As language models accelerate scientific research by automating hypothesis generation and implementation, a new bottleneck emerges: evaluating and filtering hundreds of AI-generated ideas without exhaustive experimentation. We ask whether LMs can learn to forecast the empirical success of research ideas before any experiments are run. We study comparative empirical forecasting: given a benchmark-specific research goal and two candidate ideas, predict which will achieve better benchmark performance. We construct a dataset of 11,488 idea pairs grounded in objective outcomes from PapersWithCode. While off-the-shelf 8B-parameter models struggle (30% acc.), SFT dramatically boosts performance to 77.1%, outperforming GPT-5 (61.1%). By framing evaluation as a reasoning task via Reinforcement Learning with Verifiable Rewards (RLVR), we train models to discover latent reasoning paths, achieving 71.35% acc. with interpretable justifications. Through additional ablations and out-of-distribution tests, we show robustness to surface-level heuristics and transfer to both a cross-domain time-split test set and an independently constructed test set. Our results demonstrate that compute-efficient small language models can serve as effective, objective verifiers, offering a scalable path for autonomous scientific discovery.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Scientific peer reviews frequently contain conflicting expert judgments, and the increasing scale of conference submissions makes it challenging for Area Chairs and editors to reliably identify and interpret such disagreements. Existing approaches typically frame reviewer disagreement as binary contradiction detection over isolated sentence pairs, abstracting away the review-level context and obscuring differences in the severity of evaluative conflict. In this work, we introduce a fine-grained formulation of reviewer contradiction analysis that operates over full peer reviews by explicitly identifying contradiction evidence spans and assigning graded disagreement intensity scores. To support this task, we present RevCI, an expert-annotated benchmark of peer-review pairs with evidence-level contradiction annotations with graded intensity labels. We further propose IMPACT, a structured multi-agent framework that integrates aspect-conditioned evidence extraction, deliberative reasoning, and adjudication to model reviewer contradictions and their intensity. To support efficient deployment, we distill IMPACT into TIDE, a small language model that predicts contradiction evidence and intensity in a single forward pass. Experimental results show that IMPACT substantially outperforms strong single-agent and generic multi-agent baselines in both evidence identification and intensity agreement, while TIDE achieves competitive performance at significantly lower inference cost.
Assessing how well a large language model (LLM) understands human, rather than merely text, remains an open challenge.To bridge the gap, we introduce Sentient Agent as a Judge(SAGE), an automated evaluation framework that measures an LLM’s higher-order social cognition.SAGE instantiates a “Sentient Agent” – an LLM-powered agent that simulates human-like emotional changes and inner thoughts to provide a more realistic evaluation of the tested model in multi-turn conversations.At every turn, the agent reasons about (i) how its emotion changes, (ii) how it feels, and (iii) how it should reply, yielding a numerical emotion trajectory and interpretable inner thoughts.Experiments on 100 supportive-dialogue scenarios show that the final Sentient emotion score correlates strongly with Barrett-Lennard Relationship Inventory (BLRI) ratings and utterance-level empathy metrics, validating psychological fidelity. Human evaluation further demonstrates 85.3% consistency between the agent’s emotional reasoning and human judgments. We also build a public Sentient Leaderboard covering 18 commercial and open-source models that uncovers substantial gaps (up to 4\times) between frontier systems (GPT-4o-Latest, Gemini2.5-Pro) and earlier baselines, gaps not reflected in conventional leaderboards (e.g. Arena). SAGE thus provides a principled, scalable, and interpretable tool for tracking progress toward genuinely empathetic and socially adept language agents.
Med-SRAF: A Multi-Agent Framework for Medical Reasoning via Semantic Routing and Agentic Fusion
PDF ↗While Retrieval-Augmented Generation (RAG) has become a standard paradigm for mitigating hallucinations in Large Language Models (LLMs), its effectiveness in complex medical reasoning remains limited. Existing RAG methods suffer from two main challenges: First, **Semantic Drift**: without explicit domain constraints, LLM-driven query decomposition often deviates from the original clinical intent, introducing substantial noise that degrades retrieval relevance. Second, **Concatenation Fallacy**: retrieved evidence from different semantic aspects is aggregated in a naive, unstructured manner, without modeling their inter-dependencies and potential conflicts, which ultimately undermines downstream reasoning. To address these challenges, we propose **Med-SRAF**, a multi-agent retrieval augmentation framework guided by medical domain knowledge. This framework reconstructs the traditional RAG process through two core mechanisms: (1) Intent-driven Semantic Routing, where a UMLS-based NavigationAgent dynamically maps queries to medical dimensions for strategic search space pruning; and (2) Evidence-based Agentic Fusion, where a FusionAgent resolves conflicts among dimension-specific evidence to build logically consistent reasoning chains. Extensive experiments on five widely used medical benchmarks show that Med-SRAF consistently outperforms existing general RAG baselines, achieving an average accuracy improvement of over **4.9%**, highlighting its effectiveness in robust and interpretable medical reasoning. Our code is at https://anonymous.4open.science/r/MultiAgent_RAG-F6DC.
Enrich, Aggregate, and Generate: Three-stage Biomedical Data-to-Text Generation Using Large Language Models in Low-resource Scenarios
PDF ↗Biomedical data-to-text generation aims at generating textual natural language descriptions that can fluently and precisely describe the biomedical structured data. However, biomedical data-to-text generation faces the dilemma of a lack of labeled data due to the privacy and scarcity of medical data. Large language models (LLMs) have demonstrated the ability to solve few-shot tasks through in-context learning (ICL). In this paper, we are the first to explore the performance of different LLMs in the biomedical data-to-text generation task.To address the issues of semantic sparsity and misinterpretation of numerical values in biomedical structured data, we propose an EAG (Enrich, Aggregate, and Generate) framework, a simple but efficient LLM-based three-stage biomedical D2T approach in low-resource scenarios. We conduct extensive evaluations of closed-source general LLMs, open-source general LLMs, and open-source medical LLMs. The results show that the EAG framework provides good interpretability and superior performance, achieving state-of-the-art performance on the BioLeaflets dataset. The code and data will be released at https://github.com/FXLP/EAG.
Despite the pivotal role of numerical reasoning as the cornerstone of mathematical capabilities in large language models (LLMs) across applications, few benchmarks evaluate LLMs by integrating numerical processing and mathematical reasoning, hindering the interpretability of failures in math tasks. We introduce PyraMathBench, a comprehensive hierarchical benchmark with 27,215 questions derived from 7,404 math word problems, spanning 4 key cognitive aspects, 14 subcategories, and 2 modalities. Experiments reveal that LLMs’ performance is severely compromised by inadequate numerical computation and weak handling of abstract numerical questions. To address this, we propose the Smart Optimization Learning-based VErsatile module (SOLVE) and Interactive Relative Policy Optimization (IRPO), which enhance LLMs’ numerical-mathematical synergy via efficient tool calls (fuzzy matching and low-quality call rejection). Comparative experiments show Qwen-2.5 achieves a 5.0 score improvement with SOLVE and IRPO training.
Interpretability from the Ground Up: Stakeholder-Centric Design of Automated Scoring in Educational Assessments
PDF ↗AI-driven automated scoring systems offer scalable and efficient means of evaluating complex student-generated responses. Yet, despite increasing demand for transparency and interpretability, the field has yet to develop a widely accepted solution for interpretable automated scoring to be used in large-scale real-world assessments. This work takes a principled approach to address this challenge. We analyze the needs and potential benefits of interpretable automated scoring for various assessment stakeholder groups and develop four principles of interpretability – (F)aithfulness, (G)roundedness, (T)raceability, and (I)nterchangeability (FGTI) – targeted at those needs. To illustrate the feasibility of implementing these principles, we develop the AnalyticScore framework as a baseline reference framework. When applied to the domain of text-based constructed-response scoring, AnalyticScore outperforms many uninterpretable scoring methods in terms of scoring accuracy and is, on average, within 0.06 QWK of the uninterpretable SOTA across 10 items from the ASAP-SAS dataset. By comparing against human annotators conducting the same featurization task, we further demonstrate that the featurization behavior of AnalyticScore aligns well with that of humans.
ToolSafe: Enhancing Tool Invocation Safety of LLM-based agents via Proactive Step-level Guardrail and Feedback
PDF ↗While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains underexplored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action–attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, We introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65% on average and improves benign task completion by approximately 10% under prompt injection attacks.
Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization
PDF ↗Hallucination in Large Language Models (LLMs)—characterized by the generation of content inconsistent with contextual facts or logical constraints—remains a persistent challenge for reliable deployment. In this work, we address this issue through a geometric framework rooted in the linear representation hypothesis. We propose that hallucinations manifest as orthogonal noise relative to the semantic manifold of the residual stream. Specifically, we hypothesize that while attention heads ideally propagate information congruent with the context subspace, hallucinations arise when specific heads introduce components orthogonal to this subspace, disrupting the coherence of the latent representation. Based on this formulation, we introduce Dynamic Contextual Orthogonalization (DCO), an inference-time intervention method. DCO utilizes the input residual stream as a dynamic context anchor to perform orthogonal decomposition on attention head outputs. To distinguish between context-aligned semantic updates and divergent noise, DCO employs a layer-wise Z-score suppression mechanism that selectively attenuates outlier orthogonal components based on statistical distributions. Evaluations on Llama-3-8B and 70B across benchmarks such as XSum, NQ-Swap, and IFEval demonstrate that DCO achieves superior contextual faithfulness compared to state-of-the-art intervention baselines. Furthermore, DCO maintains high performance on knowledge-intensive tasks like TriviaQA and TruthfulQA, effectively mitigating the trade-off between hallucination suppression and parametric knowledge retention often observed in existing methods. Our findings validate the geometric interpretation of hallucinations and establish DCO as a computationally efficient approach for enforcing manifold alignment.Our code is available at https://anonymous.4open.science/r/DCO-4AB0
Emergency response is a safety-critical public governance task that demands accurate and timely decision-making based on complex event information. This process involves multiple stages, including information collection, integration, analysis, risk assessment, and decision recommendation. Existing research has predominantly concentrated on the earlier stages, while studies focusing on the decision support phase remain underexplored, primarily due to the lack of suitable datasets for reliable and compliance-aware decision-oriented modeling and evaluation. To bridge this gap, we introduce the first real-world Emergency Decision-Making dataset EDM-Bench, comprising 1,179 instances spanning diverse task formats, including judgment, choice, short-answer, and structured emergency report generation. We also construct a structured rule repository, EDM-R², which contains 3,406 parsed emergency regulations to enhance decision reliability. Building on these resources, we propose a rule-enhanced reasoning framework, R³V-EDM, which integrates external regulatory knowledge with constrained inference mechanisms to improve both decision safety and interpretability. Extensive experiments demonstrate the inherent complexity of emergency decision-making and validate the effectiveness of our approach in enabling more reliable and trustworthy decisions.
Recent text-to-video (T2V) diffusion models have made remarkable progress in generating high-quality videos. However, they often struggle to align with complex text prompts, particularly when multiple objects, attributes, or spatial relations are specified. We introduce VideoRepair, the first self-correcting, training-free, and model-agnostic video refinement framework that automatically detects fine- grained text–video misalignments and performs targeted, localized corrections. Our key insight is that even misaligned videos usually contain correctly generated regions that should be preserved rather than regenerated. Building on this observation, VideoRepair proposes a novel region-preserving refinement strategy with three stages: (i) misalignment detection, where MLLM-based evaluation with automatically generated evaluation questions identifies misaligned regions; (ii) refinement planning, which preserves correctly generated entities, segments their regions across frames, and constructs targeted prompts for misaligned areas; and (iii) localized refinement, which selectively regenerates problematic regions while preserving faithful content through joint optimization of preserved and newly generated areas. On two benchmarks, EvalCrafter and T2V-CompBench with four recent T2V backbones, VideoRepair achieves substantial improvements over recent baselines across diverse alignment metrics. Comprehensive ablations further demonstrate the efficiency, robustness, and interpretability of our framework.
Recent work has increasingly explored neuron-level interpretation in vision-language models (VLMs) to identify neurons critical to final predictions. However, existing neuron analyses generally focus on single tasks, limiting the comparability of neuron importance across tasks. Moreover, ranking strategies tend to score neurons in isolation, overlooking how task-dependent information pathways shape the write-in effects of feed-forward network (FFN) neurons. This oversight can exacerbate neuron polysemanticity in multi-task settings, introducing noise into the identification and intervention of task-critical neurons. In this study, we propose HONES (**H**ead-**O**riented **N**euron **E**xplanation **S**teering), a gradient-free framework for task-aware neuron attribution and steering in multi-task VLMs. HONES ranks FFN neurons by their causal write-in contributions conditioned on task-relevant attention heads, and further modulates salient neurons via lightweight scaling. Experiments on four diverse multimodal tasks and two popular VLMs show that HONES outperforms existing methods in identifying task-critical neurons and improves model performance after steering. Our source code is released at: https://github.com/petergit1/HONES.
Retrieval-Augmented Generation (RAG) remains unreliable in specialized domains due to semantic and lexical mismatch between lay queries and professional terminology, and existing generative expansion often introduces redundancy or hallucinations that cause semantic drift. We propose Generative Query Condensation (GQC), a query rewriting strategy that reframes rewriting as semantic condensation rather than expansion. To operationalize GQC, we introduce Query-to-Entity Inference (Q2EI), an entity-centric rewriting method that realizes semantic condensation through explicit inference of the underlying target entity. By moving semantic alignment from retrieval-time vector matching to the rewriting stage, Q2EI produces information-dense query representations. Experimental results on medical and legal benchmarks show that Q2EI consistently outperforms strong baselines across retrievers, improving retrieval effectiveness while substantially reducing rewriting token consumption compared to generative expansion methods. Further analysis confirms that these gains primarily arise from accurate entity inference, and that Q2EI’s semantic condensation design limits error amplification when inference is imperfect, leading to more stable and interpretable retrieval behavior.
Save the Good Prefix: Precise Error Penalization via Process-Supervised RL to Enhance LLM Reasoning
PDF ↗Reinforcement learning (RL) has emerged as a powerful framework for improving the reasoning capabilities of large language models (LLMs). However, most existing RL approaches rely on sparse outcome rewards, which fail to credit correct intermediate steps in partially successful solutions. Process reward models (PRMs) offer fine-grained step-level supervision, but their scores are often noisy and difficult to evaluate. As a result, recent PRM benchmarks focus on a more objective capability: detecting the first incorrect step in a reasoning path. However, this evaluation target is misaligned with how PRMs are typically used in RL, where their step-wise scores are treated as raw rewards to maximize. To bridge this gap, we propose Verifiable Prefix Policy Optimization (VPPO), which uses PRMs only to localize the first error during RL. Given an incorrect rollout, VPPO partitions the trajectory into a verified correct prefix and an erroneous suffix based on the first error, rewarding the former while applying targeted penalties only after the detected mistake. This design yields stable, interpretable learning signals and improves credit assignment. Across multiple reasoning benchmarks, VPPO consistently outperforms sparse-reward RL and prior PRM-guided baselines on both Pass@1 and Pass@K.
Charts are widely used to present complex information. Deriving meaningful insights in real-world contexts often requires interpreting multiple related charts together. Research on understanding multi-chart images has not been extensively explored. We introduce PolyChartQA, a mid-scale dataset specifically designed for question answering over multi-chart images. PolyChartQA comprises 534 multi-chart images (with a total of 2,297 sub-charts) sourced from peer-reviewed computer science research publications and 2,694 QA pairs. We evaluate the performance of nine state-of-the-art Multimodal Language Models (MLMs) on PolyChartQA across question type, difficulty, question source, and key structural characteristics of multi-charts. Our results show a 27.4% LLM-based accuracy (L-Accuracy) drop on human-authored questions compared to MLM-generated questions, and a 5.39% L-accuracy gain with our proposed prompting method.
Uncovering hidden symbolic laws from time series data, as an aspiration dating back to Kepler’s discovery of planetary motion, remains a core challenge in scientific discovery and artificial intelligence. While Large Language Models show promise in structured reasoning tasks, their ability to infer interpretable, context-aligned symbolic structures from time series data is still underexplored. To systematically evaluate this capability, we introduce SymbolBench, a comprehensive benchmark designed to assess symbolic reasoning over real-world time series across three tasks: multivariate symbolic regression, Boolean network inference, and causal discovery. Unlike prior efforts limited to simple algebraic equations, SymbolBench spans a diverse set of symbolic forms with varying complexity. We further propose a unified framework that integrates LLMs with genetic programming to form a closed-loop symbolic reasoning system. Our empirical results reveal key strengths and limitations of current models, highlighting the importance of combining domain knowledge, context alignment, and reasoning structure to improve LLMs in automated scientific discovery.
Despite demonstrating remarkable performance across a wide range of tasks, large language models (LLMs) have also been found to frequently produce outputs that are incomplete or selectively omit key information. In sensitive domains, such omissions can result in significant harm comparable to that posed by factual inaccuracies, including hallucinations. In this study, we address the challenge of evaluating the comprehensiveness of LLM-generated texts, focusing on the detection of missing information or underrepresented viewpoints. We investigate three automated evaluation metrics: (1) an NLI-based method that decomposes texts into atomic statements and uses natural language inference (NLI) to identify missing facts, (2) a Q A-based metric that extracts question-answer pairs and compares responses across sources, and (3) an end-to-end approach that directly identifies missing content using LLMs. Our experiments demonstrate the surprising effectiveness of the simple end-to-end metric compared to more complex metrics, though at the cost of reduced robustness, interpretability and result granularity. We further assess the comprehensiveness of responses from several popular open-weight LLMs when answering user queries based on multiple sources.
Time series captioning, the task of describing time series in natural language, requires numeric and temporal reasoning, trend interpretation, and contextual understanding. Existing benchmarks, however, often rely on fully synthetic or generic captions, and typically neglect metadata and visual representations. We introduce CaTS-Bench, a comprehensive benchmark for Context-aware Time Series reasoning across 11 diverse domains, centered on a gold-standard evaluation set of 1746 human-rewritten captions that measure how effectively models translate numeric trends into immediately interpretable narratives. To address the scarcity of human-annotated data, we also propose a scalable pipeline for generating high-fidelity synthetic captions, the quality of which we validate. We evaluate leading Vision-Language Models on our benchmark, revealing that even proprietary models struggle to capture numeric nuances in temporal descriptions, while finetuning open-source models on synthetic data yields substantial performance gains. Finally, we release a diagnostic suite of 910 multiple-choice questions and use tailored numeric metrics to gauge time-series-specific reasoning capabilities, establishing CaTS-Bench as a reliable foundation for grounded, multimodal text generation in numeric domains.
Neuronal Insights into LLM Attacks: Targeted Neuron Tuning for Precise and Robust Vulnerability Patching
PDF ↗Despite recent advances in safety alignment, large language models (LLMs) remain highly susceptible to adversarial attacks, while the internal mechanisms behind such vulnerabilities are still poorly understood. Existing gradient-based attribution methods offer valuable interpretability for analyzing information storage and processing in LLMs. However, they are inapplicable to adversarial attacks, which typically occur in open-ended generation settings without fixed ground-truth outputs. To address these challenges, we propose a novel similarity-based gradient attribution method to identify key neurons sensitive to adversarial behaviors in open-ended generation tasks. The detected neurons, termed targeted neurons, play a critical role in safety training. Building on this neuron-level perspective, we uncover two key neuronal patterns: (i) universal neurons that are consistently exploited across multiple attack strategies, and (ii) interference neurons that hinder safety improvements when fine-tuned indiscriminately, providing mechanistic insights into the interpretability of adversarial vulnerabilities. Inspired by these findings, we propose a neuron-level defense strategy, Targeted Neuron Tuning (TNT), which selectively fine-tunes the identified targeted neurons for specific attacks. Experimental evaluations across multiple LLM architectures and scales demonstrate that TNT substantially improves model robustness against a wide range of jailbreak attacks, achieving safe rates exceeding 90% and even approaching 100%, while preserving general task performance, enabling precise and robust safety interventions. Warning: This paper contains example data that may be harmful.
Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is crucial for improving their real-world behavior. A common approach is to use reward models that enable reinforcement-learning post-training. However, traditional reward modeling requires finetuning on large preference datasets, limiting adaptability to new preferences. We introduce Activation Reward Models (Activation RMs)—the first mechanistic interpretability approach that steers LLM activations to align with few-shot preference data without finetuning. Our method combines activation denoising and output token likelihood scoring, achieving state-of-the-art performance on standard reward modeling benchmarks, surpassing zero-shot, few-shot, and voting-based baselines. We further demonstrate that Activation RMs mitigate reward hacking behaviors and remain robust to noisy exemplars and spurious reward signals. To evaluate this, we propose PreferenceHack, a novel few-shot benchmark testing reward models on reward hacking in a paired preference format, where Activation RMs achieve state-of-the-art performance, surpassing GPT-4o.