Graph-based retrieval-augmented generation (RAG) enables large language models (LLMs) to incorporate structured knowledge via graph retrieval as contextual input, enhancing more accurate and context-aware reasoning. We observe that for different queries, it could retrieve similar subgraphs as prompts, and thus we propose SubGCache, which aims to reduce inference latency by reusing computation across queries with similar structural prompts (i.e., subgraphs). Specifically, SubGCache clusters queries based on subgraph embeddings, constructs a representative subgraph for each cluster, and pre-computes the key-value (KV) cache of the representative subgraph. For each query with its retrieved subgraph within a cluster, it reuses the pre-computed KV cache of the representative subgraph of the cluster without computing the KV tensors again for saving computation. Extensive experiments on three datasets across multiple LLM backbones and graph-based RAG frameworks demonstrate that SubGCache consistently reduces inference latency with comparable and even improved generation quality, achieving up to 6.68x reduction in time-to-first-token (TTFT).
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
In-Token Rationality Optimization: Towards Accurate and Concise LLM Reasoning via Self-Feedback
PDF ↗Training Large Language Models (LLMs) for chain-of-thought reasoning presents a significant challenge: supervised fine-tuning on a single "golden" rationale hurts generalization as it penalizes equally valid alternatives, whereas reinforcement learning with verifiable rewards struggles with credit assignment and prohibitive computational cost. To tackle these limitations, we introduce InTRO (In-Token Rationality Optimization), a new framework that enables both token-level exploration and self-feedback for accurate and concise reasoning. Instead of directly optimizing an intractable objective over all valid reasoning paths, InTRO leverages correction factors—token-wise importance weights estimated by the information discrepancy between the generative policy and its answer-conditioned counterpart, for informative next-token selection. This approach allows the model to perform token-level exploration and receive self-generated feedback within a single forward pass, ultimately encouraging accurate and concise rationales. Across six math-reasoning benchmarks, InTRO consistently outperforms other baselines, raising solution accuracy by up to 20% relative to the base model. Its chains of thought are also notably more concise, exhibiting reduced verbosity. Beyond this, InTRO enables cross-domain transfer, successfully adapting to out-of-domain reasoning tasks that extend beyond the realm of mathematics, demonstrating robust generalization.
Effective customer support requires not only accurate problem-solving but also structured and empathetic communication aligned with professional standards. However, existing dialogue datasets often lack strategic guidance, and real-world service data is difficult to access and annotate. To address this, we introduce the task of Customer Support Conversation (CSC), aimed at training customer service supporters to respond using well-defined support strategies. We propose a structured CSC framework grounded in COPC guidelines, defining five conversational stages and twelve strategies to guide high-quality interactions. Based on this, we construct CSConv, an evaluation dataset of 1,855 real-world customer–agent conversations rewritten using LLMs to reflect deliberate strategy use, and annotated accordingly. Additionally, we develop a role-playing approach that simulates strategy-rich conversations using LLM-powered roles aligned with the CSC framework, resulting in the training dataset RoleCS. Experiments show that fine-tuning strong LLMs on RoleCS significantly improves their ability to generate high-quality, strategy-aligned responses on CSConv. Human evaluations further confirm gains in problem resolution.
Large Language Model-based Multi-Agent Systems (LLM-based MAS), where multiple LLM agents collaborate to solve complex tasks, have shown impressive performance in many areas. However, MAS are typically distributed across different devices or environments, making them vulnerable to perturbations such as agent failures. While existing works have studied the adversarial attacks and corresponding defense strategies, they mainly focus on reactively detecting and mitigating attacks after they occur rather than proactively designing inherently resilient systems. In this work, we study the resilience of LLM-based MAS under perturbations and find that both the communication topology and prompt design significantly influence system resilience. Motivated by these findings, we propose ResMAS: a two-stage framework for enhancing MAS resilience. First, we train a reward model to predict the MAS’s resilience, based on which we train a topology generator to automatically design resilient topology for specific tasks through reinforcement learning. Second, we introduce a topology-aware prompt optimization method that refines each agent’s prompt based on its connections and interactions with other agents. Extensive experiments across a range of tasks show that our approach substantially improves MAS resilience under various constraints. Moreover, our framework demonstrates strong generalization ability to new tasks and models, highlighting its potential for building resilient MASs.
Comprehension of ancient texts plays an important role in archaeology and understanding of Chinese history and civilization. The rapid development of large language models needs benchmarks that can evaluate their comprehension of ancient characters. Existing Chinese benchmarks are mostly targeted at modern Chinese and transmitted documents in ancient Chinese, but the part of excavated documents in ancient Chinese is not covered. To meet this need, we propose the AncientBench, which aims to evaluate the comprehension of ancient characters, especially in the scenario of excavated documents. The AncientBench is divided into four dimensions, which correspond to the four competencies of ancient character comprehension: glyph comprehension, pronunciation comprehension, meaning comprehension, and contextual comprehension. The benchmark also contains ten tasks, including radical, phonetic radical, homophone, cloze, translation, and more, providing a comprehensive framework for evaluation. We convened archaeological researchers to conduct experimental evaluations, proposed an ancient model as baseline, and conducted extensive experiments on the currently best-performing large language models. The experimental results reveal the great potential of large language models in ancient textual scenarios as well as the gap with humans. Our research aims to promote the development and application of large language models in the field of archaeology and ancient Chinese language.
Existing large language models (LLMs) occasionally generate plausible yet factually incorrect responses, known as hallucinations. Two main approaches have been proposed to mitigate hallucinations: retrieval-augmented language models (RALMs) and refusal post-training. However, current research predominantly focuses on their individual effectiveness while overlooking the evaluation of the refusal capability of RALMs. Ideally, if RALMs know when they do not know, they should refuse to answer. In this study, we ask the fundamental question: Do RALMs know when they don’t know? Specifically, we investigate three questions. First, are RALMs well calibrated with respect to different internal and external knowledge states? We examine the influence of various factors. Contrary to expectations, when all retrieved documents are irrelevant, RALMs still tend to refuse questions they could have answered correctly. Next, given the model's pronounced over-refusal behavior, we raise a second question: How does a RALM's refusal ability align with its calibration quality? Our results show that the over-refusal problem can be mitigated through in-context fine-tuning. However, we observe that improved refusal behavior does not necessarily imply better calibration or higher overall accuracy. Finally, we ask: Can we combine refusal-aware RALMs with uncertainty-based answer abstention to mitigate over-refusal? We develop a simple yet effective refusal mechanism for refusal-post-trained RALMs that improves their overall answer quality by balancing refusal and correct answers. Our study provides a more comprehensive understanding of the factors influencing RALM behavior. Meanwhile, we emphasize that uncertainty estimation for RALMs remains an open problem deserving deeper investigation.
Effective clinical history taking is a foundational yet underexplored component of clinical reasoning. While large language models (LLMs) have shown promise on static benchmarks, they often fall short in dynamic, multi-turn diagnostic settings that require iterative questioning and hypothesis refinement. To address this gap, we propose Note2Chat, a note-driven framework that trains LLMs to conduct structured history taking and diagnosis by learning from widely available medical notes. Instead of relying on scarce and sensitive dialogue data, we convert real-world medical notes into high-quality doctor-patient dialogues using a decision tree-guided generation and refinement pipeline. We then propose a three-stage fine-tuning strategy combining supervised learning, simulated data augmentation, and preference learning. Furthermore, we propose a novel single-turn reasoning paradigm that reframes history taking as a sequence of single-turn reasoning problems. This design enhances interpretability and enables local supervision, dynamic adaptation, and greater sample efficiency. Experimental results show that our method substantially improves clinical reasoning, achieving gains of +16.9 F1 and +21.0 Top-1 diagnostic accuracy over GPT-4o.
IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech
PDF ↗Existing autoregressive large-scale text-to-speech (TTS) models have advantages in speech naturalness, but their token-by-token generation mechanism makes it difficult to precisely control the duration of synthesized speech. This becomes a significant limitation in applications requiring strict audio-visual synchronization, such as video dubbing. This paper introduces IndexTTS2, which proposes a novel, general, and autoregressive model-friendly method for speech duration control. The method supports two generation modes: one explicitly specifies the number of generated tokens to precisely control speech duration; the other freely generates speech in an autoregressive manner without specifying the number of tokens, while faithfully reproducing the prosodic features of the input prompt. Furthermore, IndexTTS2 achieves disentanglement between emotional expression and speaker identity, enabling independent control over timbre and emotion. In the zero-shot setting, the model can accurately reconstruct the target timbre (from the timbre prompt) while perfectly reproducing the specified emotional tone (from the style prompt). To enhance speech clarity in highly emotional expressions, we incorporate GPT latent representations and design a novel three-stage training paradigm to improve the stability of the generated speech. Additionally, to lower the barrier for emotional control, we designed a soft instruction mechanism based on text descriptions by fine-tuning Qwen3, effectively guiding the generation of speech with the desired emotional orientation. Finally, experimental results on multiple datasets show that IndexTTS2 outperforms state-of-the-art zero-shot TTS models in terms of word error rate, speaker similarity, and emotional fidelity.
We investigate the capacity of Large Language Models (LLMs) for imaginative reasoning—the proactive construction, testing, and revision of hypotheses in information-sparse environments. Existing benchmarks, often static or focused on social deduction, fail to capture the dynamic, exploratory nature of this reasoning process. To address this gap, we introduce a comprehensive research framework based on the classic "Turtle Soup" game, integrating a benchmark, an agent, and an evaluation protocol. We present TurtleSoup-Bench, the first large-scale, bilingual, interactive benchmark for imaginative reasoning, comprising 800 turtle soup stories sourced from both the Internet and expert authors. We also propose Mosaic-Agent, a novel agent designed to assess LLMs' performance in this setting. To evaluate reasoning quality, we develop a multi-dimensional protocol measuring logical consistency, detail completion, and conclusion alignment. Experiments with leading LLMs reveal clear capability limits, common failure patterns, and a significant performance gap compared to humans. Our work offers new insights into LLMs' imaginative reasoning and establishes a foundation for future research on exploratory agent behavior.
Large Reasoning Models (LRMs) have advanced automated multi-step reasoning, but their ability to generate complex Chain-of-Thought (CoT) trajectories introduces severe privacy risks, as sensitive information may be deeply embedded throughout the reasoning process. Existing Large Language Models (LLMs) unlearning approaches that typically focus on modifying only final answers are insufficient for LRMs, as they fail to remove sensitive content from intermediate steps, leading to persistent privacy leakage and degraded security. To address these challenges, we propose Sensitive Trajectory Regulation (STaR), a parameter-free, inference-time unlearning framework that achieves robust privacy protection throughout the reasoning process. Specifically, we first identify sensitive content via semantic-aware detection. Then, we inject global safety constraints through secure prompt encoder. Next, we perform trajectory-aware suppression to dynamically block sensitive content across the entire reasoning chain. Finally, we apply token-level adaptive filtering to prevent both exact and paraphrased sensitive tokens during generation. Furthermore, to overcome the inadequacies of existing evaluation protocols, we introduce two metrics: Multi-Decoding Consistency Assessment (MCS), which measures the consistency of unlearning across diverse decoding strategies, and Multi-Granularity Membership Inference Attack (MIA) Evaluation, which quantifies privacy protection at both answer and reasoning-chain levels. Experiments on the R-TOFU benchmark demonstrate that STaR achieves comprehensive and stable unlearning with minimal utility loss, setting a new standard for privacy-preserving reasoning in LRMs.
Recent advances in large language models (LLMs) have shown remarkable capabilities across textual and multimodal domains. In parallel, large language diffusion models have emerged as a promising alternative to the autoregressive paradigm, offering improved controllability, bidirectional context modeling, and robust generation. However, their application to the audio modality remains underexplored. In this work, we introduce DIFFA, the first diffusion-based large audio-language model designed to perform spoken language understanding. DIFFA integrates a frozen diffusion language model with a lightweight dual-adapter architecture that bridges speech understanding and natural language reasoning. We employ a two-stage training pipeline: first, aligning semantic representations via an ASR objective; then, learning instruction-following abilities through synthetic audio-caption pairs automatically generated by prompting LLMs. Despite being trained on only 960 hours of ASR and 127 hours of synthetic instruction data, DIFFA demonstrates competitive performance on major benchmarks, including MMSU, MMAU, and VoiceBench, outperforming several autoregressive open-source baselines. Our results reveal the potential of large language diffusion models for efficient and scalable audio understanding, opening a new direction for speech-driven AI.
In collaborative filtering, learning effective embeddings for users and items from interaction data remains a central challenge. While recent efforts leverage large language models (LLMs) to enhance collaborative filtering, two critical limitations persist: (1) Efficiency: LLM-based inference is significantly slower than traditional embedding-based search; and (2) Topological Modeling: LLMs struggle to capture graph structures, which are essential for modeling multi-order user-item interactions. To address these limitations, we propose New Language Collaborative Filtering (NLCF), a framework that aligns LLMs with collaborative filtering by conceptualizing user-item graphs as new languages. This approach is based on two key insights: (1) LLMs excel at mastering new languages when trained on suitable corpora, and (2) the empirical conditional probability between tokens in corpora converges to the transition probabilities between nodes in graphs. NLCF translates user-item graphs into corpora, where users and items are treated as tokens. These corpora are used to fine-tune LLMs, and the learned representations are aggregated to construct user and item embeddings that encode multi-order interactions. Unlike methods that deploy LLMs for inference, NLCF distills LLM knowledge learned from corpora into compact embeddings, enabling both efficient training and real-time inference. The framework has been deployed on a billion-scale e-commerce platform for several months. Extensive experiments demonstrate that NLCF outperforms traditional graph CF models and LLM-based baselines while achieving significant training and inference efficiency improvement over LLM-based baselines.
HyperGLLM: An Efficient Framework for Endpoint Threat Detection via Hypergraph-Enhanced Large Language Models
PDF ↗Endpoint Detection and Response (EDR) systems are a cornerstone of modern threat detection and endpoint protection. However, conventional heuristic- and learning-based approaches often fail to address sophisticated and continuously evolving attack patterns. Recent advances in large language models (LLMs) offer promising capabilities for behavioral analysis in EDR logs, yet their effectiveness is hindered by the high volume of events and the interleaved nature of behavior sequences---posing significant challenges for long-context modeling and stealthy threat detection. To address these issues, we propose HyperGLLM, a novel detection framework that introduces hypergraph reasoning into LLMs. It first constructs an attribute-value level relation-aware graph to model low-order structural semantics while reducing textual redundancy. Then, it introduces a differential hypergraph module with multi-granularity clustering to capture high-order behavioral dependencies embedded in interleaved events and reinforce threat semantics. Finally, the hypergraph representations are aligned with an LLM for efficient contextual reasoning over potential malicious behaviors. To facilitate empirical evaluation, we curate EDR3.6B-63F, a large-scale EDR dataset containing 3.6 billion events across 63 distinct behavior families. Extensive experiments demonstrate that HyperGLLM significantly outperforms state-of-the-art methods by reducing the false alarm rate to 1.67%, achieving 94.65% accuracy across 63 behavior families, and improving the modeling efficiency of LLMs on long EDR logs. Our framework and dataset provide a solid foundation for future research and support the development of advanced detection solutions in endpoint security.
The evaluation of large language models (LLMs) via benchmarks is widespread, yet inconsistencies between different leaderboards and poor separability among top models raise concerns about their ability to accurately reflect authentic model capabilities. This paper provides a critical analysis of benchmark effectiveness, examining mainstream prominent LLM benchmarks using results from diverse models. We first propose Pseudo-Siamese Network for Item Response Theory (PSN-IRT), an enhanced Item Response Theory framework that incorporates a rich set of item parameters within an IRT-grounded architecture. PSN-IRT can be utilized for accurate and reliable estimations of item characteristics and model abilities. Based on PSN-IRT, we conduct extensive analysis on 11 LLM benchmarks comprising 41,871 items, revealing significant and varied shortcomings in their measurement quality. Furthermore, we demonstrate that leveraging PSN-IRT is able to construct smaller benchmarks while maintaining stronger alignment with human preference.
While reinforcement learning with verifiable rewards (RLVR) has advanced LLM reasoning in structured domains like mathematics and programming, its application to general-domain reasoning tasks remains challenging due to the absence of verifiable reward signals. To this end, methods like Reinforcement Learning with Reference Probability Reward (RLPR) have emerged, leveraging the probability of generating the final answer as a reward signal. However, these outcome-focused approaches neglect crucial step-by-step supervision of the reasoning process itself. To address this gap, we introduce Probabilistic Process Supervision (P2S), a novel self-supervision framework that provides fine-grained process rewards without requiring a separate reward model or human-annotated reasoning steps. During reinforcement learning, P2S synthesizes and filters a high-quality reference reasoning chain (gold-CoT). The core of our method is to calculate a Path Faithfulness Reward (PFR) for each reasoning step, which is derived from the conditional probability of generating the gold-CoT's suffix, given the model's current reasoning prefix. Crucially, this PFR can be flexibly integrated with any outcome-based reward, directly tackling the reward sparsity problem by providing dense guidance. Extensive experiments on reading comprehension and medical Question Answering benchmarks show that P2S significantly outperforms strong baselines.
Deep learning has significantly advanced numerous fields by training on extensive annotated datasets. However, this data-driven paradigm faces limitations such as limited adaptability and high annotation costs, particularly when precise adherence to detailed, domain-specific guidelines is required in annotation. This challenge raises a critical question: Can models effectively shift from data-driven learning to autonomously leveraging guidelines with minimal annotated examples? To address this, we propose the Guideline-Driven Prompt (GDP) optimization framework, which shifts the learning paradigm from data-driven training to guideline-driven reasoning. GDP leverages Retrieval Augmented Generation (RAG) to retrieve essential fragments from complex guidelines and synthesize them into structured, executable prompts. A tree-based optimization algorithm systematically constructs and refines these prompts, explicitly capturing the intricate logic embedded in professional guidelines through a latent pipeline structure. Empirical evaluations on four datasets ranging from diverse domains and different tasks demonstrate that GDP effectively transitions the learning process from data-intensive methods to a guideline-driven approach in tasks requiring detailed and complex guideline adherence, reducing dependence on extensive annotated datasets.
Compress-then-Rank: Faster and Better Listwise Reranking with Large Language Models via Ranking-Aware Passage Compression
PDF ↗Listwise reranking with Large Language Models (LLMs) has emerged as the state-of-the-art approach, consistently establishing new performance benchmarks in passage reranking. However, their practical application faces two critical hurdles: the prohibitive computational overhead and high latency of processing long token sequences, and the performance degradation caused by phenomena like "lost in the middle" in long contexts. To address these challenges, we introduce Compress-then-Rank (C2R), an efficient framework that performs listwise reranking not on original passages, but on their compact multi-vector surrogates. These surrogates can be pre-computed and cached for all passages in the corpus. The effectiveness of C2R hinges on three key innovations. First, the compressor model is pre-trained on a combination of text restoration and continuation objectives, enabling high-fidelity compressed vector sequences that mitigate the semantic loss common in single-vector methods. Second, a novel input scheme prepends embeddings of each ordinal index (e.g., [1]:) to its corresponding compressed vector sequence, which both delineates passage boundaries and guides the reranker LLM to generate a ranked list. Finally, the compressor and reranker are jointly optimized, making the compression explicitly ranking-aware for the ranking objective. Extensive experiments on major reranking benchmarks demonstrate that C2R provides substantial speedups while achieving competitive and even superior ranking performance compared to full-text reranking methods.
Short texts present significant challenges for clustering due to semantic sparsity, limited contextual information, and ambiguous category boundaries. While recent studies incorporating contrastive learning and cluster structure optimization have improved performance, their reliance on augmented samples often introduces noise and weakens the capacity of pretrained language models to capture fine-grained semantics. To address these issues, we propose a Graph-augmented and Over-smoothing-resistant Contrastive Clustering framework (GOCC). Specifically, GOCC constructs sentence-level and cluster-level graphs to capture local semantic similarity and global structural patterns, incorporating these signals into sentence representations to enhance representational quality and clustering suitability. Moreover, we introduce a contrastive mechanism based on intermediate layer representations within graph-augmented contrastive learning to alleviate semantic over-smoothing caused by deep networks. Finally, a target-distribution-driven clustering optimization strategy is employed to leverage high-confidence samples in guiding cluster assignments. Experimental results on several benchmark short text datasets demonstrate that GOCC consistently outperforms state-of-the-art methods in terms of clustering accuracy and normalized mutual information.
Navigating Through Paper Flood: Advancing LLM-Based Paper Evaluation Through Domain-Aware Retrieval and Latent Reasoning
PDF ↗With the rapid and continuous increase in academic publications, identifying high-quality research has become an increasingly pressing challenge. While recent methods leveraging Large Language Models (LLMs) for automated paper evaluation have shown great promise, they are often constrained by outdated domain knowledge and limited reasoning capabilities. In this work, we present PaperEval, a novel LLM-based framework for automated paper evaluation that addresses these limitations through two key components: 1) a domain-aware paper retrieval module that retrieves relevant concurrent work to support contextualized assessments of novelty and contributions, and 2) a latent reasoning mechanism that enables deep understanding of complex motivations and methodologies, along with comprehensive comparison against concurrently related work, to support more accurate and reliable evaluation. To guide the reasoning process, we introduce a progressive ranking optimization strategy that encourages the LLM to iteratively refine its predictions with an emphasis on relative comparison. Experiments on two datasets demonstrate that PaperEval consistently outperforms existing methods in both academic impact and paper quality evaluation. In addition, we deploy PaperEval in a real-world paper recommendation system for filtering high-quality papers, which has gained strong engagement on social media---amassing over 8,000 subscribers and attracting over 10,000 views for many filtered high-quality papers---demonstrating the practical effectiveness of PaperEval.
Understanding multimodal metaphors represents a crucial pathway for machines to comprehend human cognition. However, current research remains constrained by superficial dataset annotations, insufficient systematic evaluation of large language models, and fragmented task frameworks. To bridge these gaps, the paper proposes a systematic solution featuring: (I) We present the largest fine-grained Multi-task Multimodal Metaphor Understanding Challenge Dataset (M3UCD) built via multi-perspective collaborative annotation. It contains 15,345 samples, each annotated with 12 manual attribute labels. (II) Systematic benchmarking of LLMs' capacity boundaries in metaphor understanding. Evaluation results reveal the persistent challenges LLMs face in this domain while validating M3UCD's effectiveness and potential. (III) A concise and unified multi-task baseline framework was developed and demonstrated its effectiveness in enhancing the metaphor understanding capabilities of MLLMs.