High-stakes domains such as finance, law, and biomedicine demand both accurate results and rigorous reasoning. Current reinforcement learning paradigms primarily rely on outcome-based rewards, often overlooking latent logical fallacies in intermediate steps. Leveraging the cognitive asymmetry where falsifying local errors is more efficient than generating global correctness, we propose RADO (Reasoning Audit-Driven Optimization). RADO introduces a specialized audit model augmented with external tools to identify local logical ruptures and calibrate reward signals. By integrating Direct Preference Optimization (DPO) with Group Relative Policy Optimization (GRPO), our framework enables explicit supervision over reasoning paths. Experimental results demonstrate that RADO consistently improves final accuracy while significantly enhancing logical rigor in high-stakes domains.
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
Knowledge distillation is crucial for compressing Large Language Models (LLMs), enabling smaller student models to learn from larger teacher models. However, existing LLM distillation methods overly rely on student-generated outputs, which may introduce generation errors and misguide the distillation process. Moreover, existing distillation loss functions struggle to align the most informative part due to the complex output distributions of LLMs. To address these problems, we propose a multi-granularity semantic revision method for LLM distillation. At the sequence level, we propose a sequence correction and re-generation (SCRG) strategy. SCRG identifies error tokens by calculating the semantic cognitive difference between teacher and student outputs, corrects them using teacher-generated tokens, and re-generates the sequence to minimize errors. At the token level, we design a distribution adaptive clipping Kullback-Leibler (DAC-KL) loss, which uses a learnable sub-network to focus on semantically dense areas of the teacher’s output, reducing the impact of redundant information. At the span level, we utilize span priors to compute probability correlations within sequences, ensuring consistency between teacher and student outputs to enhance semantic information transfer. Extensive experiments on models ranging from 0.1B to 13B parameters demonstrate the effectiveness of our approach compared to existing methods.
As large language models (LLMs) increasingly tackle complex reasoning tasks, test-time scaling has become critical for enhancing capabilities. However, in agentic scenarios with frequent tool calls, the traditional generation-length-based definition breaks down: tool latency decouples inference time from generation length. We propose Timely Machine, redefining test-time as wall-clock time, where models dynamically adjust strategies based on time budgets. We introduce Timely-Eval, a benchmark spanning high-frequency tool calls, low-frequency tool calls, and time-constrained reasoning. By varying tool latency, we find smaller models excel with fast feedback through more interactions, while larger models dominate high-latency settings via superior interaction quality. Moreover, existing models fail to adapt reasoning to time budgets. We propose Timely-RL to address this gap. After cold-start supervised fine-tuning, we use reinforcement learning to enhance temporal planning. Timely-RL improves time budget awareness and consistently boosts performance across Timely-Eval. We hope our work offers a new perspective on test-time scaling for the agentic era.
We present SemanticQA, an evaluation suite designed to assess language models (LMs) in semantic phrase processing tasks. The benchmark consolidates existing multiword expression (MwE) resources and reorganizes them into a unified testbed. It covers both general lexical phenomena, such as lexical collocations, and three fine-grained categories: idiomatic expressions, noun compounds, and verbal constructions. Through SemanticQA, we assess LMs of diverse architectures and scales in extraction, classification, and interpretation tasks, as well as sequential task compositions. We reveal substantial performance variation, particularly on tasks requiring semantic reasoning, highlighting differences in reasoning efficacy and semantic understanding of LMs, providing insights for pushing LMs with stronger comprehension on non-trivial semantic phrases. The evaluation harness and data of SemanticQA are available at https://github.com/jacklanda/SemanticQA.
Traditional tabular data synthesis methods often overlook the cross-modal heterogeneity of real-world tables, where structured continuous and discrete attributes coexist with unstructured long-text columns. Existing synthesis approaches struggle to simultaneously achieve accurate statistical fidelity for non-textual attributes and consistent semantic constraints between textual and non-textual attributes. In this work, we establish the first benchmark for long-text tabular data synthesis and introduce a novel metric, termed Textual Column Correlation Fidelity (TCCF), to quantify cross-modal semantic alignment. We propose AFT-Tab, an adversarial fine-tuning framework that synergistically trains an LLM-based text generator and a deep-learning-based non-textual generator. Through a dual-feedback mechanism guided by an LLM discriminator, AFT-Tab ensures both precise statistical distributions and rigorous semantic constraints. Experimental results show that AFT-Tab significantly outperforms state-of-the-art baselines in statistical fidelity, TCCF, diversity, and downstream task utility.
Open Schrödinger’s Closed Box: Identifying Retrieval Augmented Generation in API-Accessible Large Language Model Services
PDF ↗Large language models (LLMs) are powerful at question-answering but prone to hallucinations due to limited domain-specific or up-to-date knowledge. Retrieval augmented generation (RAG) mitigates this by adding an external retriever and knowledge database, yet RAG remains vulnerable to targeted attacks that degrade outputs or manipulate opinions. Prior attacks typically assume adversaries know the service is RAG-enhanced and may even know deployment details, an assumption often invalid for real-world commercial LLMs that expose only black-box APIs.This opacity also risks misleading users about system capabilities. This work aims to bridge this gap by proposing RAG-ID, a framework for \underline{\text{ID}}entifying \underline{\text{RAG}} properties in LLM services.We classify adversaries into three knowledge levels and design six attack methods. Experiments show these attacks reliably detect RAG — up to 99.97% accuracy with partial or no optional knowledge, and nearly 100% when the LLM and database are known. After detection, RAG-ID can infer finer RAG properties (e.g., deployed LLM and knowledge database). We consider RAG-ID a reconnaissance tool for attackers, a way to facilitate users’ transparent selection of LLM services, and a guide for RAG developers in refining security measures.
Adapting large language models (LLMs) to low-resource languages (LRLs) is constrained by the scarcity of task data and computational resources. Although Proxy Tuning offers a logit-level strategy for introducing scaling effects, it often fails in LRL settings because the large model’s weak LRL competence might overwhelm the knowledge of specialized smaller models. We thus propose TriMix, a test-time logit fusion framework that dynamically balances capabilities from three different sources: LRL competence from a continually pretrained small model, task competence from high-resource language instruction tuning, and the scaling benefits of large models. It is data- and compute-efficient, requiring no LRL task annotations, and only continual pretraining on a small model. Experiments across four model families and eight LRLs show that TriMix consistently outperforms single-model baselines and Proxy Tuning. Our analysis reveals that prioritizing the small LRL-specialized model’s logits is crucial for success, challenging the prevalent large-model-dominant assumption.
Reinforcement Learning with Verifiable Rewards (RLVR) has markedly enhanced the reasoning abilities of large language models (LLMs). Its success, however, largely depends on strong base models with rich world knowledge, yielding only modest improvements for small-size language models (SLMs). To address this limitation, we investigate Guided GRPO, which injects ground-truth reasoning steps into roll-out trajectories to compensate for SLMs’ inherent weaknesses. Through a comprehensive study of various guidance configurations, we find that naively adding guidance delivers limited gains. These insights motivate G^2RPO-A, an adaptive algorithm that automatically adjusts guidance strength in response to the model’s evolving training dynamics. Experiments on mathematical reasoning and code-generation benchmarks confirm that G^2RPO-A substantially outperforms vanilla GRPO. Our code and models at available at https://github.com/T-Lab-CUHKSZ/G2RPO-A.
Large Language Models (LLMs) have significantly advanced Machine Translation (MT), applying them to linguistically complex domains-such as Social Network Services, literature etc. In these scenarios, translations often require handling non-literal expressions, leading to the inaccuracy of MT metrics. To systematically investigate the reliability of MT metrics, we first curate a meta-evaluation dataset focused on non-literal translations, namely MENT. MENT encompasses four non-literal translation domains and features source sentences paired with translations from diverse MT systems, with 7,530 human-annotated scores on translation quality. Experimental results reveal the inaccuracies of traditional MT metrics and the limitations of LLM-as-a-Judge, particularly the knowledge cutoff and score inconsistency problem. To mitigate these limitations, we propose RATE, a novel agentic translation evaluation framework, centered by a reflective Core Agent that dynamically invokes specialized sub-agents. Experimental results indicate the efficacy of RATE, achieving an improvement of at least 3.2 points in combined system- and segment-level correlation with human judgments compared with current methods. Further experiments demonstrate the robustness of RATE to general-domain MT evaluation. Code and dataset are available at: https://github.com/BITHLP/RATE.
Scaling Beyond Context: A Survey of Multimodal Retrieval-Augmented Generation for Document Understanding
PDF ↗Document understanding is critical for applications from financial analysis to scientific discovery. Current approaches, whether OCR-based pipelines feeding Large Language Models (LLMs) or native Multimodal LLMs (MLLMs), face key limitations: the former loses structural detail, while the latter struggles with context modeling. Retrieval-Augmented Generation (RAG) helps ground models in external data, but documents’ multimodal nature, i.e., combining text, tables, charts, and layout, demands a more advanced paradigm: Multimodal RAG. This approach enables holistic retrieval and reasoning across all modalities, unlocking comprehensive document intelligence. Recognizing its importance, this paper presents a systematic survey of Multimodal RAG for document understanding. We propose a taxonomy based on domain, retrieval modality, and granularity, and review advances involving graph structures and agentic frameworks. We also summarize key datasets, benchmarks, and applications, and highlight open challenges in efficiency, fine-grained representation, and robustness, providing a roadmap for future progress in document AI.
As Large Language Models (LLMs) are increasingly deployed in real-world settings, correctness alone is insufficient. Reliable deployment requires maintaining truthful beliefs under contextual perturbations. Existing evaluations largely rely on point-wise confidence like Self-Consistency, which can mask brittle belief. We show that even facts answered with perfect self-consistency can rapidly collapse under mild contextual interference. To address this gap, we propose Neighbor-Consistency Belief (NCB), a structural measure of belief robustness that evaluates response coherence across a conceptual neighborhood. To validate the efficiency of NCB, we introduce a new cognitive stress-testing protocol that probes outputs stability under contextual interference. Experiments across multiple LLMs show that the performance of high-NCB data is relatively more resistant to interference. Finally, we present Structure-Aware Training (SAT), which optimizes context-invariant belief structure and reduces long-tail knowledge brittleness by approximately 30%.
The prevalence of fake news on social media calls for automated fact-checking systems that deliver not only accurate verdicts but also faithful explanations. However, existing large language model (LLM)-based methods often overlook deceptive misinformation styles in generated explanations, producing unfaithful rationales that may mislead human judgment. They also rely heavily on external knowledge sources, which can introduce hallucinations and incur substantial latency, undermining both reliability and responsiveness in real-time settings. To address these limitations, we propose REason-guided Fact-checking with Latent EXplanations (REFLEX), a self-refining framework that explicitly controls reasoning style by anchoring explanations to the predicted verdict. REFLEX leverages self-disagreement veracity signals between a backbone model and its fine-tuned variant to construct steering vectors, thereby naturally disentangling factual content from stylistic cues. Experiments on a real-world benchmark show that REFLEX achieves state-of-the-art performance under LLaMA-series models using only 465 self-refined samples. Owing to its transferability, REFLEX also yields gains of up to 7.54 Macro-F1 points on in-the-wild data. Further analysis shows that our method effectively mitigates faithful hallucination, leading to both more reliable explanations and more accurate verdicts than prior explainable fact-checking approaches.
Modern software development demands code that is maintainable, testable, and scalable by organizing the implementation into modular components with iterative reuse of existing codes. We formalize this iterative, multi-turn paradigm as codeflow and introduce CodeFlowBench, the first benchmark designed to comprehensively evaluate LLMs’ ability to perform codeflow - implementing new functionality by reusing existing functions over multiple turns. CodeFlowBench comprises two complementary components: CodeFlowBench-Comp, a core collection of 5,000+ competitive programming problems from Codeforces updated via an automated pipeline and CodeFlowBench-Repo, which is sourced from GitHub repositories to better reflect real-world scenarios. Furthermore, a novel evaluation framework featured dual assessment protocol and structural metrics derived from dependency trees is introduced. Extensive experiments reveal significant performance degradation in multi-turn codeflow scenarios. Furthermore, our in-depth analysis illustrates that model performance inversely correlates with dependency complexity. These findings not only highlight the critical challenges for supporting real-world workflows, but also establish CodeFlowBench as an essential tool for advancing code generation research.
Improving the exploration of reasoning is essential for advancing Large Language Models’ (LLMs) problem-solving performance. Current methods primarily rely on output-level stochasticity, which decode within fixed reasoning patterns of LLM and suffer from insufficient exploration. In this paper, we introduce adjusting attention temperature to directly modulate the model’s internal focus during reasoning, which enables a dynamic shift between exploratory and focused processing. We reveal that moderate adjustments preserve LLM’s reasoning capability while producing problem hardness-dependent benefits: higher temperatures facilitate solving complex tasks by encouraging wider exploration, whereas lower temperatures mitigate overthinking on simpler problems. Leveraging this insight, we propose a two-stage inference strategy: first, attention temperature scaling modulates the LLM’s reasoning patterns to diversify the reasoning traces; then, a difficulty-aware aggregation scheme is introduced to effectively identify the most reliable solution from the generated candidates. Extensive evaluations show that our method improves Pass@10 by 6.78–14.20% and aggregation accuracy by 9.74% across 7 reasoning benchmarks.
Large language models with search capabilities frequently exhibit miscalibrated confidence, producing incorrect answers with high certainty. We present Deliberative Searcher, a reasoning-primary framework that integrates search operations into chain-of-thought generation while maintaining explicit confidence calibration. Our method employs constrained reinforcement learning with adaptive Lagrangian multipliers to jointly optimize correctness and reliability. Experiments across five benchmarks demonstrate substantial improvements: our 7B model reduces average false-certain rates from 54% in baselines to 2%, while our 72B variant achieves competitive accuracy with closed-source models and reduces false-certain rates to 9%. The well-calibrated confidence scores also enable more efficient test-time compute: instead of standard majority voting, we use confidence-weighted aggregation and match the performance of 16-sample majority voting with only 4 samples, a 4\times reduction in inference compute. These results establish calibrated confidence as a foundation for both trustworthy outputs and adaptive test-time compute, demonstrating the value of the proposed constrained RL framework in search-augmented language models.
Large language models have improved dialogue systems, but often process conversational turns in isolation, overlooking the event structures that guide natural interactions. Hence we introduce EventWeave, a framework that explicitly models relationships between conversational events to generate more contextually appropriate dialogue responses. EventWeave constructs a dynamic event graph that distinguishes between core events (main goals) and supporting events (interconnected details), employing a multi-head attention mechanism to selectively determine which events are most relevant to the current turn. Unlike summarization or standard graph-based approaches, our method captures three distinct relationship types between events, allowing for more nuanced context modeling. Experiments on three dialogue datasets demonstrate that EventWeave produces more natural and contextually appropriate responses while requiring less computational overhead than models processing the entire dialogue history. Ablation studies confirm improvements stem from better event relationship modeling rather than increased information density. Our approach effectively balances comprehensive context understanding with generating concise responses, maintaining strong performance across various dialogue lengths through targeted optimization techniques.
Between a Rock and a Hard Place: The Tension Between Ethical Reasoning and Safety Alignment in LLMs
PDF ↗Large Language Model safety alignment predominantly operates on a binary assumption that requests are either safe or unsafe. This classification proves insufficient when models encounter ethical dilemmas, where the capacity to reason through moral trade-offs creates a distinct attack surface. We formalize this vulnerability through TRIAL, a multi-turn red-teaming methodology that embeds harmful requests within ethical framings. TRIAL achieves consistently high attack success rates across models by exploiting the model’s own ethical reasoning to frame harmful actions as morally necessary compromises. Building on these insights, we introduce ERR (Ethical Reasoning Robustness), a defense framework that distinguishes between instrumental responses that enable harmful outcomes and explanatory responses that analyze ethical frameworks without endorsing harmful acts. ERR employs a Layer-Stratified Harm-Gated LoRA architecture, achieving robust defense against reasoning-based attacks while preserving model utility.
We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express retrieval decisions within token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of Retrieval as Generation, we propose GRIP (Generation-guided Retrieval with Information Planning), a unified framework in which the model regulates retrieval behavior through control-token emission. Central to GRIP is Self-Triggered Information Planning, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific token patterns. Experiments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters. Code and resources are provided in the supplementary materials.
Are We Using the Right Benchmark: An Evaluation Framework for Visual Token Compression Methods
PDF ↗Recent efforts to accelerate inference in Multimodal Large Language Models (MLLMs) have largely focused on visual token compression. The effectiveness of these methods is commonly evaluated by measuring the accuracy drop on existing MLLM benchmarks before and after compression. However, these benchmarks are originally designed to assess general perception and reasoning abilities, rather than the specific challenges posed by visual token compression, leading to a fundamental task mismatch. In this work, we uncover a counterintuitive yet consistent phenomenon: simple image downsampling outperforms many advanced visual token compression methods across multiple widely used benchmarks. Through a comprehensive empirical study spanning eight popular benchmarks and multiple state-of-the-art compression techniques, we show that (i) current benchmarks contain substantial noise (task-irrelevant samples) for evaluating visual token compression, and (ii) downsampling can act as an effective data filter that distinguishes between simple and difficult samples with respect to compression sensitivity. Motivated by these findings, we propose VTC-Bench, an evaluation framework that explicitly leverages downsampling as a discriminator to denoise existing benchmarks, enabling a fairer and more meaningful additional assessment of visual token compression methods.
Modeling fine-grained speaking styles remains challenging for language-speech representation pre-training, as existing speech-text models are typically trained with coarse captions or task-specific supervision, and scalable fine-grained style annotations are unavailable. We present FCaps, a large-scale dataset with fine-grained free-text style descriptions, encompassing 47k hours of speech and 19M fine-grained captions annotated via a novel end-to-end pipeline that directly grounds detailed captions in audio, thereby avoiding the error propagation caused by LLM-based rewriting in existing cascaded pipelines. Evaluations using LLM-as-a-judge demonstrate that our annotations surpass existing cascaded annotations in terms of correctness, coverage, and naturalness. Building on FCaps, we propose CLSP, a contrastive language-speech pre-trained model that integrates global and fine-grained supervision, enabling unified representations across multiple granularities. Extensive experiments demonstrate that CLSP learns fine-grained and multi-granular speech-text representations that perform reliably across global and fine-grained speech-text retrieval, zero-shot paralinguistic classification, and speech style similarity scoring, with strong alignment to human judgments. Code and dataset are publicly available at https://github.com/yfyeung/CLSP.