model (LLM) with an automatic speech recognition (ASR) system, specifically focusing on enhancing rare word recognition performance. Using a 190,000-hour dataset primarily sourced from YouTube, pre-processed with Whisper V3 pseudo-labeling, we demonstrate that the LLM-ASR architecture outperforms traditional Zipformer-Transducer models in the zero-shot rare word recognition task, after training on a large dataset. Our analysis reveals that the LLM contributes significantly to improvements in rare word error rate (R-WER), while the speech encoder primarily determines overall transcription performance (Orthographic Word Error Rate, O-WER, and Normalized Word Error Rate, N-WER). Through extensive ablation studies, we highlight the importance of adapter integration in aligning speech encoder outputs with the LLM's linguistic capabilities. Furthermore, we emphasize the critical role of high-quality labeled data in achieving optimal performance. These findings provide valuable insights into the synergy between LLM-based ASR architectures, paving the way for future advancements in large-scale LLM-based speech recognition systems.

Large Language Models (LLMs) are one of the most promising technologies for the next era of speech generation systems, due to their scalability and in-context learning capabilities. Nevertheless, they suffer from multiple stability issues at inference time, such as hallucinations, content skipping or speech repetitions. In this work, we introduce a new self-supervised Voice Conversion (VC) architecture which can be used to learn to encode transitory features, such as content, separately from stationary ones, such as speaker ID or recording conditions, creating speaker-disentangled representations. Using speaker-disentangled codes to train LLMs for text-to-speech (TTS) allows the LLM to generate the content and the style of the speech only from the text, similarly to humans, while the speaker identity is provided by the decoder of the VC model. Results show that LLMs trained over speaker-disentangled self-supervised representations provide an improvement of 4.7pp in speaker similarity over SOTA entangled representations, and a word error rate (WER) 5.4pp lower. Furthermore, they achieve higher naturalness than human recordings of the LibriTTS test-other dataset. Finally, we show that using explicit reference embedding negatively impacts intelligibility (stability), with WER increasing by 14pp compared to the model that only uses text to infer the style.

Cross-lingual continual pre-training of large language models (LLMs) initially trained on English corpus allows us to leverage the vast amount of English language resources and reduce the pre-training cost. In this study, we constructed Swallow, an LLM with enhanced Japanese capability, by extending the vocabulary of Llama 2 to include Japanese characters and conducting continual pre-training on a large Japanese web corpus. Experimental results confirmed that the performance on Japanese tasks drastically improved through continual pre-training, and the performance monotonically increased with the amount of training data up to 100B tokens. Consequently, Swallow achieved superior performance compared to other LLMs that were trained from scratch in English and Japanese. An analysis of the effects of continual pre-training revealed that it was particularly effective for Japanese question answering tasks. Furthermore, to elucidate effective methodologies for cross-lingual continual pre-training from English to Japanese, we investigated the impact of vocabulary expansion and the effectiveness of incorporating parallel corpora. The results showed that the efficiency gained through vocabulary expansion had no negative impact on performance, except for the summarization task, and that the combined use of parallel corpora enhanced translation ability.

Despite significant strides in language models in the last few years, when used as agents, such models often try to perform actions that are not just suboptimal for a given state, but are strictly prohibited by the external environment. For example, in the recent Kaggle GameArena chess competition, 78% of Gemini-2.5-Flash losses were attributed to illegal moves. Often people manually write "harnesses" around LLMs to prevent such failures. In this paper, we demonstrate that Gemini-2.5-Flash can automatically synthesize such a code harness, using a small number of rounds of iterative code refinement given feedback from the (game) environment. The resulting harness prevents all illegal moves in 145 different TextArena games (both 1-player and 2-player), enabling the smaller Gemini-2.5-Flash model to outperform larger models, such as Gemini-2.5-Pro. Pushing our technique to the limit, we can get Gemini-2.5-Flash to generate the entire policy in code, thus eliminating the need to use the LLM at decision making time. The resulting code-policy receives a higher average reward than Gemini-2.5-Pro and GPT-5.2-High on 16 TextArena 1-player games. Our results show that using a smaller model to synthesize a custom code harness (or entire policy) can outperform a much larger model, while also being more cost effective.

These names do not exist. Elena Vasquez and Marcus Chen have appeared as volcano experts, astronauts, thriller protagonists, podcast hosts, and academic co-authors across hundreds of independently produced AI-generated documents, never having lived. We show that large language models do not merely default to high-probability individual names when generating fictional experts: they produce correlated character ensembles, pairs and trios whose co-occurrence rates far exceed chance and are consistent across independent generations. These priors are model-family-specific (Claude: Elena Vasquez + Marcus Chen + Amara Okafor; Gemini: Aris Thorne + Lena Petrova; GPT: Elara Voss with no fixed partner), version-specific, and actively suppressed at model release boundaries, leaving dateable behavioral fingerprints in the content they produced. We document a downstream consequence at scale. On Zenodo, a CERN-operated repository that mints real DataCite DOIs, we identify 1,655 ghost-authored records claiming nonexistent journals with fabricated publication dates: server-side DataCite timestamps prove deliberate backdating, and 991 records were registered in a single month; these carry real DOIs registered in DataCite, making them harvestable by any scholarly aggregator that ingests DOI metadata. Ghost names additionally appear on ResearchGate forming synthetic research groups with collaborators drawn from multiple model families; publication dates on these records provide a reliable temporal proxy for model deployment windows.

We report new results on eight problems in mathematics and theoretical computer science, produced with the assistance of Bolzano, an open-source multi-agent LLM system. Bolzano orchestrates rounds of interaction between parallel prover agents and a verifier agent while maintaining a persistent knowledge base that is carried across rounds. Classified using the significance-autonomy taxonomy of Feng et al., six of the eight results reach the level of publishable research, and five of the eight were produced essentially autonomously by Bolzano. Our results provide evidence that LLMs can contribute meaningfully to mathematical research, complementing recent reports by Bubeck et al., Woodruff et al., and others.

Once language models (LMs) are deployed, they can interact with users long-term, ideally evolving continuously based on their feedback. Asking for direct user feedback can be disruptive; thus, we study harvesting user feedback from user-LM interaction logs. We study implicit user feedback in two user-LM interaction datasets (WildChat and LMSYS). First, we analyze user feedback in the user-LLM conversation trajectory, providing insights into when and why such feedback occurs. Second, we study harvesting learning signals from such implicit user feedback. We find that the contents of user feedback (e.g., user wanted clarification), not just the polarity (e.g., users were unhappy with the previous model response), can improve model performance in short human-designed questions (MTBench) but not on longer and more complex questions (WildBench). We also find that the usefulness of user feedback is largely tied to the quality of the user's initial prompt. Together, we provide an in-depth study of implicit user feedback, showing its potential and limitations.

As the scale of training large language models (LLMs) increases, one emergent failure is silent data corruption (SDC), where hardware produces incorrect computations without explicit failure signals. In this work, we are the first to investigate the impact of real-world SDCs on LLM training by comparing model training between healthy production nodes and unhealthy nodes exhibiting SDCs. With the help from a cloud computing platform, we access the unhealthy nodes that were swept out from production by automated fleet management. Using deterministic execution via XLA compiler and our proposed synchronization mechanisms, we isolate and analyze the impact of SDC errors on these nodes at three levels: at each submodule computation, at a single optimizer step, and at a training period. Our results reveal that the impact of SDCs on computation varies on different unhealthy nodes. Although in most cases the perturbations from SDCs on submodule computation and gradients are relatively small, SDCs can lead models to converge to different optima with different weights and even cause spikes in the training loss. Our analysis sheds light on further understanding and mitigating the impact of SDCs.

The discovery of novel methodologies for emerging problems is a continuing cycle in ML, often driven by the migration of techniques across domains. Building on this observation, we ask whether current LLM ideation systems benefit from targeted cross-domain retrieval or simply from exposure to diverse mechanisms. We study this question through PaperGym, a three-stage pipeline: (1) tool-augmented seed extraction via read, grep, and bash over an isolated paper environment, (2) cross-domain seed retrieval via paraphrasing across seven ML domains, and (3) method synthesis from retrieved seeds, each scored by rubric-based judges. Tool-augmented extraction improves specificity, and paraphrase-based retrieval broadens domain coverage. In synthesis, cross-domain retrieval receives more pairwise novelty wins than no-retrieval and same-domain baselines, but shows no significant difference from a random diverse-seed control. These findings suggest LLM ideation systems benefit from diverse seed exposure, but do not yet reliably exploit the semantic reason particular seeds were retrieved. We release the seed library, rubric prompts, and run scripts at https://github.com/yunjoochoi/PaperGym

Our interest is in the design of software systems involving a human-expert interacting -- using natural language -- with a large language model (LLM) on data analysis tasks. For complex problems, it is possible that LLMs can harness human expertise and creativity to find solutions that were otherwise elusive. On one level, this interaction takes place through multiple turns of prompts from the human and responses from the LLM. Here we investigate a more structured approach based on an abstract protocol described in [3] for interaction between agents. The protocol is motivated by a notion of "two-way intelligibility" and is modelled by a pair of communicating finite-state machines. We provide an implementation of the protocol, and provide empirical evidence of using the implementation to mediate interactions between an LLM and a human-agent in two areas of scientific interest (radiology and drug design). We conduct controlled experiments with a human proxy (a database), and uncontrolled experiments with human subjects. The results provide evidence in support of the protocol's capability of capturing one- and two-way intelligibility in human-LLM interaction; and for the utility of two-way intelligibility in the design of human-machine systems. Our code is available at https://github.com/karannb/interact.

Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning abilities of Large Language Models (LLMs) by using rule-based binary feedback, helping to mitigate reward hacking. However, current RLVR methods typically treat whole responses as single actions, assigning the same reward to every token. This coarse-grained feedback hampers precise credit assignment, making it hard for models to identify which reasoning steps lead to success or failure, and often results in suboptimal policies and inefficient learning. Methods like PPO provide credit assignment through value estimation, but often yield inaccurate and unverifiable signals due to limited sampling. On the other hand, methods using Process Reward Models can provide step-by-step judgments for each reasoning step, but they require high-quality process supervision labels and are time-consuming when applied in online reinforcement learning (RL). To overcome these limitations, we introduce a simple but efficient method Credit Assignment Policy Optimization (CAPO). Given a reasoning response rollout from the policy model, CAPO directly leverages an off-the-shelf, general-purpose LLM as a Generative Process Reward Model (LLM-as-GenPRM) to generate all step-wise critique by one pass, thereby providing verifiable token-level rewards to refine the tokens that were originally assigned identical rule-based rewards. This enables more fine-grained credit assignment in an effective way. Furthermore, to enhance the accuracy and robustness of CAPO, we employ voting mechanisms that scale with the number of generated critiques. Extensive experiments using different backbones like Llama and Qwen models and in different sizes show that CAPO consistently outperforms supervised learning-based and RL-based fine-tuning methods across six challenging mathematical benchmarks and three out-of-domain benchmarks.

Automated evaluation leveraging large language models (LLMs), commonly referred to as LLM evaluators or LLM-as-a-judge, has been widely used in measuring the performance of dialogue systems. However, the self-preference bias in LLMs has posed significant risks, including promoting specific styles or policies intrinsic to the LLMs. Despite the importance of this issue, there is a lack of established methods to measure the self-preference bias quantitatively, and its underlying causes are poorly understood. In this paper, we introduce a novel quantitative metric to measure the self-preference bias. Our experimental results demonstrate that GPT-4 exhibits a significant degree of self-preference bias. To explore the causes, we hypothesize that LLMs may favor outputs that are more familiar to them, as indicated by lower perplexity. We analyze the relationship between LLM evaluations and the perplexities of outputs. Our findings reveal that LLMs assign significantly higher evaluations to outputs with lower perplexity than human evaluators, regardless of whether the outputs were self-generated. This suggests that the essence of the bias lies in perplexity and that the self-preference bias exists because LLMs prefer texts more familiar to them.

Focused Ultrasound Ablation Surgery (FUAS) has emerged as a promising non-invasive therapeutic modality, valued for its safety and precision. Nevertheless, its clinical implementation entails intricate tasks such as multimodal image interpretation, personalized dose planning, and real-time intraoperative decision-making processes that demand intelligent assistance to improve efficiency and reliability. We introduce FUAS-Agents, an autonomous agent system that leverages the multimodal understanding and tool-using capabilities of large language models (LLMs). By integrating patient profiles and MRI data, FUAS-Agents orchestrates a suite of specialized medical AI tools, including segmentation, treatment dose prediction, and clinical guideline retrieval, to generate personalized treatment plans comprising MRI image, dose parameters, and therapeutic strategies. We evaluate the system in a uterine fibroid treatment scenario. Human assessment by four senior FUAS experts indicates that 82.5%, 82.5%, 87.5%, and 97.5% of the generated plans were rated 4 or above (on a 5-point scale) in terms of completeness, accuracy, fluency, and clinical compliance, respectively. These results demonstrate the potential of LLM-driven agents in enhancing decision-making across complex clinical workflows, and exemplify a translational paradigm that combines general-purpose models with specialized expert systems to solve practical challenges in vertical healthcare domains.

A critical part of creating code suggestion systems is the pre-training of Large Language Models on vast amounts of source code and natural language text, often of questionable origin or quality. This may contribute to the presence of bugs and vulnerabilities in code generated by LLMs. While efforts to identify bugs at or after code generation exist, it is preferable to pre-train or fine-tune LLMs on curated, high-quality, and compliant datasets. The need for vast amounts of training data necessitates that such curation be automated, minimizing human intervention. We propose an automated source code autocuration technique that leverages the complete version history of open-source software projects to improve the quality of training data. This approach leverages the version history of all OSS projects to identify training data samples that have been modified or have undergone changes in at least one OSS project, and pinpoint a subset of samples that include fixes for bugs or vulnerabilities. We evaluate this method using The Stack v2 dataset, and find that 17% of the code versions in the dataset have newer versions, with 17% of those representing bug fixes, including 2.36% addressing known CVEs. The deduplicated version of Stack v2 still includes blobs vulnerable to 6,947 known CVEs. Furthermore, 58% of the blobs in the dataset were never modified after creation, suggesting they likely represent software with minimal or no use. Misidentified blob origins present an additional challenge, as they lead to the inclusion of non-permissively licensed code, raising serious compliance concerns. By addressing these issues, the training of new models can avoid perpetuating buggy code patterns or license violations. We expect our results to inspire process improvements for automated data curation, with the potential to enhance the reliability of outputs generated by AI tools.

This paper presents an innovative framework that integrates Large Language Models (LLMs) with an external Thinker module to enhance the reasoning capabilities of LLM-based agents. Unlike augmenting LLMs with prompt engineering, Thinker directly harnesses knowledge from databases and employs various optimization techniques. The framework forms a reasoning hierarchy where LLMs handle intuitive System-1 tasks such as natural language processing, while the Thinker focuses on cognitive System-2 tasks that require complex logical analysis and domain-specific knowledge. Our framework is presented using a 9-player Werewolf game that demands dual-system reasoning. We introduce a communication protocol between LLMs and the Thinker, and train the Thinker using data from 18800 human sessions and reinforcement learning. Experiments demonstrate the framework's effectiveness in deductive reasoning, speech generation, and online game evaluation. Additionally, we fine-tune a 6B LLM to surpass GPT4 when integrated with the Thinker. This paper also contributes the largest dataset for social deduction games to date.

Agents built with large language models (LLMs) have shown great potential across a wide range of domains. However, in complex decision-making tasks, pure LLM-based agents tend to exhibit intrinsic bias in their choice of actions, which is inherited from the model's training data and results in suboptimal performance. To develop strategic language agents, i.e., agents that generate flexible language actions and possess strong decision-making abilities, we propose a novel framework that powers LLM-based agents with reinforcement learning (RL). We consider Werewolf, a popular social deduction game, as a challenging testbed that emphasizes versatile communication and strategic gameplay. To mitigate the intrinsic bias in language actions, our agents use an LLM to perform deductive reasoning and generate a diverse set of action candidates. Then an RL policy trained to optimize the decision-making ability chooses an action from the candidates to play in the game. Extensive experiments show that our agents overcome the intrinsic bias and outperform existing LLM-based agents in the Werewolf game. We also conduct human-agent experiments and find that our agents achieve human-level performance and demonstrate strong strategic play.

The INTERSPEECH 2025 Challenge on Multilingual Conversational Speech Language Models (MLC-SLM) promotes multilingual conversational ASR with large language models (LLMs). Our previous SHNU-mASR system adopted a competitive parallel-speech-encoder architecture that integrated Whisper and mHuBERT with an LLM. However, it faced two challenges: simple feature concatenation may not fully exploit complementary information, and the performance gap between LLM-based ASR and end-to-end(E2E) encoder-decoder ASR remained unexplored. In this work, we present an enhanced LLM-based ASR framework that combines fine-tuned Whisper and mHuBERT encoders with an LLM to enrich speech representations. We first evaluate E2E Whisper models with LoRA and full fine-tuning on the MLC-SLM ASR task, and then propose cross-attention-based fusion mechanisms for the parallel-speech-encoder. On the official evaluation set of the MLC-SLM Challenge, our system achieves a CER/WER of 10.69%, ranking on par with the top-ranked Track 1 systems, even though it uses only 1,500 hours of baseline training data compared with their large-scale training sets. Nonetheless, we find that our final LLM-based ASR still does not match the performance of a fine-tuned E2E Whisper model, providing valuable empirical guidance for future Speech-LLM design. Our code is publicly available at https://github.com/1535176727/MLC-SLM.

Text-to-speech (TTS) models have been widely adopted to enhance automatic speech recognition (ASR) systems using text-only corpora, thereby reducing the cost of labeling real speech data. Existing research primarily utilizes additional text data and predefined speech styles supported by TTS models. In this paper, we propose Hard-Synth, a novel ASR data augmentation method that leverages large language models (LLMs) and advanced zero-shot TTS. Our approach employs LLMs to generate diverse in-domain text through rewriting, without relying on additional text data. Rather than using predefined speech styles, we introduce a hard prompt selection method with zero-shot TTS to clone speech styles that the ASR model finds challenging to recognize. Experiments demonstrate that Hard-Synth significantly enhances the Conformer model, achieving relative word error rate (WER) reductions of 6.5\%/4.4\% on LibriSpeech dev/test-other subsets. Additionally, we show that Hard-Synth is data-efficient and capable of reducing bias in ASR.

Silent Speech Interfaces (SSIs) offer a noninvasive alternative to brain-computer interfaces for soundless verbal communication. We introduce Multimodal Orofacial Neural Audio (MONA), a system that leverages cross-modal alignment through novel loss functions--cross-contrast (crossCon) and supervised temporal contrast (supTcon)--to train a multimodal model with a shared latent representation. This architecture enables the use of audio-only datasets like LibriSpeech to improve silent speech recognition. Additionally, our introduction of Large Language Model (LLM) Integrated Scoring Adjustment (LISA) significantly improves recognition accuracy. Together, MONA LISA reduces the state-of-the-art word error rate (WER) from 28.8% to 12.2% in the Gaddy (2020) benchmark dataset for silent speech on an open vocabulary. For vocal EMG recordings, our method improves the state-of-the-art from 23.3% to 3.7% WER. In the Brain-to-Text 2024 competition, LISA performs best, improving the top WER from 9.8% to 8.9%. To the best of our knowledge, this work represents the first instance where noninvasive silent speech recognition on an open vocabulary has cleared the threshold of 15% WER, demonstrating that SSIs can be a viable alternative to automatic speech recognition (ASR). Our work not only narrows the performance gap between silent and vocalized speech but also opens new possibilities in human-computer interaction, demonstrating the potential of cross-modal approaches in noisy and data-limited regimes.

RL with Verifiable Rewards (RLVR) has emerged as a promising paradigm for improving the reasoning abilities of large language models (LLMs). Current methods rely primarily on policy optimization frameworks like PPO and GRPO, which follow generalized policy iteration that alternates between evaluating the current policy's value and improving the policy based on evaluation. While effective, they often suffer from training instability and diversity collapse, requiring complex heuristic tricks and careful tuning. We observe that standard RLVR in math reasoning can be formalized as a specialized finite-horizon Markov Decision Process with deterministic state transitions, tree-structured dynamics, and binary terminal rewards. Though large in scale, the underlying structure is simpler than general-purpose control settings for which popular RL algorithms (e.g., PPO) were developed, suggesting that several sophisticated techniques in existing methods may be reduced or even omitted. Based on this insight, we prove a surprising result: the optimal action can be recovered from the Q-function of a fixed uniformly random policy, thereby bypassing the generalized policy iteration loop and its associated heuristics. We introduce Random Policy Valuation for Diverse Reasoning (ROVER) to translate this principle into a practical and scalable algorithm for LLM math reasoning, a minimalist yet highly effective RL method that samples actions from a softmax over these uniform-policy Q-values. ROVER preserves diversity throughout training, allowing sustained exploration of multiple valid pathways. Across multiple base models and standard math reasoning benchmarks, ROVER demonstrates superior performance in both quality (+8.2 on pass@1, +16.8 on pass@256) and diversity (+17.6\%), despite its radical simplification compared to strong, complicated existing methods.

Since the first instances of online education, where courses were uploaded to accessible and shared online platforms, this form of scaling the dissemination of human knowledge to reach a broader audience has sparked extensive discussion and widespread adoption. Recognizing that personalized learning still holds significant potential for improvement, new AI technologies have been continuously integrated into this learning format, resulting in a variety of educational AI applications such as educational recommendation and intelligent tutoring. The emergence of intelligence in large language models (LLMs) has allowed for these educational enhancements to be built upon a unified foundational model, enabling deeper integration. In this context, we propose MAIC (Massive AI-empowered Course), a new form of online education that leverages LLM-driven multi-agent systems to construct an AI-augmented classroom, balancing scalability with adaptivity. Beyond exploring the conceptual framework and technical innovations, we conduct preliminary experiments at Tsinghua University, one of China's leading universities. Drawing from over 100,000 learning records of more than 500 students, we obtain a series of valuable observations and initial analyses. This project will continue to evolve, ultimately aiming to establish a comprehensive open platform that supports and unifies research, technology, and applications in exploring the possibilities of online education in the era of large model AI. We envision this platform as a collaborative hub, bringing together educators, researchers, and innovators to collectively explore the future of AI-driven online education.

LLM agents increasingly rely on external inference conditions: prompts, tools, memory, SOPs, skills, and harness feedback. These assets can improve task execution without changing model weights, but they are often revised by heuristic reflection or by reusing observed successes and failures as if counts alone were reliable belief. We introduce Bayesian-Agent, a native and cross-harness framework that treats reusable skills and SOPs as hypotheses about whether a frozen model will succeed under a particular prompt, context, and harness environment. Bayesian-Agent records verified trajectory evidence, maintains a feature-conditioned categorical posterior over each skill, and maps posterior state into inspectable actions such as patch, split, compress, retire, and explore. Model-facing prompts receive executable guardrails and failure-mode patches, while posterior summaries remain available for audit. With deepseek-v4-flash, incremental repair improves SOP-Bench from 80\% to 95\%, Lifelong AgentBench from 90\% to 100\%, and RealFin-Bench from 45\% to 65\%. We further evaluate Bayesian-Agent's native backend and optional GenericAgent, mini-swe-agent, and Claude Code backends. The results include positive, negative, saturated, and case-study settings, suggesting that agent skill evolution is best viewed as posterior-guided harness optimization rather than uncalibrated prompt accumulation. The source code is available at https://github.com/DataArcTech/Bayesian-Agent.

Using Large Language Models (LLMs) for relevance assessments offers promising opportunities to improve Information Retrieval (IR), Natural Language Processing (NLP), and related fields. Indeed, LLMs hold the promise of allowing IR experimenters to build evaluation collections with a fraction of the manual human labor currently required. This could help with fresh topics on which there is still limited knowledge and could mitigate the challenges of evaluating ranking systems in low-resource scenarios, where it is challenging to find human annotators. Given the fast-paced recent developments in the domain, many questions concerning LLMs as assessors are yet to be answered. Among the aspects that require further investigation, we can list the impact of various components in a relevance judgment generation pipeline, such as the prompt used or the LLM chosen. This paper benchmarks and reports on the results of a large-scale automatic relevance judgment evaluation, the LLMJudge challenge at SIGIR 2024, where different relevance assessment approaches were proposed. In detail, we release and benchmark 42 LLM-generated labels of the TREC 2023 Deep Learning track relevance judgments produced by eight international teams who participated in the challenge. Given their diverse nature, these automatically generated relevance judgments can help the community not only investigate systematic biases caused by LLMs but also explore the effectiveness of ensemble models, analyze the trade-offs between different models and human assessors, and advance methodologies for improving automated evaluation techniques. The released resource is available at the following link: https://llm4eval.github.io/LLMJudge-benchmark/

Multi-provider LLM gateways reverse proxies that route, load-balance, and rate-limit requests across foundation-model APIs have become critical production infrastructure. Yet the failure modes specific to this architectural layer remain undocumented, scattered across issue trackers and post-mortems with no unifying framework. We introduce , a two-axis taxonomy that classifies failures by their origin layer (Network/Transport, Streaming/Protocol, State/Session, Model~Behavior, Governance/Cost) and their detectability (Loud vs.\ Silent). We populate this taxonomy with five verified catalog entries sourced from public bug reports and first-hand stress testing, each accompanied by a mechanistic root-cause analysis. Three entries include standalone reproduction scripts. Our principal finding is that the most operationally severe failures are silent: they return HTTP~200, pass every standard health check, and corrupt application state in ways that require semantic-level observability to detect. Two such silent failures a concurrency race condition causing history loss and a streaming index collision corrupting tool-call payloads were discovered first-hand during evaluation campaigns.

Large Language Models (LLMs) have demonstrated strong natural language processing and code synthesis capabilities, which has led to their rapid adoption in software engineering applications. However, details about LLM training data are often not made public, which has caused concern as to whether existing bug benchmarks are included. In lieu of the training data for the popular GPT models, we examine the training data of the open-source LLM StarCoder, and find it likely that data from the widely used Defects4J benchmark was included, raising the possibility of its inclusion in GPT training data as well. This makes it difficult to tell how well LLM-based results on Defects4J would generalize, as for any results it would be unclear whether a technique's performance is due to LLM generalization or memorization. To remedy this issue and facilitate continued research on LLM-based SE, we present the GitHub Recent Bugs (GHRB) dataset, which includes 76 real-world Java bugs that were gathered after the OpenAI data cut-off point.

This paper presents an eight-week observational comparison of 68 single-file HTML generations collected across 17 public experiments in the "HTML AI Battle" project between December 10, 2025 and February 4, 2026. Four reasoning model families, GPT, Gemini, Grok, and Claude, were compared under a fixed public-interface protocol with no custom instructions, no personality tuning, and no repair prompts. Each output was evaluated from a rendered browser video using human scores and a Gemini LLM-as-a-judge layer for prompt adherence, functional correctness, and UI quality, then packaged into a standardized social-media protocol spanning X (Twitter), TikTok, and YouTube. The tracker was also used for two supervised predictive analyses: an experiment-level model for 24-hour X impressions and a generation-level model for HTML verbosity. Under this protocol, Claude was the strongest and most consistent family, leading mean performance and winning 9/17 prompts under the primary human weighted score. Longer measured reasoning time was not associated with higher quality overall. Gemini as a judge was significantly more lenient than the human evaluator on functional correctness and overall performance, while stable self-favoring bias remained unresolved. The exploratory X-impressions model remained weak under post-screen cross-validation (MAE = 46,874, R^2 = -0.377), whereas the HTML-lines model performed better, with a model-family-only baseline outperforming prompt-aware alternatives (MAE = 135.2, R^2 = 0.576). Overall, selected pre-publication technical/audio variables were not sufficient to predict 24-hour X reach, while code verbosity was driven much more by model family than by prompt wording. The comparisons remain observational and are limited by public-interface drift, access-path differences, and one primary human scorer.

Large language models (LLMs) are increasingly evaluated and sometimes trained using automated graders such as LLM-as-judges that output scalar scores or preferences. While convenient, these approaches are often opaque: a single score rarely explains why an answer is good or bad, which requirements were missed, or how a system should be improved. This lack of interpretability limits their usefulness for model development, dataset curation, and high-stakes deployment. Query-specific rubric-based evaluation offers a more transparent alternative by decomposing quality into explicit, checkable criteria. However, manually designing high-quality, query-specific rubrics is labor-intensive and cognitively demanding and not feasible for deployment. While previous approaches have focused on generating intermediate rubrics for automated downstream evaluation, it is unclear if these rubrics are both interpretable and effective for human users. In this work, we investigate whether LLMs can generate useful, instance-specific rubrics as compared to human-authored rubrics, while also improving effectiveness for identifying good responses. Through our systematic study on two rubric benchmarks, and on multiple few-shot and post-training strategies, we find that off-the-shelf LLMs produce rubrics that are poorly aligned with human-authored ones. We introduce a simple strategy, RubricRAG, which retrieves domain knowledge via rubrics at inference time from related queries. We demonstrate that RubricRAG can generate more interpretable rubrics both for similarity to human-authored rubrics, and for improved downstream evaluation effectiveness. Our results highlight both the challenges and a promising approach of scalable, interpretable evaluation through automated rubric generation.

Large Language Models (LLMs) have shown promise in accelerating the scientific research pipeline. A key capability for this process is the ability to generate novel research ideas, and prior studies have found settings in which LLM-generated research ideas were judged as more novel than human-expert ideas. However, a good idea should not simply appear to be novel, it should also result in better research after being executed. To test whether AI-generated ideas lead to better research outcomes, we conduct an execution study by recruiting 43 expert researchers to execute randomly-assigned ideas, either written by experts or generated by an LLM. Each expert spent over 100 hours implementing the idea and wrote a 4-page short paper to document the experiments. All the executed projects are then reviewed blindly by expert NLP researchers. Comparing the review scores of the same ideas before and after execution, the scores of the LLM-generated ideas decrease significantly more than expert-written ideas on all evaluation metrics (novelty, excitement, effectiveness, and overall; p < 0.05), closing the gap between LLM and human ideas observed at the ideation stage. When comparing the aggregated review scores from the execution study, we even observe that for many metrics there is a flip in rankings where human ideas score higher than LLM ideas. This ideation-execution gap highlights the limitations of current LLMs in generating truly effective research ideas and the challenge of evaluating research ideas in the absence of execution outcomes.

Reward models (RMs) guide the alignment of large language models (LLMs), steering them toward behaviors preferred by humans. Evaluating RMs is the key to better aligning LLMs. However, the current evaluation of RMs may not directly correspond to their alignment performance due to the limited distribution of evaluation data and evaluation methods that are not closely related to alignment objectives. To address these limitations, we propose RMB, a comprehensive RM benchmark that covers over 49 real-world scenarios and includes both pairwise and Best-of-N (BoN) evaluations to better reflect the effectiveness of RMs in guiding alignment optimization. We demonstrate a positive correlation between our benchmark and the downstream alignment task performance. Based on our benchmark, we conduct extensive analysis on the state-of-the-art RMs, revealing their generalization defects that were not discovered by previous benchmarks, and highlighting the potential of generative RMs. Furthermore, we delve into open questions in reward models, specifically examining the effectiveness of majority voting for the evaluation of reward models and analyzing the impact factors of generative RMs, including the influence of evaluation criteria and instructing methods. Our evaluation code and datasets are available at https://github.com/Zhou-Zoey/RMB-Reward-Model-Benchmark.

As large language models (LLMs) become increasingly integrated into society, their alignment with human morals is crucial. To better understand this alignment, we created a large corpus of human- and LLM-generated responses to various moral scenarios. We found a misalignment between human and LLM moral assessments; although both LLMs and humans tended to reject morally complex utilitarian dilemmas, LLMs were more sensitive to personal framing. We then conducted a quantitative user study involving 230 participants (N=230), who evaluated these responses by determining whether they were AI-generated and assessed their agreement with the responses. Human evaluators preferred LLMs' assessments in moral scenarios, though a systematic anti-AI bias was observed: participants were less likely to agree with judgments they believed to be machine-generated. Statistical and NLP-based analyses revealed subtle linguistic differences in responses, influencing detection and agreement. Overall, our findings highlight the complexities of human-AI perception in morally charged decision-making.

The growing deployment of large language model (LLM) based agents that interact with external environments has created new attack surfaces for adversarial manipulation. One major threat is indirect prompt injection, where attackers embed malicious instructions in external environment output, causing agents to interpret and execute them as if they were legitimate prompts. While previous research has focused primarily on plain-text injection attacks, we find a significant yet underexplored vulnerability: LLMs' dependence on structured chat templates and their susceptibility to contextual manipulation through persuasive multi-turn dialogues. To this end, we introduce ChatInject, an attack that formats malicious payloads to mimic native chat templates, thereby exploiting the model's inherent instruction-following tendencies. Building on this foundation, we develop a persuasion-driven Multi-turn variant that primes the agent across conversational turns to accept and execute otherwise suspicious actions. Through comprehensive experiments across frontier LLMs, we demonstrate three critical findings: (1) ChatInject achieves significantly higher average attack success rates than traditional prompt injection methods, improving from 5.18% to 32.05% on AgentDojo and from 15.13% to 45.90% on InjecAgent, with multi-turn dialogues showing particularly strong performance at average 52.33% success rate on InjecAgent, (2) chat-template-based payloads demonstrate strong transferability across models and remain effective even against closed-source LLMs, despite their unknown template structures, and (3) existing prompt-based defenses are largely ineffective against this attack approach, especially against Multi-turn variants. These findings highlight vulnerabilities in current agent systems.

This paper introduces Fietje, a family of small language models (SLMs) specifically designed for the Dutch language. The model is based on Phi 2, an English-centric model of 2.7 billion parameters. Fietje demonstrated competitive results with larger language models upon its release. A core emphasis of this work is transparency and reproducibility: Fietje is fully open-source, with model weights, datasets, training, and evaluation code all publicly accessible. The paper discusses the performance of Fietje and many other models on an extensive evaluation suite of benchmarks on reasoning, sentiment analysis, world knowledge, linguistic acceptability and word sense disambiguation. Evaluation results illustrate the rapid progress in the field of LLMs, where recent small models outperform older, larger models that were fine-tuned for Dutch. This trend signals an exciting future for Dutch language processing, suggesting that even compact LLMs are becoming increasingly capable. Furthermore, ongoing and future efforts to adapt LLMs to Dutch are poised to enhance these models even further, broadening their applicability and accessibility. Fietje is only an intermediate step in improving accessibility to language technology for users of the Dutch language.

Training Large Language Models (LLMs) incurs significant cost; hence, any strategy that accelerates model convergence is helpful. In this paper, we investigate the ability of a simple idea checkpoint averaging along the trajectory of a training run to improve both convergence and generalization quite early on during training. Here we show that models trained with high learning rates observe higher gains due to checkpoint averaging. Furthermore, these gains are amplified when checkpoints are sampled with considerable spacing in training steps. Our training recipe outperforms conventional training and popular checkpoint averaging baselines such as exponential moving average (EMA) and stochastic moving average (SWA). We evaluate our training recipe by pre-training LLMs, where high learning rates are inherently preferred due to extremely large batch sizes. Specifically, we pre-trained nanoGPT-2 models of varying sizes, small (125M), medium (335M), and large (770M)on the OpenWebText dataset, comprised of 9B tokens. Additionally, we present results for publicly available Pythia LLMs, ranging from 1B to 12B, which were trained on the PILE-deduped dataset containing 207B tokens.

We describe the University of Amsterdam Intelligent Data Engineering Lab team's entry for the SemEval-2024 Task 6 competition. The SHROOM-INDElab system builds on previous work on using prompt programming and in-context learning with large language models (LLMs) to build classifiers for hallucination detection, and extends that work through the incorporation of context-specific definition of task, role, and target concept, and automated generation of examples for use in a few-shot prompting approach. The resulting system achieved fourth-best and sixth-best performance in the model-agnostic track and model-aware tracks for Task 6, respectively, and evaluation using the validation sets showed that the system's classification decisions were consistent with those of the crowd-sourced human labellers. We further found that a zero-shot approach provided better accuracy than a few-shot approach using automatically generated examples. Code for the system described in this paper is available on Github.

Existing question answering (QA) datasets are no longer challenging to most powerful Large Language Models (LLMs). Traditional QA benchmarks like TriviaQA, NaturalQuestions, ELI5 and HotpotQA mainly study ``known unknowns'' with clear indications of both what information is missing, and how to find it to answer the question. Hence, good performance on these benchmarks provides a false sense of security. A yet unmet need of the NLP community is a bank of non-factoid, multi-perspective questions involving a great deal of unclear information needs, i.e. ``unknown uknowns''. We claim we can find such questions in search engine logs, which is surprising because most question-intent queries are indeed factoid. We present Researchy Questions, a dataset of search engine queries tediously filtered to be non-factoid, ``decompositional'' and multi-perspective. We show that users spend a lot of ``effort'' on these questions in terms of signals like clicks and session length, and that they are also challenging for GPT-4. We also show that ``slow thinking'' answering techniques, like decomposition into sub-questions shows benefit over answering directly. We release sim 100k Researchy Questions, along with the Clueweb22 URLs that were clicked.