Recently, long-thought reasoning LLMs, such as OpenAI’s O1, adopt extended reasoning processes similar to how humans ponder over complex problems. This reasoning paradigm significantly enhances the model’s problem-solving abilities and achieves promising results. However, long-thought reasoning process leads to a substantial increase in inference time. A pressing challenge is reducing the inference overhead of long-thought LLMs while ensuring accuracy. In this paper, we identify that long-thought reasoning models struggle to effectively allocate token budgets based on problem difficulty and reasoning redundancies. To address this, we propose Length-Harmonizing Fine-Tuning (O1-Pruner), aiming at minimizing reasoning overhead while maintaining accuracy. This effective fine-tuning method first estimates the LLM’s baseline performance through pre-sampling and then uses RL-style fine-tuning to encourage the model to generate shorter reasoning processes under accuracy constraints. This allows the model to achieve efficient reasoning with lower redundancy while maintaining accuracy. Experiments on various mathematical reasoning benchmarks show that O1-Pruner not only significantly reduces inference overhead but also achieves higher accuracy, providing a novel and promising solution to this challenge.
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
BRIEF-Pro: Universal Context Compression with Short-to-Long Synthesis for Fast and Accurate Multi-Hop Reasoning
PDF ↗As retrieval-augmented generation (RAG) tackles complex tasks, increasingly expanded contexts offer richer information, but at the cost of higher latency and increased cognitive load on the model. To mitigate this bottleneck, especially for intricate multi-hop questions, we introduce BRIEF-Pro. It is a universal, lightweight compressor that distills relevant evidence for a given query from retrieved documents into a concise summary for seamless integration into in-context RAG. Using seed data consisting of relatively short contexts (fewer than 1k words), BRIEF-Pro is trained to perform abstractive compression of extended contexts exceeding 10k words across a wide range of scenarios. Furthermore, BRIEF-Pro offers flexible user control over summary length by allowing users to specify the desired number of sentences. Experiments on four open-domain multi-hop question-answering datasets show that BRIEF-Pro generates more concise and relevant summaries, enhancing performance across small, large, and proprietary language models. With the 70B reader model, 32× compression by BRIEF-Pro improves QA performance by 4.67% on average over LongLLMLingua’s 9×, while requiring only 23% of its computational overhead.
Optimizing Native Sparse Attention with Latent Attention and Local Global Alternating Strategies
PDF ↗In this work, we conduct a systematic analysis of Native Sparse Attention (NSA) and propose targeted improvements that enhance long-context modeling. A key insight is that alternating between local (sliding-window) and global (compression/selective) attention across layers, rather than using fixed patterns, enables more effective propagation of long-range dependencies and substantially boosts performance on long-sequence tasks. Meanwhile, we further refine NSA’s branches with Latent Attention that the sliding-window branch is enhanced with Multi-head Latent Attention (MLA) while compression and selective branches adopt Group-head Latent Attention (GLA). These changes reduce KV-cache memory by 50% versus NSA while improving the model’s common-sense reasoning and long-text understanding capabilities. Experiments on models from 340M to 1.3B parameters (trained on 15B and 100B tokens) show our method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks.
Large language models (LLMs) are increasingly applied in financial scenarios. However, they may produce harmful outputs, including facilitating illegal activities or unethical behavior, posing serious compliance risks. To systematically evaluate LLM safety in finance, we propose FinSafetyBench, a bilingual (English-Chinese) red-teaming benchmark designed to test an LLM’s refusal of requests that violate financial compliance. Grounded in real-world financial crime cases and ethics standards, the benchmark comprises 14 subcategories spanning financial crimes and ethical violations. Through extensive experiments on general-purpose and finance-specialized LLMs under three representative attack settings, we identify critical vulnerabilities that allow adversarial prompts to bypass compliance safeguards. Further analysis reveals stronger susceptibility in Chinese contexts and highlights the limitations of prompt-level defenses against sophisticated or implicit manipulation strategies.
MedScore: Generalizable Factuality Evaluation of Open-ended Long-form Medical Answers by Domain-adapted Claim Decomposition and Verification
PDF ↗While Large Language Models (LLMs) can generate fluent and convincing responses, they are not necessarily correct. This is especially apparent in the popular decompose-then-verify factuality evaluation pipeline, where LLMs evaluate generated text by decomposing it into individual, valid claims. Factuality evaluation is especially important for medical answers, since incorrect medical information could seriously harm the patient. However, existing factuality systems are a poor match for the medical domain, as they are typically only evaluated on objective, entity-centric, formulaic texts such as biographies and historical topics. This differs from condition-dependent, conversational, hypothetical, sentence-structure diverse, and subjective medical answers, making decomposition into valid facts challenging. We propose MedScore, a new pipeline to decompose medical answers into condition-aware valid facts and verify against in-domain corpora. Our method extracts up to three times as many valid facts as existing methods, reducing hallucination and vague references, and retaining condition-dependency in facts. We also find MedScore is generalizable to non-medical domains without any specific tuning. The resulting factuality score substantially varies by decomposition method, verification corpus, and used backbone LLM, highlighting the importance of customizing each step for reliable factuality evaluation by using our generalizable and modularized pipeline for domain adaptation.
Simulating Crisis Cognition: A Computational Framework for Hypothesis Generation in Crisis Communication
PDF ↗Large Language Models (LLMs) have demonstrated remarkable fidelity in simulating social dynamics, yet using them to inform high-stakes crisis policy requires rigorous causal evaluation. We introduce CRISIS COGNITION, a framework rooted in generative Structural Causal Models (SCM) that functions as an in-silico hypothesis generator. By coupling real-world telemetry with 1,813 agents, we conduct a counterfactual simulation to evaluate communication strategies. Unlike prior descriptive work, we employ a Stratified Analysis to strictly control for personality confounders. Our simulations generate a computational hypothesis: within the LLM’s generative process, emotional scaffolding serves as a functional prerequisite to unlock valid reasoning paths for high-neuroticism agents. Crucially, we identify a “Sedative Effect” in simultaneous interventions, confirming that the sequence of support is as vital as the content. This framework provides a rigorous testbed for evaluating strategies before human-subject trials.
Large Language Models (LLMs) have demonstrated remarkable capabilities in various reasoning-intensive tasks. However, these models exhibit unexpected brittleness, often failing on simple variations of the same underlying task. Existing robustness evaluations predominantly rely on hand-crafted templates or a limited set of perturbation rules. Consequently, such approaches lack the adaptability to probe latent vulnerabilities unique to specific models and remain susceptible to data contamination. To address this, we propose the Math Stress Tester (MaSTer), an automated framework inspired by software stress testing. MaSTer generates adversarial variants via a multi-round rewrite-verify loop, ensuring semantic consistency while successfully inducing model failure. Our framework generates benchmark variants dynamically for each LLM, thus minimizing the risk of data contamination. Experiments on GSM8K and MATH-500 demonstrate the effectiveness of MaSTer on mathematical tasks. Additionally, we validate the framework’s extensibility to non-mathematical tasks, highlighting its broad applicability. Furthermore, we demonstrate that the synthesized variants generated by MaSTer can be utilized as a fine-tuning dataset to significantly enhance the model’s robustness.
Large Language Models have shown promise in translating natural language into executable optimization models, yet they often suffer from the Sisyphus Dilemma: a memoryless cycle where identical errors are repeated across structurally similar problems. Existing retrieval-augmented strategies primarily fetch static problem-model pairs as few-shot demonstrators, failing to capture the dynamic reasoning required to resolve execution failures. To bridge this gap, we propose EOM, a framework that implements Experience Replay to transform transient rectification steps into persistent knowledge. EOM distills interaction histories into Causal Correction Mappings, indexing both diagnostic insights and prohibitive traps. By utilizing a structure-aware retrieval mechanism that aligns semantic intent with abstract syntax trees and solver tracebacks, the system enables models to recall specific correction strategies for isomorphic errors. Extensive experiments across seven benchmarks demonstrate that EOM improves modeling accuracy by 8.45% on complex tasks while reducing token consumption by 28.65% and interaction turns by 25.82%, validating the efficiency of a “Rectify Once, Solve Many” paradigm.
Recognizing whether outputs from large language models (LLMs) contain faithfulness hallucination is crucial for real-world applications, e.g., retrieval-augmented generation and summarization. In this paper, we introduce FaithLens, a cost-efficient and effective faithfulness hallucination detection model that can jointly provide binary predictions and corresponding explanations to improve trustworthiness. To achieve this, we first synthesize training data with explanations via advanced LLMs and apply a well-defined data filtering strategy to ensure label correctness, explanation quality, and data diversity. Subsequently, we fine-tune the model on these well-curated training data as a cold start and further optimize it with rule-based reinforcement learning, using rewards for both prediction correctness and explanation quality. Results on 12 diverse tasks show that the 8B-parameter FaithLens outperforms advanced models such as GPT-5.2 and o3. Also, FaithLens can produce high-quality explanations, delivering a distinctive balance of trustworthiness, efficiency, and effectiveness.
As text-to-music models gain widespread adoption, the prompts used to guide these systems have become valuable intellectual property. This shift has given rise to a new form of attack: prompt stealing, aiming to reconstruct the high-value prompts that guide the music generation. However, unlike prior work in text and image generation, prompt stealing in text-to-music systems faces unique challenges due to the entangled and diffuse nature of semantic representations in audio, which complicates the decoupling of specific textual tokens from acoustic outputs. To address these challenges, we present AudioStealer, the first targeted study of prompt inversion in the audio domain. AudioStealer operates via a two-stage black-box attack framework: first, a heuristic search guided by audio-language embeddings identifies initial candidates; then, these candidates are refined using a game-theoretic strategy based on Shapley value estimation to attribute precise semantic contributions. Our method requires no direct access to the target model and relies solely on a shadow model, making it broadly applicable. Through extensive experiments, we demonstrate that AudioStealer recovers prompts with high textual consistency to the ground truth, while the regenerated audio maintains strong perceptual similarity to the target recordings. These results expose critical vulnerabilities in the text-to-audio market ecosystem and underscore the urgent need for intellectual property protections in generative audio technologies.
Large Language Models (LLMs), despite their success in question answering, exhibit limitations in complex multi-hop question answering (MQA) tasks that necessitate non-linear, structured reasoning. This limitation stems from their inability to adequately capture deep conceptual relationships between entities. To overcome this challenge, we present ORACLE (Ontology-driven Reasoning And Chain for Logical Elucidation), a training-free framework that combines LLMs’ generative capabilities with the structural benefits of knowledge graphs. Our approach operates through three stages: (1) dynamic construction of question-specific knowledge ontologies using LLMs, (2) transformation of these ontologies into First-Order Logic (FOL) reasoning chains, and (3) systematic decomposition of the original query into logically coherent sub-questions. Extensive experiments across a diverse set of models and standard MQA benchmarks demonstrate that our framework achieves competitive performance while producing more interpretable reasoning chains.
With the proliferation of LLM-driven multi-agent systems (MAS), the security of Web links has become a critical concern. Once MAS is induced to trust a malicious link, attackers can use it as a springboard to expand the attack surface. In this paper, we propose Web Fraud Attacks, a novel type of attack manipulating unique structures of web links to deceive MAS. We design 12 representative attack variants that encompass various methods, such as homoglyph deception, sub-directory nesting, and parameter obfuscation. Through extensive experiments on these attack vectors, we demonstrate that Web fraud attacks not only exhibit significant destructive potential across different MAS architectures but also possess a distinct advantage in evasion: they circumvent the need for complex input design, lowering the threshold for attacks significantly. These results underscore the importance of addressing Web fraud attacks, providing new insights into MAS safety.
Post-training LLMs with Reinforcement Learning, specifically Group Relative Policy Optimization (GRPO), has emerged as a paradigm for enhancing mathematical reasoning. However, standard GRPO relies on scalar correctness rewards that are often non-injective with respect to semantic content: distinct reasoning paths receive identical rewards. This leads to a Diversity-Quality Inconsistency, where the policy collapses into a narrow set of dominant modes while ignoring equally valid but structurally novel strategies.To bridge this gap, we propose Diversity-aware Reward Adjustment (DRA), a theoretically grounded framework that calibrates the reward signal using the semantic density of sampled groups. By leveraging Submodular Mutual Information (SMI), DRA implements an Inverse Propensity Scoring (IPS) mechanism that effectively de-biases the gradient estimation. This creates a repulsive force against redundancy, driving the policy to achieve better coverage of the high-reward landscape.Our method is plug-and-play and integrates seamlessly with GRPO variants. Empirical evaluations on five math benchmarks demonstrate that DRA-GRPO consistently outperforms strong baselines, achieving an average accuracy of 58.2% on DeepSeek-R1-Distill-Qwen-1.5B with only 7,000 training samples and 55 cost, highlighting the critical role of diversity calibration in data-efficient alignment.
Large Language Models (LLMs) can generate factually inaccurate content even if they have corresponding knowledge, which critically undermines their reliability. Existing approaches attempt to mitigate this by incorporating uncertainty in QA prompt during training, but these numerical scores lack the semantic richness for LLM to properly understand its internal states of trustworthiness and honestness, leading to insufficient factuality alignment. We introduce FAITH (Factuality Alignment through Integrating Trustworthiness and Honestness), a post-training framework for factuality alignment that integrates natural-language uncertainty signals with external knowledge. Specifically, we augment training datasets by computing confidence scores and semantic entropy from LLM outputs and mapping them into a knowledge state quadrant that describes the model’s internal knowledge possession (trustworthiness) and answering behaviors (honestness) in natural language. Based on this enhanced data, we design a reward function that considers both correctness and uncertainty signals, and fine-tune the LLM using the Proximal Policy Optimization (PPO) algorithm. To further mitigate weakly grounded responses, we design a retrieval-augmented module that retrieves relevant external passages, improving the consistency between internal and external knowledge representations. Extensive experiments on four knowledge-intensive benchmarks demonstrate that FAITH enhances the factual accuracy and truthfulness of LLMs.
Large Language Models (LLMs) have demonstrated remarkable proficiency in code generation, yet their application to Property-Based Testing (PBT) remains fraught with a superficiality gap. While LLMs can readily generate syntactically correct tests, they often struggle to bridge the semantic gap between code implementation and its intended invariant logic, resulting in weak properties that provide a false sense of security. To address this, we introduce PROBE, an agentic framework that hardens software properties through Adversarial Refinement. Unlike traditional generation approaches, PROBE treats test generation as a game of semantic asymmetry: it employs a Validator agent to actively generate counter-implementations, which are semantically incorrect codes that satisfy the generated property, to expose loopholes in the specification. Furthermore, PROBE constructs a cross-functional semantic graph to capture deep dependencies often missed by local analysis. Extensive evaluation reveals that PROBE increases mutation scores by 9.79% over baselines. In real-world deployment, PROBE identified 45 previously unknown bugs in top-tier libraries that have been confirmed by developers, demonstrating its ability to uncover deep semantic defects.
Large language models (LLMs) are increasingly used in daily applications, from content generation to code writing, where each interaction treats the model as stateless, generating responses independently without memory. Yet human writing is inherently longitudinal: authors’ styles and cognitive states evolve across months and years. This raises a central question: can LLMs reproduce such temporal structure across extended time periods? We construct and publicly release a longitudinal dataset of 412 human authors and 6,086 documents spanning 2012–2024 across three domains (academic abstracts, blogs, news) and compare them to trajectories generated by three representative LLMs under standard and history-conditioned generation settings. Using drift and variance-based metrics over semantic, lexical, and cognitive–emotional representations, we find temporal flattening in LLM-generated text. LLMs produce greater lexical diversity but exhibit substantially reduced semantic and cognitive–emotional drift relative to humans. These differences are highly predictive: temporal variability patterns alone achieve 94% accuracy and 98% ROC-AUC in distinguishing human from LLM trajectories. Our results demonstrate that temporal flattening persists regardless of whether LLMs generate independently or with access to incremental history, revealing a fundamental property of current deployment paradigms. This gap has direct implications for applications requiring authentic temporal structure, such as synthetic training data and longitudinal text modeling.
Graphical User Interface (GUI) agents aim to automate a wide spectrum of human tasks by emulating user interaction. Despite rapid advancements, current approaches are hindered by several critical challenges: data bottleneck in end-to-end training, high cost of delayed error detection, and risk of contradictory guidance. Inspired by the human cognitive loop of Thinking, Alignment, and Reflection, we present D-Artemis—a novel deliberative framework in this paper. D-Artemis leverages a fine-grained, app-specific tip retrieval mechanism to inform its decision-making process. It also employs a proactive Pre-execution Alignment stage, where Thought-Action Consistency (TAC) Check module and Action Correction Agent (ACA) work in concert to mitigate the risk of execution failures. A post-execution Status Reflection Agent (SRA) completes the cognitive loop, enabling strategic learning from experience. Crucially, D-Artemis enhances the capabilities of general-purpose Multimodal large language models (MLLMs) for GUI tasks without the need for training on complex trajectory datasets, demonstrating strong generalization. D-Artemis achieves SOTA among open-source general models on AndroidWorld (75.8%) and ScreenSpot-V2 (96.8%). Extensive ablation studies further demonstrate the significant contribution of each proposed component.
CRPS: Curriculum Replay via Progressive Suffixes from Successful Trajectories for Long-Horizon LLM Agents
PDF ↗Long-horizon LLM agents trained with sparse terminal rewards tend to experience slow and unstable learning, and the issue is amplified by group-normalized on-policy objectives commonly used for LLM training (e.g., GRPO). When rollout groups collapse to nearly all failures early in training, within-group normalization yields degenerate advantages and weak learning signals. To address this, we propose Curriculum Replay via Progressive Suffixes from Successful Trajectories (CRPS), a lightweight RL-training strategy that turns serendipitous terminal successes into a within-trajectory curriculum. CRPS maintains a buffer of successful trajectories and restarts rollouts from suffix states, with an online controller adapting k to match agent competence and keep replay outcomes informative. Across ALFWorld and WebShop with different foundation models, CRPS consistently outperforms full-episode GRPO and naive experience replay. Group-level diagnostics further show that CRPS reduces degenerate groups ratio and increases within-group outcome diversity, aligning with faster and more stable training.
Graph-based Retrieval-Augmented Generation (GraphRAG) organizes external knowledge as a hierarchical graph, enabling efficient retrieval and aggregation of scattered evidence across multiple documents. However, many existing benchmarks for GraphRAG rely on short, curated passages as external knowledge, failing to adequately evaluate systems in realistic settings involving long contexts and large-scale heterogeneous documents. To bridge this gap, we introduce , a benchmark designed to assess GraphRAG performance in the wild. We leverage Wikipedia’s unique structure, where cohesive narratives are grounded in long and heterogeneous external reference documents, to construct a benchmark reflecting real-word scenarios. Specifically, we sample articles across 12 top-level topics, using their external references as the retrieval corpus and citation-linked statements as ground truth, resulting in 1,100 questions spanning three levels of complexity: single-fact QA, multi-fact QA, and section-level summarization. Experiments across multiple baselines reveal that current GraphRAG pipelines help on multi-fact aggregation when evidence comes from a moderate number of sources, but this aggregation paradigm may overemphasize high-level statements at the expense of fine-grained details, leading to weaker performance on summarization tasks.
Large Reasoning Models(LRMs) have achieved significant breakthroughs over prior large language models (LLMs), but they also entail greater potential safety risks. Existing alignment methods often remain at a shallow level of protection, making them insufficient to address deeper risks and strategic attacks in complex reasoning processes. To bridge this gap, we move beyond the conventional paradigm that treats safety alignment merely as a preventive measure to reduce harmful outputs. Drawing inspiration from human-like introspection and self-correction, we propose Self-Reflection, a technique that introduces a special Self-Reflection token, enabling LRMs to perform Self-Reflection during generation and recover from harmful outputs. Our approach integrates seamlessly into standard post-training paradigms , further enhancing both helpfulness and safety. The experimental results demonstrate that models trained with Self-Reflection not only consistently outperform the baseline in terms of safety (reducing the HCR from 13.8% to 4.1%, nearly a threefold improvement over mainstream approaches), but also achieve substantial advantages in both helpfulness and the safety–helpfulness balance. More importantly, under evaluations involving various adversarial attacks, including a specially designed adaptive attack, the Self-Reflection mechanism significantly enhances model safety without targeted adversarial training.Notice: This paper contains harmful content.