Large language models (LLMs) frequently generate toxic content, posing significant risks for safe deployment. Current mitigation strategies often degrade generation quality or require costly human annotation. We propose CausalDetox, a framework that identifies and intervenes on the specific attention heads causally responsible for toxic generation. Using the Probability of Necessity and Sufficiency (PNS), we isolate a minimal set of heads that are necessary and sufficient for toxicity. We utilize these components via two complementary strategies: (1) Local Inference-Time Intervention, which constructs dynamic, input-specific steering vectors for context-aware detoxification, and (2) PNS-Guided Fine-Tuning, which permanently unlearns toxic representations. We also introduceParaTox, a novel benchmark of aligned toxic/non-toxic sentence pairs enabling controlled counterfactual evaluation. Experiments on ToxiGen, ImplicitHate, and ParaDetox show that CausalDetox achieves up to 5.34% greater toxicity reduction compared to baselines while preserving linguistic fluency, and offers a 7\times speedup in head selection.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
CoDA: Restoring Contextual Dominance via Copy-Encouraged Attention Intervention for Mitigating RAG Hallucinations
PDF ↗Retrieval-augmented generation reduces hallucination by grounding model outputs in external evidence, yet hallucinations can still occur even when the retrieved context is accurate and sufficient. From the perspective of information routing in the residual stream, this reflects an imbalance where internal parametric knowledge overwhelms external context during generation. We present an attention-centric analysis of RAG hallucination under valid evidence, showing that hallucinated and factual tokens diverge in mid-to-late Transformer layers as context-selective attention routing weakens, allowing parametric influence to dominate the residual stream. Motivated by prior studies showing that some attention heads—often referred to as copying heads—exhibit stronger information transport capacity, we aim to extend similar evidence-carrying behavior to a broader set of attention heads. To this end, we introduce CoDA, a lightweight inference-time attention intervention that amplifies evidence-aligned value states, enabling more attention heads to transport reliable external evidence in a copy-encouraged manner. Experiments demonstrate that CoDA improves contextual faithfulness, reduces hallucination, and remains robust under long and noisy contexts with modest and stable inference overhead.
Dialectal Arabic datasets embody a range of domain, dialect, and quality. To better understand the landscape of these datasets, we perform a computational analysis of the ‘dialectness’ and a set of measures of audio quality. This analysis of the training splits of dialectal Arabic datasets, provides a valuable complement to existing literature surveys of dialectal Arabic.To further address inconsistencies between datasets, we also introduce Arab Voices, a standardized framework for supporting Automatic Speech Recognition in dialectal Arabic. This framework provide access to 31 datasets covering 14 dialects, to better address the limited data availability encountered in dialectal Arabic speech processing. Our benchmark further provides a current evaluation of SOTA tools as well as modern multimodal LLMs at dialectal Arabic ASR.
One critical challenge for large language models (LLMs) in making complex reasoning is their reliance on matching reasoning patterns from training data, instead of proactively selecting the most appropriate cognitive strategy to solve a given task. Existing approaches impose fixed cognitive structures that enhance performance in specific tasks but lack adaptability across diverse scenarios. To address this limitation, we introduce MetaScale, a test-time scaling framework based on meta-thoughts, i.e., adaptive thinking strategies tailored to each task. MetaScale initializes a pool of candidate meta-thoughts, then iteratively selects and evaluates them using a multi-armed bandit algorithm with upper confidence bound selection, guided by a reward model. To further enhance adaptability, a genetic algorithm evolves high-reward meta-thoughts, refining and extending the strategy pool over time. By dynamically proposing and optimizing meta-thoughts at inference time, MetaScale improves both accuracy and generalization across a wide range of tasks. Experimental results demonstrate that MetaScale consistently outperforms standard inference approaches, achieving an 11% performance gain in win rate on Arena-Hard with GPT-4o, improving from 82.14% to 93.14% against GPT-4. Notably, MetaScale scales more effectively with increasing sampling budgets and produces more structured, expert-level responses.
MedErrBench: A Fine-Grained Multilingual Benchmark for Medical Error Detection and Correction with Clinical Expert Annotations
PDF ↗Inaccuracies in existing or generated clinical text may lead to serious adverse consequences, especially if it is a misdiagnosis or incorrect treatment suggestion. With Large Language Models (LLMs) increasingly being used across diverse healthcare applications, comprehensive evaluation through dedicated benchmarks is crucial. However, such datasets remain scarce, especially across diverse languages and contexts. In this paper, we introduce MedErrBench, the first multilingual benchmark for error detection, localization, and correction, developed under the guidance of experienced clinicians. Based on an expanded taxonomy of ten common error types, MedErrBench covers English, Arabic and Chinese, with natural medical cases annotated and reviewed by domain experts. We assessed the performance of a range of general-purpose, language-specific, and medical-domain language models across all three tasks. Our results reveal notable performance gaps, particularly in non-English settings, highlighting the need for clinically grounded, language-aware systems. By making MedErrBench and our evaluation protocols publicly-available, we aim to advance multilingual clinical NLP to promote safer and more equitable AI-based healthcare globally. The dataset is publicly available at: https://github.com/congboma/MedErrBench.
Inference attacks have been widely studied and offer a systematic risk assessment of ML services; however, their implementation and the attack parameters for optimal estimation are challenging for non-experts. The emergence of advanced large language models presents a promising yet largely unexplored opportunity to develop autonomous agents as inference attack experts, helping address this challenge. In this paper, we propose InferPilot, an autonomous agent capable of independently conducting inference attacks without human intervention. We evaluate it on 20 target services. The evaluation shows that our agent, using GPT-4o, achieves a 100.0% task completion rate and near-expert attack performance, with an average token cost of only 0.627 per run. The agent can also be powered by many other representative LLMs and can adaptively optimize its strategy under service constraints. We further perform trace analysis, demonstrating that design choices, such as a multi-agent framework and task-specific action spaces, effectively mitigate errors such as bad plans, inability to follow instructions, task context loss, and hallucinations. We anticipate that such agents could empower non-expert ML service providers, auditors, or regulators to systematically assess the risks of ML services without requiring deep domain expertise.
While recent safety guardrails effectively suppress overtly biased outputs, subtler forms of social bias emerge during complex logical reasoning tasks that evade current evaluation benchmarks. To fill this gap, we introduce a new evaluation framework, PRIME (\textbf{P}uzzle \textbf{R}easoning for \textbf{I}mplicit Biases in \textbf{M}odel \textbf{E}valuation), that uses logic grid puzzles to systematically probe the influence of social stereotypes on logical reasoning and decision making in LLMs. Our use of logic puzzles enables automatic generation and verification, as well as variability in complexity and biased settings. PRIME includes stereotypical, anti-stereotypical, and neutral puzzle variants generated from a shared puzzle structure, allowing for controlled and fine-grained comparisons. We evaluate multiple model families across puzzle sizes and test the effectiveness of prompt-based mitigation strategies. Focusing our experiments on gender stereotypes, our findings highlight that models consistently reason more accurately when solutions align with stereotypical associations. This demonstrates the significance of PRIME for diagnosing and quantifying social biases perpetuated in the deductive reasoning of LLMs, where fairness is critical.
We present VoiceStar, the first zero-shot TTS model that achieves both output duration control and extrapolation. VoiceStar is an autoregressive encoder-decoder neural codec language model, that leverages a novel Progress-Monitoring Rotary Position Embedding (PM-RoPE) and is trained with Continuation-Prompt Mixed (CPM) training. PM-RoPE enables the model to better align text and speech tokens, indicates the target duration for the generated speech, and also allows the model to generate speech waveforms much longer in duration than those seen during training. CPM training also helps to mitigate the training/inference mismatch, and significantly improves the quality of the generated speech in terms of speaker similarity and intelligibility. VoiceStar outperforms or is on par with current state-of-the-art models on short-form benchmarks such as LibriSpeech and Seed-TTS, and significantly outperforms these models on long-form/extrapolation benchmarks (20-50s) in terms of intelligibility and naturalness. Code and model: https://github.com/jasonppy/VoiceStar. Audio samples: https://jasonppy.github.io/VoiceStar_web.
Large language models (LLMs) are increasingly used for emotional support and mental health–related interactions outside clinical settings, yet little is known about how people evaluate and relate to these systems in everyday use. We analyze 5,126 Reddit posts from 47 mental health communities describing experiential or exploratory use of AI for emotional support or therapy. Grounded in the Technology Acceptance Model and therapeutic alliance theory, we develop a theory-informed annotation framework and apply a hybrid LLM–human pipeline to analyze evaluative language, adoption-related attitudes, and relational alignment at scale. Our results show that engagement is shaped primarily by narrated outcomes, trust, and response quality, rather than emotional bond alone. Positive sentiment is most strongly associated with task and goal alignment, while companionship-oriented use more often involves misaligned alliances and reported risks such as dependence and symptom escalation. Overall, this work demonstrates how theory-grounded constructs can be operationalized in large-scale discourse analysis and highlights the importance of studying how users interpret language technologies in sensitive, real-world contexts.
Large-language-model (LLM) agents are becoming competent at straightforward web tasks, such as opening an item page or submitting a form, but still struggle with objectives that require long-horizon navigation, large-scale information extraction, and reasoning under constraints. We present WebDART, a general framework that enables a single LLM to handle such complex chores. WebDART (i) dynamically decomposes each objective into three focused subtasks—navigation, information extraction, and execution—so the model concentrates on one skill at a time, and (ii) continuously re-plans the decomposition as new webpages are revealed, taking advantage of newly discovered filters or shortcuts and avoiding redundant exploration. Evaluated on WebChoreArena, WebDART lifts end-to-end success rates by up to 13.7 percentage points over previous state-of-the-art agents, while matching their performance on the easier WebArena suite and completing tasks with up to 14.7 fewer navigation steps. Code will be publicly available.
Recent progress in large language model (LLM) reasoning has focused on domains like mathematics and coding, where abundant high-quality data and objective evaluation metrics are readily available. In contrast, progress in scientific reasoning remains limited in domains such as medicine and materials science due to restricted dataset coverage and the inherent complexity of open-ended scientific questions. To address these challenges, we propose a general framework for sustainable scientific reasoning QA generation, and introduce WildSci, a new dataset of domain-specific science questions automatically synthesized from peer-reviewed literature, spanning 9 scientific disciplines and 26 subdomains. WildSci enables scalable training with well-defined reward signals in a multiple-choice format. We further apply reinforcement learning to finetune models on WildSci and analyze the resulting training dynamics, including domain-specific performance changes, response behaviors, and generalization trends. Experiments on a suite of scientific benchmarks demonstrate the effectiveness of our framework and dataset. We release WildSci to enable scalable and sustainable research in scientific reasoning.
With the rapid adoption of large language models (LLMs), conversational AI agents have become widely deployed across real-world applications. To enhance safety, these agents are often equipped with guardrails that moderate harmful content. Identifying the guardrails in an agent thus becomes critical for adversaries to understand the system and design guard-specific attacks. In this work, we introduce AP-Test, a novel approach that leverages guard-specific adversarial prompts to detect the identity of guardrails deployed in black-box AI agents. Our method addresses key challenges in this task, including the influence of safety-aligned LLMs and other guardrails, as well as a lack of principled decision-making strategies. AP-Test employs two complementary testing strategies, input and output guard tests, and a new metric, match score, to enable robust identification. Experiments across diverse agents and four open-source guardrails demonstrate that AP-Test achieves perfect classification accuracy in multiple scenarios. Ablation studies further highlight the necessity of our proposed components. Our findings reveal a practical path toward guardrail identification in real-world AI systems.
Understanding which parts of the retrieved context contribute to a large language model’s generated answer is essential for building interpretable and trustworthy retrieval-augmented generation. We propose a novel framework that formulates context attribution as a combinatorial multi-armed bandit problem. We utilize Linear Thompson Sampling to efficiently identify the most influential context segments while minimizing the number of model queries. Our reward function leverages token log-probabilities to measure how well a subset of segments supports the original response, making it applicable to both open-source and black-box API-based models. Unlike SHAP and other perturbation-based methods that sample subsets uniformly, our approach adaptively prioritizes informative subsets based on posterior estimates of segment relevance, reducing computational costs. Experiments on multiple QA benchmarks demonstrate that our method achieves up to 30% reduction in model queries while matching or exceeding the attribution quality of existing approaches.
We present a neuromuscular speech interface that translates silently voiced articulations directly into text. We record surface electromyographic (EMG) signals from multiple articulatory sites on the face and neck as participants *silently* articulate speech, enabling direct EMG-to-text translation. Such an interface has the potential to restore communication for individuals who have lost the ability to produce intelligible speech due to laryngectomy, neuromuscular disease, stroke, or trauma-induced damage (e.g., radiotherapy toxicity) to the speech articulators. Prior work has largely focused on mapping EMG collected during *audible* articulation to time-aligned audio targets or transferring these targets to *silent* EMG recordings, which inherently requires audio and limits applicability to patients who can no longer speak. In contrast, we propose an efficient representation of high-dimensional EMG signals and demonstrate direct sequence-to-sequence EMG-to-text conversion at the phonemic level without relying on time-aligned audio.
Knowledge-Based Visual Question Answering (KB-VQA) requires grounding visual queries to external knowledge beyond directly observable content in images.While recent multi modal large language models (MLLMs) show strong perceptual abilities, they struggle on KB-VQA tasks requiring groundings from both fine-grained entity and evidence levels.Most existing multi-modal retrieval augmented generation (MM-RAG) methods tightly couple entity discrimination and section-level evidence ranking into a single re-ranking stage, leading to high cost and limited generalization.In this work, we revisit existing MM-RAG solutions from a workflow perspective and argue both entity-level and fact-level groundings are key bottlenecks.We observe that although MLLMs often fail under open-ended entity naming, they can better identify the correct entity when selecting from a small set of candidate names.Based on this insight, we propose a simple and training-free identify-before-answer IBA framework that decouples entity identification from section-level re-ranking.Our approach prompts an MLLM to select high-confidence entities using only candidate names, followed by an off-the-shelf textual re-ranker for evidence selection.Experiments on Encyclopedic-VQA and InfoSeek show that our method consistently outperforms fine-tuned multi-modal re-ranking baselines while reducing training and inference complexity.Additional analyses reveal that the improvements arise not only from better entity identification, but also from selecting more informative evidence once correct entity is fixed.Our implementation is made public to ease reproducibility
Time series anomaly detection (TSAD) has traditionally focused on binary classification and often lacks the fine-grained categorization and explanatory reasoning required for transparent decision-making. To address these limitations, we propose Time-series Reasoning for Anomaly (Time-RA), a novel task that reformulates TSAD from a discriminative into a generative, reasoning-intensive paradigm. To facilitate this, we introduce RATs40K, the first real-world large-scale multimodal benchmark with ~40,000 samples across 10 domains, integrating raw time series, textual context, and visual plots with structured reasoning annotations. Extensive benchmarking shows that while supervised fine-tuning and visual representations boost diagnostic accuracy and reasoning consistency, performance varies across complex scenarios. Notably, fine-tuned models demonstrate strong "plug-and-play" transferability, outperforming traditional baselines on unseen real-world datasets. Our work establishes a foundation for interpretable, multimodal time series analysis. All code and the RATs40K dataset are fully open-sourced to facilitate future research.
Follow the Path: Reasoning over Knowledge Graph Paths to Improve Large Language Model Factuality
PDF ↗We introduce fs1, a simple yet effective method that improves the factuality of reasoning traces by sourcing them from large reasoning models and grounding them by conditioning on knowledge graph (KG) paths. We fine-tune eight instruction-tuned Large Language Models (LLMs) on 3.9K factually grounded reasoning traces and rigorously evaluate them on six complex open-domain question-answering (QA) benchmarks encompassing 23.9K questions. Our results demonstrate that our fs1-tuned model consistently outperforms instruction-tuned counterparts with parallel sampling by 6-14 absolute points (pass@). Our detailed analysis shows that fs1 considerably improves model performance over more complex questions (requiring 3 or more hops on KG paths) and numerical answer types compared to the baselines. Furthermore, in single-pass inference, we notice that smaller LLMs show the most improvements. While prior works demonstrate the effectiveness of reasoning traces primarily in the STEM domains, our work shows strong evidence that anchoring reasoning to factual KG paths is a critical step in transforming LLMs for reliable knowledge-intensive tasks.
LLMs show promise in code generation, yet their effectiveness for IT automation tasks, particularly for tools like Ansible, remains understudied. Existing benchmarks rely primarily on synthetic tasks that fail to capture the needs of practitioners who use IT automation tools. We present ExITBench (Execution-based IT Automation Benchmark), a benchmark of 126 diverse tasks (e.g., configuring servers and managing files) in which each task captures state reconciliation - a core property of IT automation tools. ExITBench evaluates LLMs’ ability to generate functional Ansible automation scripts via dynamic execution in controlled environments. We evaluate 14 open-source and 3 proprietary LLMs and find that GPT-4.1-Mini achieves the best pass@10 rate of 23.9%, while Claude-3.5-Sonnet achieves the best pass@1 performance. To explain the low performance, we analyze 1,517 execution failures across the evaluated LLMs and identify two prevalent semantic error categories: failures in state-reconciliation reasoning (42.117% combined from variable (12.287%), host (10.363%), path (10.511%), and template (8.956%) issues) and deficiencies in module-specific execution knowledge (26.203% combined from attribute & parameter (17.617%) and module (8.586%) errors). Our findings reveal key limitations in LLMs’ ability to address state reconciliation and apply specialized module knowledge, indicating that reliable IT automation with LLM-based agents need major advances in state reasoning and domain-specific execution.
LANTERN in the Event Stream: Training-Free Temporal Knowledge Graph Forecasting by Balancing Inertia and Shifts
PDF ↗Temporal knowledge graph forecasting(TKGF) asks a model to rank the mostplausible future entity for a query such as(s, r, ?, t) from historical events. Recenttraining-free methods use large languagemodels (LLMs) for this task, but their accuracydepends heavily on which past events areshown in the prompt under a tight contextbudget. We present LANTERN, a training-freeprompting framework that addresses thisbottleneck by combining two complementaryviews of history: a long-window strengthscore for stable interaction patterns anda short-window novelty score for suddenchanges. LANTERN first filters unhelpfulevents, then selects a compact evidence setwith Pareto-greedy selection, and finally addsone structure-aware analogical demonstration.Across ICEWS14, ICEWS05-15, ICEWS18,and GDELT, LANTERN consistently outperforms the state-of-the-art training-free baselineAnRe under the same backbone and 2-hopcandidate protocol, improving Hits@1 by upto 2.5 points and MRR by up to 1.2 points.
Large reasoning models (LRMs) achieve state-of-the-art performance on challenging benchmarks by generating long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence. One approach to reduce this latency is to evict entries from the key-value (KV) cache, thereby reducing the active context used in attention computation. However, such sparse attention methods suffer from severe accuracy degradation on reasoning tasks due to cumulative selection errors and the evolving importance of tokens over long derivations. We present DELTA, a training-free sparse attention mechanism that improves computational efficiency without sacrificing model accuracy. DELTA partitions transformer layers into three groups: initial layers that use full attention, a small set of Δ-layers that identify salient tokens via aggregated head-level attention scores, and subsequent sparse-attention layers that attend only to the selected subset. This design preserves the full KV cache in GPU memory for accuracy, while avoiding expensive full-attention computation over many layers. On reasoning benchmarks such as AIME and GPQA-Diamond, DELTA matches or surpasses full attention in accuracy, while reducing the number of attended tokens by up to 4.25× and delivering 1.54× end-to-end speedup. Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning. The code is available at https://github.com/hoenza/DELTA.