Retrieval-Augmented Generation (RAG) has demonstrated strong effectiveness in knowledge-intensive tasks by grounding language generation in external evidence. Despite its success, many existing RAG systems are built based on a ranking-centric, asymmetric dependency paradigm, where the generation quality of the generator is highly dependent on reranking results of the reranker.To overcome this limitation, we propose Cooperative Retrieval-Augmented Generation (CoRAG), a framework that treats the reranker and the generator as peer decision-makers rather than being connected through an asymmetric dependency pipeline. By jointly optimizing their behaviors toward a shared task objective, the reranker and generator are encouraged to cooperate, ensuring that document reranking and generation work in concert to improve the final response.Experimental results demonstrate good generalization and improved generation stability of CoRAG, even when the model is trained on only around 10K PopQA samples. Our model released in https://github.com/CoderrrSong/CoRAG
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
Grammatical Error Correction (GEC) for Urdu remains an under-researched area due to the lack of annotated datasets. This paper addresses the challenge of generating a robust corpus for fine-tuning deep learning models aimed at Urdu GEC. We propose a method for synthesizing a large dataset by collecting errors from the Urdu WikiEdits history, learning from them, and inserting similar errors in grammatically correct sentences to generate incorrect sentences with grammatical errors, hence creating a pair of grammatically correct and incorrect sentences. We introduce UrduGEC-Synthetic, a synthetically generated dataset produced through this pipeline. Furthermore, we introduce UrduGEC-Gold, a Gold Dataset by extracting errors from exam copies of students. Finally, we also fine-tuned various models on UrduGEC-Synthetic and evaluated them against UrduGEC-Gold to show the quality of synthetic data generation.
Human experts tackle difficult math problems by identifying and executing a few pivotal steps rather than listing every intermediate thought. In contrast, standard Chain-of-Thought (CoT) distillation trains small models on lengthy reasoning traces, encouraging a uniform overthinking style across easy and hard items alike. The result is rigid, slow solutions that sacrifice adaptivity. This approach stands in sharp contrast to human intuition. Humans naturally adapt their problem-solving strategy, dedicating significant effort to difficult problems while finding quick, simple solutions for easier ones. We argue that the root cause lies in the training data: it contains excess information and reasoning steps organized in ways misaligned with human practice. We address this with Difficulty-Aware Distillation(DAD), a procedure for producing training data that mirrors concise human reasoning. A large teacher model first assesses a problem’s difficulty and then rewrites the solution to retain only the essential steps. Using this process, we constructed LiteCoT, a 100,000-example corpus of short, clear rationales, and used it to train our Liter models. With 100k LiteCoT, we outperform models trained on 800k long CoT and cut both training and inference costs. The advantage is consistent across standard math benchmarks, showing that concise, human-aligned data delivers equal or better accuracy with much less compute. For example, on the challenging AIME24 exam, our approach reaches 74.2% Pass@1 using only about 5K inference tokens, surpassing other methods that consume many more tokens.
Standard in-context learning (ICL) assumes identical output spaces between test and retrieval datasets (fully aligned). However, in practice, these datasets can be fully aligned, partially aligned, or fully disjoint in label space (Output space), forming an information continuum from rich to scarce. Naive ICL often becomes ineffective under such mismatches. In this work, we challenge this assumption by demonstrating that the retrieval dataset need not perfectly align with the test dataset, as long as it remains related to the target task. We propose Task-Related In-Context Learning (TRICL), a unified framework for ICL under output-space mismatch, designed to cover the full continuum of scenarios. TRICL first identifies demonstrations in the mismatched retrieval dataset that are relevant to the test label space via a lightweight Bayesian probabilistic criterion, and uses them to form a related dataset. TRICL then perform ICL on the related dataset to obtain preliminary predictions; finally, TRICL leverage these intermediate predictions to reduce and transform the output space of the original test task, thereby improving the performance of LLMs. Even in the most information-scarce fully disjoint scenario, as long as the retrieval dataset is task-related to the test task, TRICL achieves state-of-the-art (SOTA) results across three LLMs, three task types, and four datasets. Moreover, TRICL remains effective in the fully aligned and partially aligned scenarios, consistently yielding strong gains over competitive baselines. Moreover, TRICL also extends to generative task.
Speculative decoding is a widely used technique for accelerating inference in large language models (LLMs), but its performance degrades as input length grows, with significant drops even at moderate lengths. Yet, this early degradation has remained largely underexplored. We introduce SpecExtend, a drop-in enhancement that improves speculative decoding on long sequences without additional training. SpecExtend integrates efficient attention mechanisms such as FlashAttention and Hybrid Tree Attention to accelerate prefill and verification steps. To improve both draft accuracy and speed on long inputs without retraining, we propose Cross-model Retrieval, a novel KV cache eviction strategy that leverages the target model’s attention scores to dynamically select relevant context for the smaller draft model. Extensive evaluations show that SpecExtend accelerates speculative decoding by up to 2.84× on 16K-token long document summarization and up to 3.86× on long-form reasoning, while preserving the short-input performance of state-of-the-art frameworks.
Legal case retrieval remains challenging due to the complexity of legal language and the need for precise lexical alignment between queries and relevant cases. Although dense retrieval models have achieved notable progress, empirical studies show that BM25 continues to serve as a strong baseline in this domain. It motivates us to propose a self-evolving framework for rule-driven query rewriting that enhances BM25 without any parameter training. The framework equips an LLM-based agent with an automatic evaluation environment, enabling it to iteratively create rewriting rules, plan validation experiments over rule combinations, and eliminate ineffective rules based on historical feedbacks. We evaluate our method on the Chinese legal case retrieval benchmark LeCaRD-v2. Experimental results demonstrate that the proposed framework outperforms non-evolutionary baselines, including human-designed rules and greedy rule selection, particularly when powered by a high-capacity core LLM. We also conduct detailed analyses to investigate the mechanisms underlying self-evolution. Our findings reveal that LLM’s capabilities to leverage previous experimental results and its intrinsic knowledge of rule elimination play critical roles in refining the rule set via self-evolution.
Understanding financial documents is critical for high-stakes decision-making yet hindered by systemic semantic implicitness: key facts are rarely explicit in surface text and often determined by global structural cues. Missing these cues invites semantic misinterpretations, such as misreading what a number refers to, an outcome unacceptable in high-stakes environments. However, existing Retrieval-Augmented Generation (RAG) systems typically treat structure as a physical navigational skeleton rather than intrinsic semantic knowledge. To address this, we introduce Fin-STAR (Financial STructure-As-Semantics Retrieval), a framework redefining hierarchy as intrinsic semantics. Fin-STAR incorporates a novel Structure-Enriched Semantic Indexing mechanism that augments the hierarchical lineage with snippet-derived virtual nodes, and injects this enriched context via a semantic cross-attention paradigm, rendering implicit cues explicit. By grounding evidence within its structural scope, we preserve factual invariance and ensure contextual integrity. Addressing the lack of granular public datasets, we conduct experiments on FinTierQA Gold, a curated expert benchmark. Results show that Fin-STAR outperforms state-of-the-art hierarchical and graph-based baselines across diverse query complexities, document types, and markets. Notably, ablations confirm that our semantic injection consistently outperforms alternative strategies. Finally, we release FinTierQA, comprising 3.9M pairs automatically constructed from 78k documents via our framework .
Providing accurate reward signals for code generated by large language models (LLMs) is a significant challenge in applying reinforcement learning (RL) to code generation. Existing methods rely on unit tests to evaluate code correctness and provide rewards, which are hindered by the difficulty of acquiring and verifying reliable unit tests at scale. In this work, we propose CodeRM-NT, a code reward model with no reliance on unit tests. Our method leverages Monte Carlo Tree Search guided by LLMs to generate code snippets and judges execution traces to annotate code with reward signals. We use the rewards to train CodeRM-NT that is capable of providing rewards for code during RL. CodeRM-NT also facilitates curriculum learning by scoring and sorting training samples based on their difficulty. Experimental results demonstrate that training with CodeRM-NT consistently outperforms synthetic unit test-based rewards, yielding superior performance on multiple code generation benchmarks. Additionally, curriculum learning based on CodeRM-NT further enhances model performance. Our code and dataset are available at: https://github.com/THUDM/CodeRM-NT.
From Experience to Skill: Multi-Agent Generative Engine Optimization via Reusable Strategy Learning
PDF ↗Generative engines (GEs) are reshaping information access by replacing ranked links with citation-grounded answers, yet current Generative Engine Optimization (GEO) methods optimize each instance in isolation, unable to accumulate or transfer effective strategies across tasks and engines. We reframe GEO as a strategy learning problem and propose MAGEO, a multi-agent framework in which coordinated planning, editing, and fidelity-aware evaluation serve as the execution layer, while validated editing patterns are progressively distilled into reusable, engine-specific optimization skills. To enable controlled assessment, we introduce a Twin Branch Evaluation Protocol for causal attribution of content edits and DSV-CF, a dual-axis metric that unifies semantic visibility with attribution accuracy. We further release MSME-GEO-Bench, a multi-scenario, multi-engine benchmark grounded in real-world queries. Experiments on three mainstream engines show that MAGEO substantially outperforms heuristic baselines in both visibility and citation fidelity, with ablations confirming that engine-specific preference modeling and strategy reuse are central to these gains, suggesting a scalable learning-driven paradigm for trustworthy GEO. Code is available at https://github.com/Wu-beining/MAGEO.
Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models
PDF ↗Historical documents act as invaluable knowledge archives but often suffer from illegibility due to physical deterioration and damage. While existing restoration methods based on masked language modeling effectively utilize local context, they struggle to restore named entities that require external historical knowledge. To address this limitation, we introduce a novel framework for historical document restoration that leverages large language models with retrieval-augmented generation (RAG). By combining the implicit knowledge of pre-trained LLMs with explicitly retrieved external context, our model ARI effectively mitigates the challenge of inferring context-dependent proper nouns. Extensive experiments on Korean historical documents demonstrate that our approach significantly outperforms baselines, achieving substantial gains in restoring both general characters and named entities. Furthermore, comprehensive evaluations including expert assessments confirm that ARI serves as a practical tool for domain experts, promising to accelerate the analysis of historical records.
Hallucinations in Speech Large Language Models (SpeechLLMs) pose significant risks, yet existing detection methods typically rely on gold-standard outputs that are costly or impractical to obtain. Moreover, hallucination detection methods developed for text-based LLMs do not directly capture audio-specific signals. We investigate four attention-derived metrics: AudioRatio, AudioConsistency, AudioEntropy, and TextEntropy, designed to capture pathological attention patterns associated with hallucination, and train lightweight logistic regression classifiers on these features for efficient inference-time detection. Across automatic speech recognition and speech-to-text translation tasks, evaluations on Qwen-2-Audio and Voxtral-3B show that our approach outperforms uncertainty-based and prior attention-based baselines on in-domain data, achieving improvements of up to +0.23 PR-AUC, and generalises to out-of-domain ASR settings. We further find that strong performance can be achieved with approximately 100 attention heads, improving out-of-domain generalisation compared to using all heads. While effectiveness is model-dependent and task-specific training is required, our results demonstrate that attention patterns provide a valuable tool for hallucination detection in SpeechLLMs
Modern Text-to-SQL systems generate multiple candidate SQL queries and rank them to judge a final prediction. However, existing methods face two limitations. First, they often score functionally equivalent SQL queries inconsistently despite identical execution results. Second, ranking cannot recover when the correct SQL is absent from the candidate pool. We propose R3 -SQL, a Text-to-SQL framework that addresses both issues through unified reward for ranking and resampling. R3 -SQL first groups candidates by execution result and ranks groups for consistency. To score each group, it combines a pairwise preference across groups with a pointwise utility from the best group rank and size, capturing relative preference, consistency, and candidate quality. To improve candidate recall, R3 -SQL introduces agentic resampling, which judges the generated candidate pool and selectively resamples when the correct SQL is likely absent. R3 -SQL achieves 75.03 execution accuracy on BIRD-dev, a new state of the art among methods using models with disclosed sizes, with consistent gains across five benchmarks.
Masked Diffusion Language Models (DLMs) have recently emerged as a promising alternative to traditional Autoregressive Models (ARMs). DLMs employ transformer encoders with bidirectional attention, enabling parallel token generation while maintaining competitive performance. Although their efficiency and effectiveness have been extensively studied, the internal mechanisms that govern DLMs remain largely unexplored. In this work, we conduct an empirical analysis of DLM attention patterns, focusing on the attention sinking phenomenon, an effect previously observed in various transformer-based architectures. Our findings reveal that DLMs also exhibit attention sinks, but with distinct characteristics. First, unlike in ARMs, the sink positions in DLMs tend to shift throughout the generation process, displaying a dynamic behaviour. Second, while ARMs are highly sensitive to the removal of attention sinks, DLMs remain robust: masking sinks leads to only a minor degradation in performance. These results provide new insights into the inner workings of diffusion-based language models and highlight fundamental differences in how they allocate and utilize attention compared to autoregressive models.
This work addresses test output prediction, a key challenge in test case generation. To improve the reliability of predicted outputs by LLMs, prior approaches generate code first to ground predictions. One grounding strategy is direct execution of generated code, but even minor errors can cause failures. To address this, we introduce LLM-based pseudocode execution, which grounds prediction on more error-resilient pseudocode and simulates execution via LLM reasoning. We further propose DUET, a dual-execution framework that combines both approaches by functional majority voting. Our analysis shows the two approaches are complementary in overcoming the limitations of direct execution suffering from code errors, and pseudocode reasoning from hallucination. On LiveCodeBench, DUET achieves the state-of-the-art performance, improving Pass@1 by 13.6 pp. For filtering candidates in code generation, DUET shows the best Pass@1 on LiveCodeBenchEasy, BigCodeBench-Hard, DevEval and HumanEval(+).
Large language models (LLMs) frequently produce source code that seems correct and well-formed, yet includes hallucinated elements that cause downstream test failures. In this study, we benchmark state-of-the-art uncertainty quantification methods and existing baselines for the task of hallucination detection in source code and introduce a diff-based pipeline to construct a code dataset annotated with line-level hallucinations. Building on this, we train a lightweight Transformer-based detector that uses LLM internal representations to identify hallucinations, substantially outperforming existing methods across several code generation domains. The detector also shows particular promise for enabling self-correction in LLM-based coding agents. We release the first publicly available dataset of line-level code hallucinations, along with the corresponding source code and trained hallucination detectors https://github.com/datapaf/CodeHallucinationDetection
The Confidence Paradox: Unveiling the Latent Discriminative Power of Diffusion Large Language Models in Mathematical Reasoning
PDF ↗Diffusion large language models (DLLMs) have emerged as a promising alternative to autoregressive (AR) generation, uniquely offering token-level probabilities under bidirectional context. However, the semantics of their native uncertainty estimates remain underexplored. In this work, we uncover a calibration paradox inherent to the bidirectional generation mechanism of state-of-the-art DLLMs. Concretely, we demonstrate that diffusion confidence is structurally distinct from AR likelihood. Notably, LLaDA-8B is highly miscalibrated (31.2% ECE) on mathematical reasoning benchmarks, yet possesses superior discriminative power (0.826 AUROC), significantly outperforming comparable AR baselines in single-pass settings (0.611 AUROC). We diagnose that this paradox arises because diffusion confidence functions less like a probability of correctness and more like a proxy for structural consistency enabled by the model’s bidirectional access to the entire solution path. We further show that lightweight post-hoc calibration can reconcile this gap, reducing ECE by over 60% while preserving the strong ranking signal. Our findings suggest that DLLMs offer a unique, cost-efficient uncertainty signal for reasoning tasks that complements expensive AR approaches.
Large language models (LLMs), like human language learners, show patterns influenced by their dominant training language. Just as humans display language patterns influenced by their native tongue (semantic accents) when learning new languages, LLMs often default to English-centric responses even when generating in other languages. However, we observe that explicitly providing cultural context in prompts significantly improves the models’ ability to generate culturally localized responses. We term this phenomenon the explicit-implicit localization gap, indicating that while cultural knowledge exists within LLMs, it may not naturally surface in multilingual interaction without explicitly including cultural context. In this paper, we (1) quantify this gap in multiple LLMs using a new cultural localization benchmark and find large (>10%) gaps in the majority of investigated models. (2) Demonstrate a fundamental trade-off between localization accuracy and output diversity. (3) Through mechanistic interpretability, we identify the underlying localization mechanisms within LLMs and show that these mechanisms are both language and task agnostic, with individual steering vectors effectively generalizing across different languages and culturally-relevant tasks.
In fine-grained sparse Mixture-of-Experts (MoE) models, a large pool of specialized experts replaces a small homogeneous set, shifting performance and throughput to be governed by inference-time expert activation. Yet most existing optimization recipes implicitly assume a fixed activation budget (e.g., a constant Top-k per layer), whose behavior in fine-grained MoEs is poorly understood. We first characterize runtime skipping strategies, quantifying the accuracy–efficiency trade-off of (i) uniform fixed activation and (ii) static layer-wise Top-k allocation found by search. Our analysis reveals that static skipping can already provide substantial throughput gains, but optimal static schedules vary significantly across models and routing mechanisms. We therefore introduce Adaptive Skipping with Entropy-Penalized Thresholding (ASET), a training-free policy that adapts token-level activation using router confidence and entropy while remaining within the model’s original budget. Across the fine-grained MoEs we study, static skipping policies yield 10–78% throughput gains with minimal performance degradation, including \ge10% improvement on DeepSeek-V3 without measurable loss. On the OLMoE testbed, ASET yields a Pareto frontier between average activation and task quality. Overall, these results identify expert skipping as a practical lever for faster fine-grained MoE inference, with adaptive activation helping when fixed budgets are too rigid.
Medical report generation from medical images is a vital AI task that helps doctors with diagnosis and marks a significant step toward creating general AI-powered medical systems. However, previous methods either fail to optimize factual accuracy or heavily depend on expert preference data. To overcome these challenges, we propose MedQPA, an automatic and generalizable report evaluation technique that uses question proposing and answering to enable controllable, structured reasoning grounded in medical domain knowledge and the factual correctness of the report. Additionally, we design MedQPA-Gen, a medical report generation pipeline that maximizes the MedQPA score through prompt engineering and reinforcement learning with MedQPA as a reward signal. We demonstrate that MedQPA is an accurate evaluation metric that closely correlates with human preferences. More importantly, MedQPA-Gen achieves higher human preference scores and better performance on downstream tasks. We open-source code at this repo https://github.com/MedQPA-gen/MedQPA-gen.
Chain-of-Relations: Faithful and Efficient LLM Reasoning over Knowledge Graphs via Relation-Centric Exploration
PDF ↗Knowledge graph question answering (KGQA) serves as an essential benchmark for KG-enhanced large language models. Among various approaches, agent-based methods have emerged as an effective solution.Existing methods adopt entity-centric exploration that incrementally constructs reasoning paths by selecting and connecting intermediate entities. However, they face two critical limitations. (1) Entity incompleteness vulnerability arises when some intermediate entities lack semantic information beyond opaque IDs, preventing relevance evaluation and leading to discarding valid reasoning paths.(2) Premature entity pruning occurs because beam search retains only top-ranked entities at each step, eliminating candidates before their relevance can be verified.To address these challenges, this paper proposes Chain-of-Relations (CoR) with relation-centric exploration and global entity filtering, reducing dependence on entity completeness and ensuring complete candidate retrieval before constraint validation.Experiments on three benchmark datasets show that CoR consistently outperforms strong baselines in both F1 score and KG-grounded Rate.