Large language models (LLMs) often struggle to use tools reliably in domain-specific settings, where APIs may be idiosyncratic, under-documented, or tailored to private workflows. This highlights the need for effective adaptation to task-specific tools. We propose RIMRULE, a neuro-symbolic approach for LLM adaptation based on dynamic rule injection. Compact, interpretable rules are distilled from failure traces and injected into the prompt during inference to improve task performance. These rules are proposed by the LLM itself and consolidated using a Minimum Description Length (MDL) objective that favors generality and conciseness. Each rule is stored in both natural language and a structured symbolic form, supporting efficient retrieval at inference time. Experiments on tool-use benchmarks show that this approach improves accuracy on both seen and unseen tools without modifying LLM weights. It outperforms prompting-based adaptation methods and complements finetuning, offering an interpretable layer of inference-time generalization. Moreover, rules learned from one LLM can be reused to improve others, including long reasoning LLMs, highlighting the portability of symbolic knowledge across architectures.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
From Individual to Common: An Early Exploration of Consensus in Non-verifiable Data for Balanced Preference Optimization
PDF ↗Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated remarkable effectiveness in boosting the objective performance (e.g., reasoning) of Large Language Models (LLMs) through rule-based, on-policy self-improvement strategies. However, optimizing LLMs for subjective capabilities and alignment with human preferences remains challenging due to the non-verifiable nature. Most prior works use datasets comprising response pairs with substantial quality gaps labeled by a strong external judge. While effective for preference metrics, this paradigm often incurs an “alignment tax”, where the model’s objective performance on downstream tasks degrades as it overfits to subjective preferences. In this work, we introduce Donkey, a high-quality, non-verifiable dataset where response pairs differ only by subtle nuances. We find that LLMs optimized on Donkey via preference learning outperform those trained on data with explicit quality gaps, while simultaneously maintaining their objective capabilities. Furthermore, we observe that preference signals on Donkey can be decomposed into consensus preferences and individual preferences. Our analysis reveals that distilling consensus preferences provides a significantly more data-efficient signal for preference optimization. Our findings underscore the importance of leveraging nuanced preference signals and the consensus of multiple judges for advancing subjective LLM alignment. Our code and data will be available at https://github.com/SJY8460/Donkey.
Human conversation relies heavily on *conversational implicature*, in which speakers convey meanings that are suggested rather than explicitly stated. Although recent large language models (LLMs) exhibit strong conversational fluency, they remain unreliable when interpretation depends on reasoning that integrates social and contextual cues, a process rarely articulated in text. We introduce **DRinQ**, a benchmark for evaluating pragmatic reasoning about conversational implicature in question utterances, designed to isolate pragmatic variation while holding each question’s surface form fixed. To support scalable evaluation, we propose a semi-automated pipeline that produces question-context-interpretation instances with systematic variation. Across evaluations, we find a consistent generation-inference asymmetry: while state-of-the-art models can generate plausible pragmatic scenarios when guided, they often fail to recover the intended implication at inference time. For smaller models, structured prompting improves alignment with human judgments. A comparative writing study further reveals complementary strengths: human authors tend to produce safer, predictable contexts, whereas models generate varied scenarios with interpretations that sometimes exceed contextual support. These findings highlight persistent challenges in modeling conversational implicature and motivate more context-sensitive evaluation frameworks.
We propose a corpus-dependent alternative to byte encoding that learns fixed-length atomic codes for characters directly from text, which we refer to as Latom (Learned Atom-based Encoding).We instantiate this framework by training an HMM on N-repeated character sequences to estimate "atom" posteriors, followed by a Hungarian assignment yielding a globally optimal one-to-one character-code mapping.Across 14 languages, the encodings improve intrinsic metrics, including token counts after subword tokenization and bigram perplexity, with appropriate code lengths.On Amazon Reviews in six languages, Latom improves text classification accuracy and reduces decoding errors in language model generation.Overall, these results demonstrate that character encodings can be learned from corpus statistics while remaining reversible and compatible with standard tokenization pipelines.
Traditional reinforcement learning from human feedback (RLHF) for large language models (LLMs) relies on expensive human-annotated datasets, while Reinforcement Learning from AI Feedback (RLAIF) also incurs significant costs, requiring the collection of diverse prompts and corresponding responses, often necessitating external reward models or proprietary models like GPT-4 to annotate preference pairs. In this work, we introduce Self-Alignment Optimization (SAO), a fully self-synthetic framework for LLM alignment, where all training data, including prompts (i.e., user queries), responses, and preferences, are generated by the model itself. Specifically, SAO first instructs the LLM to engage in persona role-play and generate diverse prompts and responses, which are then self-evaluated for preference optimization. Extensive experiments demonstrate that SAO effectively enhances the model’s chat capabilities on standard benchmarks like AlpacaEval 2.0, while maintaining strong performance on downstream objective tasks (i.e.,, question-answering, math reasoning). Our work provides a practical solution for self-improvement in aligning LLMs, and the code for reproducing our results is available at: https://github.com/SJY8460/SAO.
Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection
PDF ↗We investigate the efficacy of Large Language Models (LLMs) in detecting implicit and explicit hate speech, examining how models with minimal safety alignment (uncensored) compare with more heavily aligned (censored) counterparts in a deployed-model setting when deployed using political personas. While uncensored models are often framed as offering a less constrained perspective, our results reveal a trade-off: censored models outperform their uncensored counterparts in both accuracy and robustness, achieving 69.0% versus 64.1% strict accuracy. However, this higher performance is also associated with greater resistance to persona-based influence, while uncensored models are more malleable to ideological framing. Furthermore, we identify critical failures across all models in understanding nuanced language such as irony. We also find alarming fairness disparities in performance across different targeted groups and systemic overconfidence that renders self-reported certainty unreliable. These findings challenge the notion of LLMs as objective arbiters and highlight the need for more sophisticated auditing frameworks that account for fairness, calibration, and ideological consistency. Taken together, these results point to censorship-as-deployed rather than safety alignment in isolation as the more appropriate frame for interpreting model differences.
English financial NLP has progressed rapidly through benchmarks for sentiment, document understanding, and financial question answering, while Arabic financial NLP remains comparatively under-explored despite strong practical demand for trustworthy finance and Islamic-finance assistants. We introduce SAHM, a document-grounded benchmark and instruction-tuning dataset for Arabic financial NLP and Shari’ah-compliant reasoning. SAHM contains 14,380 expert-verified instances spanning seven tasks: AAOIFI standards QA, fatwa-based QA/MCQ, accounting and business exams, financial sentiment analysis, extractive summarization, and event–cause reasoning, curated from authentic regulatory, juristic, and corporate sources. We evaluate 19 strong open and proprietary LLMs using task-specific metrics and rubric-based scoring for open-ended outputs, and find that Arabic fluency does not reliably translate to evidence-grounded financial reasoning: models are substantially stronger on recognition-style tasks than on generation and causal reasoning, with the largest gaps on event–cause reasoning. We release the benchmark, evaluation framework, and an instruction-tuned model to support future research on trustworthy Arabic financial NLP.
Cross-architecture GPU code transpilation is essential for unlocking low-level hardware portability, yet no scalable solution exists. We introduce CASS, the first dataset and model suite for source- and assembly-level GPU translation (CUDA ↔ HIP, SASS ↔ RDNA3). CASS contains 60k verified host-device code pairs, enabling learning-based translation across both ISA and runtime boundaries. We generate each sample using our automated pipeline that scrapes, translates, compiles, and aligns GPU programs across vendor stacks. Leveraging CASS, we train a suite of domain-specific translation models that achieve 88.2% accuracy on CUDA → HIP and 69.1% on SASS → RDNA3, outperforming commercial baselines including GPT-5.1, Claude-4.5, and Hipify by wide margins. Generated code matches native performance in 85% of cases, preserving both runtime and memory behavior. To support rigorous evaluation, we introduce CASS-Bench, a curated benchmark spanning 18 GPU domains with ground-truth execution. All data, models, and evaluation tools will be released as open source to support progress in GPU compiler tooling, binary compatibility, and LLM-guided code translation.
What Makes a Good Curriculum? Disentangling the Effects of Data Ordering on LLM Mathematical Reasoning
PDF ↗Curriculum learning (CL), which orders training data from easy to hard, has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction—forward or reverse—is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that: (i) no curriculum strategy dominates universally—the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) Task-aligned curricula focus on shaping the model’s final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.
Social media are shifting towards pluralism — community-governed platforms where groups define their own norms. What violates rules in one community may be perfectly acceptable in another. Can AI models help moderate such pluralistic communities? We formalize the task as a multiple-choice problem, mirroring how human moderators operate in the real world: given a comment and its surrounding context, identify which specific rule, if any, is violated. We introduce PluRule, a multimodal, multilingual benchmark for detecting 13,371 rule violations across 1,989 Reddit communities spanning 2,885 rules in 9 languages. Using this benchmark, we show that state-of-the-art vision-language models struggle significantly: even GPT-5.2 with high reasoning performs only slightly better than a trivial baseline. We also find that bigger models and increased context provide marginal gains, and universal rules like civility and self-promotion are easier to detect. Our results show that moderation of pluralistic communities on social media is a fundamental challenge for language models. Our code and benchmark are publicly available.
Real-world fact-checking often involves verifying claims grounded in structured data at scale. Despite substantial progress in fact-verification benchmarks, this setting remains largely underexplored. In this work, we introduce ClaimDB, a fact-verification benchmark where the evidence for claims is derived from compositions of millions of records and multiple tables. ClaimDB consists of 80 unique real-life databases covering a wide range of domains, from governance and healthcare to media, education and the natural sciences. At this scale, verification approaches that rely on "reading" the evidence break down, forcing a timely shift toward reasoning in executable programs. We conduct extensive experiments with 30 state-of-the-art proprietary and open-source (below 70B) LLMs and find that more than half score below 55% accuracy. Our analysis also reveals that both closed- and open-source models struggle with abstention – the ability to admit that there is no evidence to decide – raising doubts about their reliability in high-stakes data analysis tasks. We release the benchmark, code, and the LLM leaderboard at https://claimdb.github.io.
Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning. Nevertheless, at the limits of model capability, CoT often proves insufficient, and its strictly sequential nature constrains test-time scalability. A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution space. Although promising, our analysis reveals a fundamental misalignment between general-purpose post-training and DAC-style inference, which limits the model’s capacity to fully leverage this potential. To bridge this gap and fully unlock LLMs’ reasoning capabilities on the most challenging tasks, we propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity. At each step, the policy decomposes a problem into a group of subproblems, solves them sequentially, and addresses the original problem conditioned on the subproblem solutions, with both decomposition and solution integrated into RL training. Under comparable training settings, our DAC-style framework endows the model with a higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks. The code is available at the [provided link](https://github.com/MasterVito/DAC-RL).
Relation extraction (RE) is a core task in natural language processing. Traditional approaches typically frame RE as a supervised learning problem, directly mapping context to labels—an approach that often suffers from poor out-of-domain (OOD) generalization. Inspired by the workflow of human annotators, we reframe RE as a reasoning task guided by annotation guidelines and introduce R1-RE, the first reinforcement learning with verifiable reward (RLVR) framework for RE tasks. Our method elicits the reasoning abilities of small language models for annotation tasks, resulting in significantly improved OOD robustness. We evaluate our approach on the public Sem-2010 dataset and a private MDKG dataset. The R1-RE-7B model attains an average OOD accuracy of approximately 70%, on par with leading proprietary models such as GPT-4o. Additionally, our comprehensive analysis provides novel insights into the training dynamics and emergent reasoning behaviors of the RLVR paradigm for RE.
Search-augmented LLM agents can produce deep research reports (DRRs), but verifying claim-level factuality remains challenging. Existing fact-checkers usually target general-domain atomic claims, and there is no benchmark to test whether such verifiers transfer to DRRs.Yet building such a benchmark for DRR fact-checkers is itself difficult because it requires expert judgments over cognitively demanding, domain-specific claims.In a controlled study with PhD-level specialists, unassisted experts achieve only 60.8% accuracy on hidden known-answer claims. We therefore propose evolving benchmarking via **Audit-then-Score** (**AtS**), in which labels and rationales remain revisable: when a verifier disagrees with the current benchmark, it submits evidence; an auditor adjudicates the dispute; and accepted revisions update the benchmark before scoring. After three additional **AtS** rounds, expert accuracy rises to 90.9%, showing that experts are better auditors than one-shot labelers.We instantiate **AtS** as **DeepFactBench**, a versioned DRR factuality benchmark with auditable rationales, and introduce **DeepFactEval**, a claim-level verifier.On the frozen **DeepFactBench** release, **DeepFactEval** achieves 83.4% accuracy, outperforming the best prior deep-research and traditional fact-checkers by 14.3 and 24.9 points, respectively, and transferring well to external factuality datasets.
GameplayQA: A Benchmarking Framework for Decision-Dense POV-Synced Multi-Video Understanding of 3D Virtual Agents
PDF ↗Multimodal LLMs are increasingly deployed as perceptual backbones for autonomous agents in 3D environments, from robotics to virtual worlds. These applications require agents to perceive rapid state changes, attribute actions to the correct entities, and reason about concurrent multi-agent behaviors from a first-person perspective, capabilities that existing benchmarks do not adequately evaluate. We introduce GameplayQA, a framework for evaluating agentic-centric perception and reasoning through video understanding. Specifically, we densely annotate multiplayer 3D gameplay videos at 1.22 labels/second, with time-synced, concurrent captions of states, actions, and events structured around a triadic system of Self, Other Agents, and the World, a natural decomposition for multi-agent environments. From these annotations, we refined 2.4K diagnostic QA pairs organized into three levels of cognitive complexity, accompanied by a structured distractor taxonomy that enables fine-grained analysis of where models hallucinate. Evaluation of frontier MLLMs reveals a substantial gap from human performance, with common failures in temporal and cross-video grounding, agent-role attribution, and handling the decision density of the game. We hope GameplayQA stimulates future research at the intersection of embodied AI, agentic perception, and world modeling.
SMARTER: A Data-efficient Framework to Improve Toxicity Detection with Explanation via Self-augmenting Large Language Models
PDF ↗To address toxic content on social media, we introduce SMARTER, a data-efficient 2-stage framework for explainable content moderation using Large Language Models (LLMs). In Stage 1, we leverage LLMs’ own outputs to generate synthetic explanations for correct and incorrect labels, enabling preference optimization with minimal supervision. In Stage 2, we refine explanation quality through cross-model training, allowing weaker models to align with stronger ones. Experiments on 3 benchmarks (HateXplain, Latent Hate, Implicit Hate) show SMARTER achieves up to 13% macro-F1 improvement over few-shot baselines using only 6-57% of training data. Our framework offers a scalable strategy for low-data settings by harnessing LLMs’ self-improvement for explainable moderation.
The ability to accurately align LLMs with diverse population groups on subjective questions would have great value. In this work, we show that adding simple supervision can more consistently improve the alignment of LLM-generated distributions with diverse population groups, as measured across three datasets spanning public health, public opinion, and values and beliefs. Beyond evaluating average alignment, we also report how alignment varies across specific groups. Our broad findings provide insights into the distributional alignment of LLM generations with diverse populations. By conducting evaluation over many LLMs and prompting strategies, we provide a benchmark to stimulate future research.
Detecting deceptive behavior in LLMs is typically done post-hoc on outputs or by probing static activations. We instead treat deception as a dynamic process, a trajectory through the model’s hidden-state space during inference. We capture layerwise activations at sparse "decision points" where the model is uncertain between competing tokens, forming activation trajectories for matched truthful vs. deceptive responses across strategic deception, sycophancy, instructed deception, and confabulation. Across GPT-2 and Llama variants, deceptive generation is associated with changes in trajectory geometry, but increases in path length are model and deception-type-dependent. Sycophancy shows the clearest signal, whereas instructed deception yields near-null signatures. With just 7 geometric features, a lightweight classifier achieves performance comparable to PCA-reduced probing at matched dimensionality for binary sycophancy detection and shows preliminary utility for 4-way deception-type classification. These findings indicate that trajectory-based monitoring can provide process-level signals associated with deceptive generation during inference, complementing methods that focus on endpoint activation states.
Audio-language pretraining (ALP) holds promise for learning general-purpose audio representation, yet remains underexplored. Crucially, there is no consensus on whether audio–language models can build effective general-purpose audio encoders, nor a systematic understanding of how pretraining objectives behave across diverse tasks and scales.We identify three key barriers: limited scale of audio-text corpora, limited coverage of audio attributes in existing caption corpora, and lack of systematic exploration and evaluation.To fill this gap, we present the first principled empirical study of ALP.We first introduce CaptionStew, a 10.7M caption dataset aggregating open-source audio-text corpora across multiple domains and captioning focuses.We then conduct the first comprehensive evaluation comparing contrastive and captioning objectives for learning audio representation across speech, music, and environmental sound tasks.Our results not only demonstrate that ALP yields competitive, transferable representations, but reveal critical trade-offs: contrastive learning offers superior data efficiency, while captioning exhibits better scalability.Furthermore, we find that the benefits of supervised initialization often diminish at larger scales, challenging common practices.By grounding these claims in empirical evidence, we establish a viable pathway toward general-purpose audio representation learning, guiding future research.
Propaganda is a widely used approach for shaping public opinion and disseminating misinformation in news media. While it has recently gained significant attention within the NLP community, research on fine grained propaganda detection remains heavily concentrated in high resource languages. To bridge this gap, we introduce KinyaProp, the first fine-grained propaganda dataset of its kind for Kinyarwanda and, to our knowledge, the first such resource created for a Bantu language. Using this dataset, we evaluate whether state-of-the-art LLMs can function as reliable annotators in a genuinely low resource and culturally grounded setting. Our results show that current multilingual LLMs do not reliably approximate human annotation behavior. Instead, they behave as conservative annotators whose performance is largely limited to lexically explicit cues, substantially under-identifying propaganda and exhibiting extremely low and unstable performance on discourse-level techniques. Our findings highlight an important limitation of recent successes in LLM based annotation reported for high resource languages, demonstrating that such results do not readily transfer to low resource settings, where scalable annotation would be most valuable. We release KinyaProp to support future research on fine grained propaganda detection and to enable more robust evaluation of multilingual models in underrepresented languages.