Reinforcement learning (RL) has shown promise for enhancing code generation capabilities in large language models (LLMs), yet its effectiveness critically depends on high-quality test suites for reliable reward signals. Current approaches suffer from inadequate test case quantity and quality, leading to false positives (incorrect solutions passing verification) and slow positives (valid but suboptimal implementations), which corrupt RL training dynamics. We address these challenges through three key contributions: (1) We systematically analyze how low-quality test suites degrade Code RL performance via reward misalignment; (2) We propose Themis, an automated framework that transforms test case generation into code synthesis—first extracting problem constraints via template-guided parsing, then generating executable test generators through LLM-powered code synthesis, and finally validating tests through constraint-aware filtering; (3) We develop an error-guided test case reduction method that preserves error detection efficacy while reducing test set cardinality, thereby enhancing reinforcement learning training efficiency. Evaluated on programming competition datasets, Themis achieves 95 percent error detection rates, outperforming original test suites in most of the cases. When integrated into RL pipelines, models trained with Themis-generated tests demonstrate consistent 3-5 percent improvements across HumanEval, MBPP, and LiveCodeBench compared to the baseline, matching performance levels achieved with manually curated test suites. Our constraint-aware test synthesis framework ensures full automation while preserving semantic validity—critical for scaling RL training to complex code generation tasks. The framework's modular design also enables seamless integration with existing code data synthesis frameworks.
论文检索
输入标题、作者或关键词，从 11,272 篇学术成果中精准定位
Seeing through the Conflict: Transparent Knowledge Conflict Handling in Retrieval-Augmented Generation
PDF ↗Large language models (LLMs) equipped with retrieval—the Retrieval-Augmented Generation (RAG) paradigm—should combine their parametric knowledge with external evidence, yet in practice they often hallucinate, over-trust noisy snippets, or ignore vital context. We introduce TCR (Transparent Conflict Resolution), a plug-and-play framework that makes this decision process observable and controllable. TCR (i) disentangles semantic match and factual consistency via dual contrastive encoders, (ii) estimates self-answerability to gauge confidence in internal memory, and (iii) feeds the three scalar signals to the generator through a lightweight soft-prompt with SNR-based weighting. Across seven benchmarks TCR improves conflict detection (+5–18 F₁), raises knowledge-gap recovery by +21.4 percentage points and cuts misleading-context overrides by –29.3 percentage points, while adding only 0.3% parameters. The signals align with human judgements and expose temporal decision patterns.
Interpreting Fedspeak with Confidence: A LLM-Based Uncertainty-Aware Framework Guided by Monetary Policy Transmission Paths
PDF ↗"Fedspeak", the stylized and often nuanced language used by the U.S. Federal Reserve, encodes implicit policy signals and strategic stances. The Federal Open Market Committee strategically employs Fedspeak as a communication tool to shape market expectations and influence both domestic and global economic conditions. As such, automatically parsing and interpreting Fedspeak presents a high-impact challenge, with significant implications for financial forecasting, algorithmic trading, and data-driven policy analysis. Technically, to enrich the semantic and contextual representation of Fedspeak texts, we incorporate domain-specific reasoning grounded in the monetary policy transmission mechanism. We further introduce a dynamic uncertainty decoding module to assess the confidence of model predictions, thereby enhancing both classification accuracy and model reliability. Experimental results demonstrate that our framework achieves state-of-the-art performance on the policy stance analysis task. Moreover, statistical analysis reveals a significant positive correlation between perceptual uncertainty and model error rates, validating the effectiveness of perceptual uncertainty as a diagnostic signal.
Through reinforcement learning (RL) with outcome correctness rewards, large reasoning models (LRMs) with scaled inference computation have demonstrated substantial success on complex reasoning tasks. However, the one-sided reward, focused solely on final correctness, limits its ability to provide detailed supervision over internal reasoning process. This deficiency leads to suboptimal internal reasoning quality, manifesting as issues like over-thinking, under-thinking, redundant-thinking, and disordered-thinking. Inspired by the recent progress in LRM self-rewarding, we introduce self-rewriting framework, where a model rewrites its own reasoning texts, and subsequently learns from the rewritten reasoning to improve the internal thought process quality. For algorithm design, we propose a selective rewriting approach wherein only "simple" samples, defined by the model's consistent correctness, are rewritten, thereby preserving all original reward signals of GRPO. For practical implementation, we compile rewriting and vanilla generation within one single batch, maintaining the scalability of the RL algorithm and introducing only 10% overhead. Extensive experiments on diverse tasks with different model sizes validate the effectiveness of self-rewriting. In terms of the accuracy-length tradeoff, the self-rewriting approach achieves improved accuracy (+0.6) with substantially shorter reasoning (-46%) even without explicit instructions in rewriting prompts to reduce reasoning length, outperforming existing strong baselines. In terms of internal reasoning quality, self-rewriting achieves significantly higher scores (+7.2) under the LLM-as-a-judge metric, successfully mitigating internal reasoning flaws.
Vision Transformer (ViT) has become one of the cornerstones of the computer vision field, demonstrating exceptional performance. However, its inherent high computational complexity and inference latency still pose significant obstacles for deployment in resource-constrained environments. Token pruning, by removing less informative tokens, offers an effective strategy to reduce computational overhead. However, existing pruning methods largely rely on static or local token importance scores. This myopic approach fundamentally overlooks the sequential dependency of pruning decisions and fails to capture the interaction effects between pruning decisions across layers, often neglecting the global interactions between mask variables. To address this limitation, we propose V-Pruner, a fast and globally-informed token pruning framework for Vision Transformer. V-Pruner first leverages Fisher information to perform an initial assessment of token importance, providing a principled initial prior for pruning decisions. Building on this, V-Pruner introduces a Reinforcement Learning (RL) Proximal Policy Optimization (PPO) algorithm, refining token pruning into a global sequential decision process. The algorithm combines a composite reward signal that incorporates both model performance and computational cost to guide policy exploration, effectively evaluating the long-term impact of different pruning decision combinations on global model performance. Extensive experiments on ViT-L, DeiT-B, DeiT-S, and DeiT-T demonstrate that V-Pruner achieves a better balance between accuracy, GFLOPs, inference speed, and training time, surpassing existing mainstream ViT pruning algorithms in overall performance.
Learning diagnosis is a critical task that monitors students' cognitive state during educational activities, with the goal of enhancing learning outcomes. With advancements in language models (LMs), many AI-driven educational studies have shifted towards conversational learning scenarios, where students engage in multi-turn interactive dialogues with tutors. However, conversational learning diagnosis remains underdeveloped, and most existing techniques acquire students' cognitive state through intuitive instructional prompts on LMs to analyze the dialogue text. This direct prompting approach lacks a solid psychological foundation and fails to ensure the reliability of the generated analytical text. In this study, we introduce ParLD, a preview-analyze-reason framework for conversational learning diagnosis, which leverages multi-agent collaboration to diagnose students' cognitive state over multiple dialogue turns. Specifically, ParLD comprises main components: (1) Behavior Previewer, which generates a student behavior schema based on previous states and learning content; (2) State Analyzer, which diagnose the tutor-student dialogue and behavior schema to update the cognitive state; and (3) Performance Reasoner, which predicts the student's future responses and provides verifiable feedback to support ParLD's self-reflection with the Chain Reflector. They operate sequentially and iteratively during each interaction turn to diagnose the student’s cognitive state. We conduct experiments to evaluate both performance prediction and tutoring support, emphasizing the effectiveness of ParLD in providing reliable and insightful learning diagnosis.
ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation
PDF ↗Training large language models (LLMs) with synthetic reasoning data has become a popular approach to enhancing their reasoning capabilities, while a key factor influencing the effectiveness of this paradigm is the quality of the generated multi-step reasoning data. To generate high-quality reasoning data, many recent methods generate synthetic reasoning paths and filter them based on final answer correctness, often overlooking flaws in intermediate reasoning steps. To enhance the verification of intermediate reasoning steps, prior work primarily resorts to code execution or symbolic reasoning engines. However, code-based validation is restricted to code or mathematical tasks, and reasoning engines require a well-structured and complete context. As a result, existing methods fail to function effectively in natural language reasoning tasks that involve ambiguous or incomplete contexts. In these tasks, synthetic data still lack reliable checks for verifying each reasoning step. To address this challenge, we introduce ORACLE, a structured data generation framework inspired by syllogistic reasoning. ORACLE integrates the generative strengths of LLMs with symbolic supervision: the LLM produces step-wise reasoning contexts, while a symbolic reasoning engine verifies the validity of each intermediate step. By employing a unified prompting template to elicit modular reasoning chains, ORACLE enables fine-grained, step-level validation, facilitating the construction of high-quality multi-step reasoning data. Across six logical, factual, and commonsense reasoning benchmarks, our ORACLE consistently outperforms strong baselines on multiple models.
Multimodal DeepResearcher: Generating Text-Chart Interleaved Reports from Scratch with Agentic Framework
PDF ↗Visualizations play a crucial part in effective communication of concepts and information. Recent advances in reasoning and retrieval augmented generation have enabled Large Language Models (LLMs) to perform deep research and generate comprehensive reports. Despite its progress, existing deep research frameworks primarily focus on generating text-only content, leaving the automated generation of interleaved texts and visualizations underexplored. This novel task poses key challenges in designing informative visualizations and effectively integrating them with text reports. To address these challenges, we propose Formal Description of Visualization (FDV), a structured textual representation of charts that enables LLMs to learn from and generate diverse, high-quality visualizations. Building on this representation, we introduce Multimodal DeepResearcher, an agentic framework that decomposes the task into four stages: (1) researching, (2) exemplar report textualization, (3) planning and (4) multimodal report generation. For the evaluation of the generated reports, we develop MultimodalReportBench which contains 100 diverse topics as inputs, and a set of dedicated metrics for report and chart evaluation. Extensive experiments across models and evaluation methods demonstrate the effectiveness of Multimodal DeepResearcher. Notably, utilizing the same Claude 3.7 Sonnet model, Multimodal DeepResearcher achieves an 82% overall win rate over the baseline method.
This work proposes a grammar-based chunking strategy that segments input streams into semantically complete units by parsing dependency relations (e.g., noun phrase boundaries, verb-object structures) and punctuation features. The method ensures chunk coherence and minimizes semantic fragmentation. Building on this mechanism, we present SASST (Syntax-Aware Simultaneous Translation), an end-to-end framework integrating frozen Whisper encoder and decoder-only LLM. The unified architecture dynamically outputs translation tokens or symbols to jointly optimize translation timing and content, with target-side reordering addressing word-order divergence. Experiments on CoVoST2 multilingual corpus (En to De/Zh/Ja) demonstrate significant translation quality improvements across languages, validating the effectiveness of syntactic structures in LLM-driven SimulST systems.
Data selection for instruction tuning is crucial for improving the performance of large language models (LLMs) while reducing training costs. In this paper, we propose Refined Contribution Measurement with In-Context Learning (RICo), a novel gradient-free method that quantifies the fine-grained contribution of individual samples to both task-level and global-level model performance. RICo enables more accurate identification of high-contribution data, leading to better instruction tuning. We also introduce a lightweight selection paradigm trained on RICo scores, enabling scalable data selection with strictly linear inference complexity. Extensive experiments on 3 LLMs across 12 benchmarks and 5 pairwise evaluation sets demonstrate the effectiveness of RICo. Remarkably, on LLaMA3.1-8B, models trained in 15% of RICo-selected data outperform full datasets by 5.42 percentage points and exceed the best performance of widely used selection methods by 1.48 percentage points. We further analyze high-contribution samples selected by RICo, which show both diverse tasks and appropriate difficulty levels, rather than merely the most difficult cases.
EHRStruct: A Comprehensive Benchmark Framework for Evaluating Large Language Models on Structured Electronic Health Record Tasks
PDF ↗Structured Electronic Health Record (EHR) data stores patient information in relational tables and plays a central role in clinical decision-making. Recent advances have explored the use of large language models (LLMs) to process such data, showing promise across various clinical tasks. However, the absence of standardized evaluation frameworks and clearly defined tasks makes it difficult to systematically assess and compare LLM performance on structured EHR data. To address these evaluation challenges, we introduce EHRStruct, a benchmark specifically designed to evaluate LLMs on structured EHR tasks. EHRStruct defines 11 representative tasks spanning diverse clinical needs and includes 2,200 task-specific evaluation samples derived from two widely used EHR datasets. We use EHRStruct to evaluate 20 advanced and representative LLMs, covering both general and medical models. We further analyze key factors influencing model performance, including input formats, few-shot generalisation, and finetuning strategies, and compare results with 11 state-of-the-art LLM-based enhancement methods for structured data reasoning. Our results indicate that many structured EHR tasks place high demands on the understanding and reasoning capabilities of LLMs. In response, we propose SEMaster, a code-augmented method that achieves state-of-the-art performance and offers practical insights to guide future research.
While large language models (LLMs) have demonstrated strong capabilities in code generation, current benchmarks primarily focus on single-turn scenarios, neglecting the complexity of multi-turn interactions and user diversity. To address this gap, we introduce Talk2Code, the first benchmark for user-stratified multi-turn dialogue code generation evaluation across algorithmic problem-solving and backend programming tasks.A distinctive feature of our benchmark is its user-stratified interaction modeling. For identical coding tasks, we construct dialogue trajectories tailored for novice, intermediate, and expert users, capturing their distinct expectations and communication patterns.To facilitate comprehensive evaluation, we propose a multi-dimensional evaluation framework assessing both code quality and interaction experience through a novel Dual-track Evaluation Method. In the Direct Generation Track, the benchmark provides golden dialogue context (excluding the final code) directly to the LLM for code generation. In contrast, the Interactive Dialogue Track simulates realistic multi-turn interactions, prompting the model to proactively clarify instructions and gather requirements before generating solutions. Code quality is evaluated in both tracks by Test Pass Rate and Success Rate, while interaction experience is assessed exclusively within the Interactive Dialogue Track through subjective and alignment indicators. Our benchmark and multi-dimensional indicator system collectively establish a new paradigm for evaluating adaptive, user-aware AI coding assistants.
Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures. We observe that perturbations orthogonal to the alignment direction—defined by weight differences between aligned (safe) and unaligned models—rapidly compromise model safety. In contrast, updates along the alignment direction largely preserve it, revealing the parameter space as a "narrow safety basin". To address this, we propose AsFT (Anchoring Safety in Fine-Tuning) to maintain safety by explicitly constraining update directions during fine-tuning. By penalizing updates orthogonal to the alignment direction, AsFT effectively constrains the model within the "narrow safety basin," thus preserving its inherent safety. Extensive experiments on multiple datasets and models show that AsFT reduces harmful behaviors by up to 7.60%, improves task performance by 3.44%, and consistently outperforms existing methods across multiple tasks.
RegionMarker: A Region-Triggered Semantic Watermarking Framework for Embedding-as-a-Service Copyright Protection
PDF ↗Embedding-as-a-Service (EaaS) is an effective and convenient deployment solution for addressing various NLP tasks. Nevertheless, recent research has shown that EaaS is vulnerable to model extraction attacks, which could lead to significant economic losses for model providers. For copyright protection, existing methods inject watermark embeddings into text embeddings and use them to detect copyright infringement. However, current watermarking methods often resist only a subset of attacks and fail to provide comprehensive protection. To this end, we present the region-triggered semantic watermarking framework called RegionMarker, which defines trigger regions within a low-dimensional space and injects watermarks into text embeddings associated with these regions. By utilizing a secret dimensionality reduction matrix to project onto this subspace and randomly selecting trigger regions, RegionMarker makes it difficult for watermark removal attacks to evade detection. Furthermore, by embedding watermarks across the entire trigger region and using the text embedding as the watermark, RegionMarker is resilient to both paraphrasing and dimension-perturbation attacks. Extensive experiments on various datasets show that RegionMarker is effective in resisting different attack methods, thereby protecting the copyright of EaaS.
Multi-character role-playing aims to equip models with the capability to simulate diverse roles. Existing methods either use one shared parameterized module across all roles or assign a separate parameterized module to each role. However, the role-shared module may ignore distinct traits of each role, weakening personality learning, while the role-specific module may overlook shared traits across multiple roles, hindering commonality modeling. In this paper, we propose a novel HyCoRA: Hyper-Contrastive Role-Adaptive learning framework, which efficiently improves multi-character role-playing agents' ability by balancing the learning of distinct and shared traits. Specifically, we propose a Hyper-Half Low-Rank Adaptation structure, where one half is a role-specific module generated by a lightweight hyper-network, and the other half is a trainable role-shared module. The role-specific module is devised to represent distinct persona signatures, while the role-shared module serves to capture common traits. Moreover, to better reflect distinct personalities across different roles, we design a hyper-contrastive learning mechanism to help the hyper-network distinguish their unique characteristics. Extensive experimental results on both English and Chinese available benchmarks demonstrate the superiority of our framework. Further GPT-4 evaluations and visual analyses also verify the capability of HyCoRA to capture role characteristics.
Multimodal retrieval-augmented generation (RAG) systems enhance large vision-language models by integrating cross-modal knowledge, enabling their increasing adoption across real-world multimodal tasks. These knowledge databases may contain sensitive information that requires privacy protection. However, multimodal RAG systems inherently grant external users indirect access to such data, making them potentially vulnerable to privacy attacks, particularly membership inference attacks (MIAs). Existing MIA methods targeting RAG systems predominantly focus on the textual modality, while the visual modality remains relatively underexplored. To bridge this gap, we propose MrM, the first black-box MIA framework targeted at multimodal RAG systems. It utilizes a multi-object data perturbation framework constrained by counterfactual attacks, which can concurrently induce the RAG systems to retrieve the target data and generate information that leaks the membership information. Our method first employs an object-aware data perturbation method to constrain the perturbation to key semantics and ensure successful retrieval. Building on this, we design a counterfact-informed mask selection strategy to prioritize the most informative masked regions, aiming to eliminate the interference of model self-knowledge and amplify attack efficacy. Finally, we perform statistical membership inference by modeling query trials to extract features that reflect the reconstruction of masked semantics from response patterns. Experiments on two visual datasets and eight mainstream commercial visual-language models (e.g., GPT-4o, Gemini-2) demonstrate that MrM achieves consistently strong performance across both sample-level and set-level evaluations, and remains robust under adaptive defenses.
Open knowledge bases (e.g., websites) are widely adopted in Retrieval-Augmented Generation (RAG) systems to provide supplementary knowledge (e.g., latest information). However, such sources inevitably contain biased or harmful content, and incorporating these untrusted contents into the RAG process introduces significant safety risks, including the degradation of LLM performance and the potential generation of harmful outputs. Recent studies have shown that this vulnerability can be further amplified by adversarial poisoning attacks specifically targeting the knowledge sources. Most existing methods primarily emphasize improving the accuracy and efficiency of RAG systems, usually overlooking these critical safety concerns. In this paper, we propose a safety-aware retrieval framework (ShieldRAG) designed to augment language model generation by jointly optimizing for both relevance and safety in the retrieved knowledge content. The core idea of ShieldRAG is to transfer the safety knowledge implicitly encoded in powerful LLMs into the retriever model through an adversarial knowledge alignment mechanism. This can empower the retriever with the safety awareness, and adapt to the diverse and unknown distribution of unsafe content encountered in practical scenarios. We evaluate ShieldRAG on seven real-world datasets using five widely-used LLMs and two state-of-the-art poisoning attack strategies. Experimental results show that our method substantially improves the robustness of RAG systems against unsafe knowledge sources, while maintaining competitive performance in terms of generation accuracy and efficiency.
Long Chain-of-Thought (CoT) reasoning has shown great promise in complex reasoning tasks, but its application to medical decision-making presents unique challenges. Unlike structured tasks relying on static verification frameworks, medical decision-making requires dynamic validation through longitudinal clinical outcomes, exhibiting temporal-causal dependencies that complicate the verification of reasoning processes. Therefore, we introduce a novel data construction framework specifically designed for medical decision-making. First, the framework analyzes real-world clinical cases to construct a timeline of medical events and identify critical decision points, including examination, diagnosis, and treatment. Subsequently, it employs a clinical causality-aware strategy to generate decision-making questions at the identified points, along with reasoning traces and corresponding answers. Finally, information drawn from future nodes serves as clinical logic-constrained criteria to re-evaluate and refine the soundness of the generated reasoning and responses. Building on this, we present OncoCoT, an oncologic decision-making dataset derived from clinical records over the past four years across eight common cancer types. Furthermore, we distill a subset of OncoCoT into a dedicated benchmark, OncoEval, to facilitate systematic evaluation of clinical reasoning capabilities in LLMs. Evaluation results show that existing state-of-the-art reasoning models, such as Deepseek-r1 and GPT-o3, exhibit limited capability in addressing clinical problems in OncoEval, highlighting the need for further improvement.
While large language models (LLMs) leverage both knowledge and reasoning during inference, the capacity to distinguish between them plays a pivotal role in model analysis, interpretability, and development. Inspired by dual-system cognitive theory, we propose a cognition attribution framework to decouple the contribution of knowledge and reasoning. In particular, the cognition of LLMs is decomposed into two distinct yet complementary phases: knowledge retrieval (Phase 1) and reasoning adjustment (Phase 2). To separate these phases, LLMs are prompted to generate answers under two different cognitive modes, fast thinking and slow thinking, respectively. The performance under different cognitive modes is analyzed to quantify the contribution of knowledge and reasoning. This architecture is employed to 15 LLMs across 3 datasets. Results reveal: (1) reasoning adjustment is domain-specific, benefiting reasoning-intensive domains (e.g., mathematics, physics, and chemistry) and potentially imparing knowledge-intensive domains. (2) Parameter scaling improves both knowledge and reasoning, with knowledge improvements being more pronounced. Additionally, parameter scaling make LLMs reasoning significantly more prudent, while moderately more intelligent. (3) Knowledge primarily resides in lower network layers, while reasoning operates in higher layers. Our framework not only helps understand LLMs from a "decoupling" perspective, but also provides new insights into existing research, including scaling laws, hierarchical knowledge editing, and limitations of small-scale-LLM reasoning.
RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior
PDF ↗Large Language Models (LLMs) with long chain-of-thought (CoT) capability, termed Reasoning Models, demonstrate superior intricate problem-solving abilities through multi-step long CoT reasoning. To create a dual-capability model with long CoT capability and domain-specific knowledge without substantial computational and data costs, model merging emerges as a highly resource-efficient method. However, significant challenges lie in merging domain-specific LLMs with long CoT ones since nowadays merging methods suffer from reasoning capability degradation, even gibberish output and output collapse. To overcome this, we introduce RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior, a novel merging framework designed to integrate domain-specific LLMs with long CoT capability, meanwhile maintaining model performance in the original domain. Treating reasoning model weights as foundational prior, our method utilizes a reasoning capability indicator to preserve core long CoT capability model weights while selectively merging essential domain-specific weights. We conducted extensive experiments on Qwen2.5-7B, Llama3.1-8B, and Qwen2.5-1.5B models in BioMedicine and Finance domains. Our results show that RCP-Merging successfully merges a reasoning model with domain-specific ones, improving domain task performance by 9.5% and 9.2% over state-of-the-art methods, without significantly harming the original long CoT reasoning capability.