With the remarkable performance of large language models (LLMs) in medicine, particularly their ability to support clinical decision-making in medical dialogues, a key limitation remains: the static reasoning patterns derived from human expert experience are often inadequate for the dynamic and diverse nature of real-world multi-turn conversations. While recent large reasoning models (such as R1) enable deeper and more complex thought processes to address such challenges, they also introduce significant redundancy. Meanwhile, recent studies on reusing atomic thoughts demonstrate a practical pathway toward dynamic and precise reasoning in general domains. In this paper, we investigate the role of atomic thought-based experience in medical dialogue tasks. First, we collect human expert clinical experience. Then, we propose a novel distillation framework that extracts atomic thoughts from teacher models and reuses them to guide reasoning and generate responses. Based on this framework, we construct training data from ReMeDi and fine-tune student models, which demonstrate enhanced performance in both static and interactive medical dialogue scenarios. Furthermore, we examine the impact of experience across various models, datasets, and scenarios. Crucially, transferring this experience empowers weaker models to generate high-quality reasoning data, matching the annotation capabilities of stronger LLMs while significantly reducing costs. The code is available in this repository https://github.com/VioletAmethystLunar/Atomic-Thoughts-Medical-Dialogue.
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
Long-context Large Language Models, despite their expanded capacity, require careful working memory management to mitigate attention dilution during long-horizon tasks. Yet existing approaches rely on external mechanisms that lack awareness of the agent’s reasoning state, leading to suboptimal decisions. We propose Memory-as-Action (MemAct), a framework that treats working memory management as learnable policy actions. By formulating context management as in-place editing operations (deletion, insertion), MemAct enables joint optimization of information retention and task performance through end-to-end reinforcement learning. To address the computational challenges of dynamic context updates, we introduce Dynamic Context Policy Optimization, which restores training efficiency without compromising reasoning integrity. Experiments show that MemAct-RL-14B matches the accuracy of models 16× larger while reducing average context length by 51%, with learned strategies that adapt to model capabilities and generalize across task complexities. The code and datasets are available at https://github.com/ADaM-BJTU/MemAct.
Retrieval-Augmented Generation (RAG) systems rely on Optical Character Recognition (OCR) to ingest knowledge from unstructured documents. However, OCR engines often struggle with complex layouts, introducing Structural Noise, such as line insertion and paragraph interleaving, which disrupts the semantic flow of the text. Existing evaluations largely overlook this dimension, operating on the assumption of structurally perfect input. To bridge this gap, we introduce StruNRAG, a dedicated benchmark for evaluating RAG robustness against OCR-induced structural perturbations. We construct a bilingual dataset of 2,132 question-answer pairs derived from complex Chinese and English documents and systematically inject three categories of real-world structural noise: line insertion, paragraph interleaving, and line interleaving. Our evaluation of mainstream retrievers and Large Language Models (LLMs) reveals a nuanced interaction between noise and pipeline stages: while structural distortions consistently degrade retrieval performance, the generation stage exhibits unexpected robustness. Advanced LLMs demonstrate robustness against local noise (e.g., line insertion), but struggle to maintain reasoning capabilities under severe structural disruption that fragments global context. These findings indicate that while LLMs are capable of compensating for minor parsing errors, future RAG optimizations must take into account the effects of structural noise. Our code and datasets are available at [https://github.com/GaoMengnana/StruNRAG](https://github.com/GaoMengnana/StruNRAG).
Low-Rank Adaptation (LoRA) is a widely used method to fine-tune large language models, extremely reducing computational and storage costs. But its fixed-rank design cannot well capture the varying importance across different layers, limiting its flexibility. Dynamic rank allocation methods mitigate this issue by adaptively allocating ranks during training, but most of them focus solely on either rank pruning or expansion, leading to redundant parameterization or insufficient representational capacity. To address this problem, we introduce Hebbian-Guided Bi-Directional Rank Adaptation (HeBiRA), a novel framework that bi-directionally reallocates low-rank capacity using Hebbian-inspired importance estimation. HeBiRA computes the contribution of each rank direction by measuring the synergy between activations and output gradients, and adjusts the rank bi-directionally by pruning uninformative directions while expanding those in critical layers. This mechanism flexibly redistributes the rank budget during training and also can be applied to PEFT methods such as DoRA, HiRA, and QLoRA. Experiments on multiple benchmarks and theoretical analysis show that HeBiRA consistently improves performance over baselines.
DeCoVec: Building Decoding Space based Task Vector for Large Language Models via In-Context Learning
PDF ↗Task vectors, representing directions in model or activation spaces that encode task-specific behaviors, have emerged as a promising tool for steering large language models (LLMs). However, existing approaches typically require fine-tuning or invasive manipulation of internal states, limiting their flexibility and scalability. We propose DeCoVec (Decoding Space based Task Vector), a training-free and non-invasive framework that constructs task vectors directly in the decoding space by leveraging in-context learning (ICL). Specifically, DeCoVec captures the task essence as the difference between the output logit distributions of few-shot and zero-shot prompts, then steers generation by injecting this vector into the decoding process. Experiments across seven LLMs (0.5B–9B) on TruthfulQA, Math-500, and AQUA-RAT show that DeCoVec consistently outperforms standard few-shot baselines, with gains up to +5.50 average accuracy. Further analysis demonstrates that DeCoVec effectively suppresses generation degeneration and logical flaws while exhibiting strong robustness to demonstration ordering, all without incurring additional input token costs. Our method offers a training-free and non-invasive solution for LLM steering without requiring weight updates or auxiliary models.
When reading foreign-language literature, non-native users often face significant challenges. Existing traditional machine translation systems tend to obscure or mistranslate key terminology, while paraphrasing aimed at lay readers often oversimplifies it, thereby hindering their ability to master domain-specific technical vocabulary. To bridge this gap, we first define a novel task, Glossing-Oriented Academic Translation (GOAT), which aims to produce translations dynamically adapted to a reader’s academic proficiency, or level. We then propose GlossaGen, a comprehensive framework to address this task. GlossaGen features two key innovations: a multi-agent data synthesis pipeline that leverages academic personas to automatically generate a large-scale, structured dataset with level-specific explanations; and a novel training strategy based on dynamic adapter merging, which balances task generalization with user-level specialization by combining a ”generalist” adapter with a fine-grained ”expert” one. We evaluate GlossaGen on our synthesized benchmark, where results from automatic metrics, large language model (LLM)-based assessments, and human evaluations consistently demonstrate that our approach achieves higher scores than strong baselines across most metrics. Our framework provides a scalable pathway to enhance the comprehensibility of scientific literature for non-native readers, delivering more accurate translations accompanied by pedagogically sound, level-specific term explanations, and we release our code and data to facilitate further research.
Reinforcement Learning with Verifiable Rewards (RLVR) has shown significant promise for enhancing the reasoning capabilities of large language models (LLMs). However, prevailing algorithms like GRPO broadcast a uniform advantage signal across all tokens in a sequence. This coarse-grained approach overlooks the pivotal role of uncertain, high-stakes decisions during reasoning, leading to inefficient exploration and the well-documented problem of entropy collapse. To address this, we introduce UnCertainty-aware Advantage Shaping (UCAS), a model-free method that refines credit assignment by leveraging the model’s internal uncertainty signals. UCAS operates in two stages: it first modulates the response-level advantage using the model’s overall self-confidence, and then applies a token-level penalty based on raw logit certainty. This dual mechanism encourages exploration of high-uncertainty paths that yield correct answers while penalizing overconfident yet erroneous reasoning, effectively balancing the exploration-exploitation trade-off. Extensive experiments on five mathematical reasoning benchmarks show that UCAS significantly outperforms strong RLVR baselines across multiple model scales, including 1.5B and 7B. Our analysis confirms that UCAS not only achieves higher rewards but also promotes greater reasoning diversity and successfully mitigates entropy collapse.
Prior work synthesizes tool-use LLM datasets by first generating a user query, followed by complex tool-use annotations like DFS. This inherently leads to inevitable annotation failures and low efficiency in data generation. We introduce ToolGrad, an agentic framework that inverts this paradigm. ToolGrad first constructs valid tool-use chains through an iterative process guided by textual "gradients", and then synthesizes corresponding user queries. This "answer-first" approach led to ToolGrad-500, a dataset generated with more complex tool use, lower cost, and almost 100% pass rate. Experiments show that models trained on ToolGrad-500 outperform those trained on expensive baseline datasets and proprietary LLMs.
Reverse Constitutional AI: A Framework for Controllable Toxic Data Generation via Probability-Clamped RLAIF
PDF ↗Ensuring the safety of large language models (LLMs) requires robust red teaming, yet the systematic synthesis of high-quality toxic data remains under-explored. We propose Reverse Constitutional AI (R-CAI), a framework for automated and controllable adversarial data generation that moves beyond isolated jailbreak prompts. By inverting a harmless constitution into a constitution of toxicity and iteratively refining model outputs through a critique–revision pipeline, R-CAI enables scalable synthesis of multi-dimensional adversarial data without human annotation. Optimizing solely for toxicity-related rewards, however, can lead to reward hacking and degraded semantic coherence. To address this challenge, we introduce probability clamping within reinforcement learning from AI feedback, which stabilizes adversarial optimization while preserving adversarial intent. Experiments demonstrate that R-CAI generates diverse, high-quality toxic data and that probability clamping substantially improves semantic coherence (15%) without sacrificing adversarial strength. Overall, R-CAI provides a fully automated framework for red teaming data generation and systematic safety evaluation of aligned language models.
Failures are inevitable when embodied agents execute complex tasks. Visual-language models (VLMs) serve as the core component of embodied agents in perceiving the environment and making decisions. Assessing the capabilities of VLMs in detecting and reasoning about failures has become increasingly important. Previous work primarily considered low-level manipulation failures (e.g., 3cm grasp offsets), neglecting high-level failures arising during long-horizon task execution (e.g., object-dropping failure in the “clean room” task) by embodied agents. In this paper, we propose FAER, a failure-aware benchmark aiming to evaluate the performance of VLMs in terms of failure detection, failure categorization, failure description, and failure correction in long-horizon tasks. FAER comprises 3,323 episodes, spanning 3 scenes, 65 tasks, and 83 objects. We assess the performance of 16 widely utilized VLMs and 4 LLMs for FAER tasks. Experimental results show that nearly all VLMs, even GPT-4o, exhibit limited performance in failure detection with a high false negative rate, meaning that they tend to ignore abnormal events, revealing notable gaps in current models’ capacity to effectively handle failures.
Selecting an appropriate LLM configuration for a given query is critical, yet existing routing frameworks operate within a single computational paradigm. To address this gap, we formalize the Cross-System Routing Problem, a hierarchical decision-making task that decomposes routing into intra-regime configuration selection and inter-regime system selection. Building on this, we propose BiCSRouter, a bi-level cross-system routing framework that integrates two orthogonal regimes: intensive reasoning via single-agent systems and extensive collaboration via multi-agent systems. BiCSRouter performs policy learning within each system and employs a lightweight inter-regime router that selects the optimal regime based on predicted performance and cost. Experiments on the MBPP and MATH benchmarks demonstrate that BiCSRouter outperforms 15 representative baselines across three types. On MBPP, compared to the performance ceiling of GPT-5, BiCSRouter achieves a 46% reduction in cost with only a 2% drop in accuracy. Finally, we show that BiCSRouter can extend to additional regimes, highlighting its generality as a cross-system routing framework.
Exploratory GUI testing is essential for software quality but suffers from high manual costs. While Multi-modal Large Language Model (MLLM) agents excel in navigation, they fail to autonomously discover defects due to two core challenges: Goal-Oriented Masking, where agents prioritize task completion over reporting anomalies, and Execution-Bias Attribution, where system defects are misidentified as agent errors. To address these, we first introduce GUITestBench, the first interactive benchmark for this task, featuring 143 tasks across 26 defects. We then propose GUITester, a multi-agent framework that decouples navigation from verification via two modules: (i) a Planning-Execution Module (PEM) that proactively probes for defects via embedded testing intents, and (ii) a Hierarchical Reflection Module (HRM) that resolves attribution ambiguity through interaction history analysis. GUITester achieves an F1-score of 48.90% (Pass@3) on GUITestBench, outperforming state-of-the-art baselines (33.35%). Our work demonstrates the feasibility of autonomous exploratory testing and provides a robust foundation for future GUI quality assurance.
Large language models (LLMs) have made progress in knowledge-intensive tasks, reasoning and planning, and collaborative problem solving, yet they exhibit intrinsic limitations such as knowledge cutoff, single-threaded reasoning that hinders finer-grained branch and aggregation, and rigid collaboration mechanisms that struggle to coordinate specialized capabilities. Graphs, with their ability to represent relational knowledge and complex dependencies, offer a natural means to address these limitations: they provide structured, high-density knowledge for augmenting or correcting LLMs’ generation; enable revisitable inference by organizing intermediate steps as graphs; and support dynamic coordination among experts or agents in collaborative settings. Motivated by these developments, we present the first systematic survey of graph-assisted LLMs from the perspective of how graph structures mitigate LLMs’ limitations. We introduce a taxonomy spanning *Graph-Assisted Knowledge Augmentation*, *Graph-Assisted Reasoning and Planning*, and *Graph-Assisted LLM Collaboration*, and analyze representative methods, summarize common design patterns, and outline open challenges and future directions for advancing LLMs with graph-based enhancements. The collected papers are available in [link here](https://github.com/FairyFali/Graph4LLM-Survey).
Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework
PDF ↗Sparse autoencoders (SAEs) enable interpretability research by decomposing entangled model activations into monosemantic features. However, under what circumstances SAEs derive most fine-grained latent features for safety—a low-frequency concept domain—remains unexplored. Two key challenges exist: identifying SAEs with the greatest potential for generating safety domain-specific features, and the prohibitively high cost of detailed feature explanation. In this paper, we propose **Safe-SAIL**, a unified framework for interpreting SAE features in safety-critical domains to advance mechanistic understanding of large language models. Safe-SAIL introduces a pre-explanation evaluation metric to efficiently identify SAEs with strong safety domain-specific interpretability, and reduces interpretation cost by 55% through a segment-level simulation strategy. Building on Safe-SAIL, we train a comprehensive suite of SAEs with human-readable explanations and systematic evaluations for 1,758 safety-related features spanning four domains: pornography, politics, violence, and terror. Using this resource, we conduct empirical analyses and provide insights on the effectiveness of Safe-SAIL for risk feature identification and how safety-critical entities and concepts are encoded across model layers. All models, explanations, and tools are publicly released in an open-source toolkit at https://anonymous.4open.science/r/Safe-SAIL/.
Large language models (LLMs) have been widely adopted in healthcare, yet they still encounter significant challenges in complex clinical decision-making scenarios. Existing benchmarks primarily assess LLM performance in single-course settings and lack systematic evaluation in multi-course scenarios, where a patient’s condition evolves over time. To address this gap, we propose ClinicalMC, a benchmark for multi-course clinical decision-making. It includes 1,275 Chinese and 5,804 English samples across four stages from admission to discharge. These stages cover triage, first-course examination/diagnosis/treatment, subsequent multi-course examination/assessment/treatment, and final diagnosis. In ClinicalMC, patients in the English dataset undergo an average of 5.11 clinical courses, whereas those in the Chinese dataset undergo 3.42. To assess LLM performance, we construct a multi-agent evaluation framework that includes patient, examiner, and doctor agents. Based on the benchmark and framework, we design two experimental settings—a single-turn static setting and a multi-turn dynamic setting—and assess three categories of LLMs: 1) closed-source LLMs like GPT-4o-mini; 2) open-source LLMs like DeepSeek-V3, and 3) medical LLMs like HuatuoGPT-o1. Through extensive evaluation, we aim to better understand LLM performance in the medical domain and support its effective deployment in healthcare.
Large language models require continuous adaptation to new tasks while preserving safety alignment. However, fine-tuning on even benign data often compromises safety behaviors. We investigate which training samples cause alignment drift through a data-centric lens. Our experiments show samples contribute unequally: high-gradient samples cause greater safety degradation and drive models toward pretrained distributions, while moderate-gradient samples enable task learning with minimal alignment loss. This connects to the elasticity phenomenon—high-gradient samples activate the reversion force pulling models toward pretrained behavior. We propose gradient-based sample selection that filters high-gradient samples during fine-tuning. Across multiple model families on continual domain tasks, our method substantially improves alignment preservation while maintaining competitive task performance, without requiring curated safe data or architectural modifications.
Zero-shot Relational Learning (ZRL) aims to perform knowledge graph completion when dealing with newly emerging relations without instances of them. However, existing ZRL methods typically depend on external knowledge beyond Knowledge Graphs (KGs), resulting in increased annotation costs and limited practical applicability. To address this issue, we propose a new **S**tructure-**A**ware paradigm for **ZRL**, termed **SAZRL**, that performs ZRL without relying on external knowledge. SAZRL leverages intrinsic structural patterns in KGs to bridge semantic correlations for new relations with existing ones. It constructs structure-aware conditional query graphs based on shared entities and adaptive relation updating module to generate representations for new relations based on the query graphs. We conduct extensive experiments on three real-world benchmarks, **NELL-ZS**, **Wiki-ZS** and **FB15K-ZS**, demonstrating that SAZRL consistently surpasses state-of-the-art ZRL methods, achieving up to **10.66%** improvement in **MRR** while reducing annotation costs and enhancing practical applicability. **The code and data are provided in supplementary materials.**
Beyond Sentence-level Labels: Integrating Conversational Context and Personal Experience for Natural Emotional Expression
PDF ↗Emotional Text-to-Speech aims to synthesize speech with human-like naturalness and expressiveness. However, existing systems rely on sentence-level labels, which fails to capture the subtle nuances of human affect. Based on cognitive appraisal theories, we argue that emotional expression is not generated in isolation but is deeply influenced by speaker’s Personal Experience and the conversational Context.To overcome the information bottleneck inherent in traditional annotations, we present Emotional-Context-Speech, a large-scale, context-aware speech corpus derived from multi-speaker audiobooks. This dataset provides not only transcriptions but also dialogue context, personal experience, open-vocabulary emotion labels, and paralinguistic descriptions.Experimental results demonstrate that TTS model trained using additional context and experience descriptions as inputs, called Emotional-Context-TTS, significantly outperforms existing methods in terms of emotional expression accuracy and naturalness.
Privacy-Preserving Reasoning with Knowledge-Distilled Parametric Retrieval Augmented Generation
PDF ↗The current RAG system requires uploading plaintext documents to the cloud, risking private data leakage. Parametric RAG (PRAG) encodes documents as LoRA parameters within LLMs, offering a possible way to reduce exposure of raw content. However, it still faces two issues: (1) PRAG demands synthesizing QA pairs and fine-tuning LLM for each individual document to create its corresponding LoRA, leading to unacceptable inference latency. (2) The performance of PRAG relies solely on synthetic QA data while lacking internal alignment with standard RAG, resulting in poor generalization on out-of-distribution (OOD) inputs. Therefore, achieving high-efficiency parameterization while maintaining RAG-level performance remains a critical challenge for privacy-preserving reasoning. In this paper, we propose DistilledPRAG, a generalizable knowledge-distilled parametric RAG model aligned with standard RAG in document structure and parameter activation. We first synthesize QA pairs from single and multi-documents to enhance cross-document reasoning. Then, we mask the plaintext documents with a special token and translate them to LoRA via a parameter generator, maintaining the standard RAG document structure. Finally, guided by synthetic QA data, we train the parameter generator to match standard RAG’s hidden states and output logits, enabling RAG-style reasoning without original documents. Experiments on four QA datasets show that DistilledPRAG outperforms baselines in accuracy and generalizes well on OOD data.
While Large Language Models (LLMs) excel in autonomous agent settings, small language models (SLMs) remain fragile, often collapsing after encountering errors. Traditional knowledge distillation focuses on imitating successful trajectories, while existing "learning from mistakes" methods treat errors as auxiliary signals rather than states requiring recoverable policies, leaving the dynamics of failure and recovery in agent settings largely unexplored. Inspired by Donald Schön’s theory of reflective practice, we propose P-BRIDGE (Pedagogical Bridge for Reflective Insight and Distillation of Guiding Errors). P-BRIDGE combines reflection-in-action with reflection-on-action, enabling agents to diagnose and correct critical errors during execution while abstracting transferable strategies from contrastive student–teacher trajectories. Experiments across eight benchmarks demonstrate that P-BRIDGE significantly elevates SLM performance—e.g., raising the 2WikiMultiHopQA accuracy of a 0.6B model from 6.2% to 34.2%.