Generating with citations is crucial for trustworthy Large Language Models (LLMs), yet even advanced LLMs often produce mismatched or irrelevant citations. Existing methods over-optimize citation fidelity while overlooking relevance to the user query, which degrades answer quality and robustness in real-world settings with noisy or irrelevant retrieved content. Moreover, the prevailing single-pass paradigm struggles to deliver optimal answers in long-form generation that requiring multiple citations. To address these limitations, we propose FineRef, a framework based on Fine-grained error Reflection, which explicitly teaches the model to self-identify and correct two key citation errors—mismatch and irrelevance—on a per-citation basis. FineRef follows a two-stage training strategy. The first stage instills an “attempt–reflect–correct” behavioral pattern via supervised fine-tuning, using fine-grained and controllable reflection data constructed by specialized lightweight models. An online self-reflective bootstrapping strategy is designed to improve generalization by iteratively enriching training data with verified, self-improving examples. To further enhance the self-reflection and correction capability, the second stage applies process-level reinforcement learning with a multi-dimensional reward scheme that promotes reflection accuracy, answer quality, and correction gain. Experiments on the ALCE benchmark demonstrate that FineRef significantly improves both citation performance and answer accuracy. Our 7B model outperforms GPT-4 by up to 18% in Citation F1 and 4% in EM Recall, while also surpassing the state-of-the-art model across key evaluation metrics. FineRef also exhibits strong generalization and robustness in domain transfer settings and noisy retrieval scenarios.
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
Large language models now draft news, legal analyses, and software code with human-level fluency. At the same time, regulations such as the EU AI Act mandate that each synthetic passage carry an imperceptible, machine-verifiable mark for provenance. Conventional logit-based watermarks satisfy this requirement by selecting a pseudorandom green vocabulary at every decoding step and boosting its logits, yet the random split can exclude the highest-probability token and thus erode fluency. WaterMod mitigates this limitation through a probability-aware modular rule. The vocabulary is first sorted in descending model probability; the resulting ranks are then partitioned by the residue rank mod k, which distributes adjacent—and therefore semantically similar—tokens across different classes. A fixed bias of small magnitude is applied to one selected class. In the zero-bit setting (k=2), an entropy-adaptive gate selects either the even or the odd parity as the green list. Because the top two ranks fall into different parities, this choice embeds a detectable signal while guaranteeing that at least one high-probability token remains available for sampling. In the multi-bit regime (k>2), the current payload digit d selects the color class whose ranks satisfy rank mod k = d. Biasing the logits of that class embeds exactly one base-k digit—equivalently log2(k) bits—per decoding step, thereby enabling fine-grained provenance tracing. The same modular arithmetic therefore supports both binary attribution and rich payloads. Experimental results demonstrate that WaterMod consistently attains strong watermark detection performance while maintaining generation quality in both zero-bit and multi-bit settings. This robustness holds across a range of tasks, including natural language generation, mathematical reasoning, and code synthesis.
Large Language Models (LLMs) demonstrate strong reasoning capabilities but still struggle with hallucinations and limited transparency. Recently, KG-enhanced LLMs that integrate knowledge graphs (KGs) have been shown to improve reasoning performance, particularly for complex, knowledge-intensive tasks. However, these methods still face significant challenges, including inaccurate retrieval and reasoning failures, often exacerbated by long input contexts that obscure relevant information. Furthermore, many of these approaches rely on LLMs to directly retrieve evidence from KGs, and to self-assess the sufficiency of this evidence, which often results in premature or incorrect reasoning. To address the retrieval and reasoning failures, we propose ProgRAG, a multi-hop knowledge graph question answering (KGQA) framework that decomposes complex questions into sub-questions, and progressively extends partial reasoning paths by answering each sub-question. At each step, external retrievers gather candidate evidence, which is then refined through uncertainty-aware pruning by the LLM. Finally, the context for LLM reasoning is optimized by organizing and rearranging the partial reasoning paths obtained from the sub-question answers. Experiments on two well-known datasets, WebQSP and CWQ, demonstrate that ProgRAG outperforms existing baselines in multi-hop KGQA, offering improved reliability and reasoning quality.
Explain with Visual Keypoints Like a Real Mentor! A Benchmark for Multimodal Solution Explanation
PDF ↗With the rapid advancement of mathematical reasoning capabilities in Large Language Models (LLMs), AI systems are increasingly being adopted in educational settings to support students’ comprehension of problem-solving processes. However, a critical component remains underexplored in current LLM-generated explanations: multimodal explanation. In real-world instructional contexts, human tutors routinely employ visual aids, such as diagrams, markings, and highlights, to enhance conceptual clarity. To bridge this gap, we introduce the multimodal solution explanation task, designed to evaluate whether models can identify visual keypoints, such as auxiliary lines, points, angles, and generate explanations that incorporate these key elements essential for understanding. To evaluate model performance on this task, we propose ME2, a multimodal benchmark consisting of 1,000 math problems annotated with visual keypoints and corresponding explanatory text that references those elements. Our empirical results show that current models struggle to identify visual keypoints. In the task of generating keypoint-based explanations, open-source models also face notable difficulties. This highlights a significant gap in current LLMs’ ability to perform mathematical visual grounding, engage in visually grounded reasoning, and provide explanations in educational contexts. We expect that the multimodal solution explanation task and the ME2 dataset will catalyze further research on LLMs in education and promote their use as effective, explanation-oriented AI tutors.
Large language models (LLMs) may generate harmful outputs on malicious inputs. Existing safety methods, including prompt engineering and model editing, rely on hand-crafted templates or target-driven parameter modifications, limiting their generalizability in unseen harmful scenarios. Post-training aims to ensure LLM safety in general domains via supervised fine-tuning (SFT) or reinforcement learning (RL) on diverse malicious inputs. SFT needs annotated refusal samples while RL learns to refuse risk by exploring diverse harmful inputs. However, these methods tend to harshly refuse over any possible risks, sacrificing potentially useful information and degrading model utility. We argue that realistic malicious inputs often mix both harmful and helpful semantics (i.e., entities and relations), and LLMs should identify and remove only harmful relations while preserving useful ones. Thus, the original malicious user inputs can shift into safe queries, to which LLMs can respond safely and helpfully. In this paper, we propose WALKSAFE, a graph-based risk-aware training framework that enables LLMs to identify potential risks of key semantics (entities and relations) in user inputs via graph structure. By filtering harmful relations, LLMs can respond to safe input queries and then generate their corresponding safe and helpful responses. First, we model all entities and relations in the inputs with a graph structure. Second, we adopt a risk-aware random walk on the graph to quantify potential risk under multiple entities and relations. Then, we reconstruct safe queries by filtering harmful relations to promote the LLM to answer safely and helpfully rather than with direct refusals. Finally, we propose Bi-GRPO to post-train LLMs. As vanilla GRPO conducts only the intra-group comparison, Bi-GRPO performs both intra-group and inter-group comparisons between different response groups. The extra inter-group rewards encourage the model to distinguish harmful and safe semantics, and thus prefer safe and helpful responses. Experiments on three LLMs show that our models obtain SOTA results.
Pre-DPO: Improving Data Utilization in Direct Preference Optimization Using a Guiding Reference Model
PDF ↗Direct Preference Optimization (DPO) simplifies reinforcement learning from human feedback (RLHF) for large language models (LLMs) by directly training on offline preference data to align with human preferences. During DPO training, the reference model serves as a data weight adjuster. However, the common practice of initializing the policy and reference models identically in DPO can lead to inefficient data utilization and impose a performance ceiling. Meanwhile, the absence of a reference model in Simple Preference Optimization (SimPO) reduces training robustness and requires stricter conditions to prevent catastrophic forgetting. In this work, we propose Pre-DPO, a simple yet effective DPO-based training paradigm that improves preference optimization by introducing a guiding reference model. This reference model provides foresight into the desired policy state achievable through the training preference data, serving as a guiding mechanism that adaptively assigns higher weights to samples more suitable for the model and lower weights to those less suitable. Extensive experiments on the AlpacaEval 2 and Arena-Hard v0.1 benchmarks demonstrate that Pre-DPO consistently improves the performance of both DPO and SimPO, without relying on external models or additional data.
Social biases embedded in Large Language Models (LLMs) raise critical concerns, resulting in representational harms -- unfair or distorted portrayals of demographic groups -- that may be expressed in subtle ways through generated language. Existing evaluation methods often depend on predefined identity-concept associations, limiting their ability to surface new or unexpected forms of bias. In this work, we present the Bias Association Discovery Framework (BADF), a systematic approach for extracting both known and previously unrecognized associations between demographic identities and descriptive concepts from open-ended LLM outputs. Through comprehensive experiments spanning multiple models and diverse real-world contexts, BADF enables robust mapping and analysis of the varied concepts that characterize demographic identities. Our findings advance the understanding of biases in open-ended generation and provide a scalable tool for identifying and analyzing bias associations in LLMs.
We introduce DSCodeBench, a new benchmark designed to evaluate large language models (LLMs) on complicated and realistic data science code generation tasks. DSCodeBench consists of 1,000 carefully constructed problems sourced from realistic problems from GitHub across ten widely used Python data science libraries. DSCodeBench offers a more challenging and representative testbed, more complex code solutions, more comprehensive data science libraries, clearer and better structured problem descriptions, and stronger test suites. To construct the DSCodeBench, we develop a robust pipeline that combines task scope selection, code construction, test case generation, and problem description synthesis. The process is paired with rigorous manual editing to ensure alignment and enhance the reliability of the evaluation. Experimental result shows that DSCodeBench exhibits robust scaling behavior, where larger models systematically outperform smaller ones, validating its ability to distinguish model capabilities. The best LLM we test, GPT-4o, has a pass@1 of 0.392, indicating that LLMs still have a large room to improve for realistic data science code generation tasks. We believe DSCodeBench will serve as a rigorous and trustworthy foundation for advancing LLM-based data science programming.
Reinforcement Learning (RL) has been shown to improve the capabilities of large language models (LLMs). However, applying RL to open-domain tasks faces two key challenges: (1) the inherent subjectivity of these tasks prevents the verifiable rewards as required by Reinforcement Learning with Verifiable Rewards (RLVR); (2) Reinforcement Learning from Human Feedback (RLHF) relies on external reward mechanisms. To overcome these limitations, we propose Self-Examining Reinforcement Learning (SERL), a novel self-improving framework where the LLM serves as both Actor and Judge. SERL introduces two synergistic reward mechanisms without any external signals. On the one hand, to improve the Actor's capability, we derive rewards from Copeland-style pairwise comparison judgments across a group of generated responses. On the other hand, a self-consistency reward that encourages coherent judgments is proposed to improve the Judge's reliability. This refinement strengthens the Judge, consequently generating a more robust training signal for the Actor. Experiments show that our method outperforms existing self-improvement training methods. SERL improves the LC win rate of Qwen3-8B on AlpacaEval 2.0 from 52.37% to 59.90%. To the best of our knowledge, our method achieves state-of-the-art performance among self-improving approaches. Furthermore, it achieves a performance comparable to significantly larger models like Qwen3-32B, demonstrating superior effectiveness and robustness on open-domain tasks.
Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, DoRA, and HiRA, enable lightweight adaptation of large pre-trained models via low-rank updates. However, existing PEFT approaches apply static, input-agnostic updates to all tokens, disregarding the varying importance and difficulty of different inputs. This uniform treatment can lead to overfitting on trivial content or under-adaptation on more informative regions, especially in autoregressive settings with distinct prefill and decoding dynamics. In this paper, we propose GateRA, a unified framework that introduces token-aware modulation to dynamically adjust the strength of PEFT updates. By incorporating adaptive gating into standard PEFT branches, GateRA enables selective, token-level adaptation—preserving pre-trained knowledge for well-modeled inputs while focusing capacity on challenging cases. Empirical visualizations reveal phase-sensitive behaviors, where GateRA automatically suppresses updates for redundant prefill tokens while emphasizing adaptation during decoding. To promote confident and efficient modulation, we further introduce an entropy-based regularization that encourages near-binary gating decisions. This regularization prevents diffuse update patterns and leads to interpretable, sparse adaptation without hard thresholding. Finally, we present a theoretical analysis showing that GateRA induces a soft gradient-masking effect over the PEFT path, enabling continuous and differentiable control over adaptation. Experiments on multiple commonsense reasoning benchmarks demonstrate that GateRA consistently outperforms or matches prior PEFT methods.
Research in Machine Learning (ML) and AI evolves rapidly. Information Extraction (IE) from scientific publications enables to identify information about research concepts and resources on a large scale and therefore is a pathway to improve understanding and reproducibility of ML-related research. To training and testing of IE models focused on fine-grained information in ML-related research, e.g. method training and data usage, we introduce GSAP-ERE. It is a manually curated fine-grained dataset of mentions of 63K ML-related entities and 35K relations distributed across 10 entity types and 18 semantically categorized relation types annoated in the full text of 100 ML publications. We show that our dataset enables fine-tuned models to automatically extract ML-related information that facilitate knowledge graph (KG) construction from scholarly papers or monitoring of computational reproducibility of AI research at scale. Additionally, we use our dataset as a test suite to explore prompting strategies for IE using Large Language Models (LLM). We observe that the performance of state-of-the-art LLM prompting methods is largely outperformed by our best fine-tuned baseline model (NER: 80.6%, RE: 54.0% for the fine-tuned model vs. NER: 44.4%, RE: 10.1% for the LLM). This disparity of performance between supervised models and unsupervised usage of LLMs suggests datasets like GSAP-ERE are needed to advance research in the domain of scholarly information extraction.
Latent Self-Consistency for Reliable Majority-Set Selection in Short- and Long-Answer Reasoning
PDF ↗Probabilistic decoding in Large Language Models (LLMs) often yields inconsistent outputs, particularly on complex or long-form questions. Self-Consistency (SC) mitigates this for short-form QA by majority voting over exact strings, whereas Universal Self-Consistency (USC) and Weighted Unigram Consistency Score (WUCS) extend to long-form responses but lose accuracy on short-form benchmarks. We introduce Latent Self-Consistency (LSC), which selects the most semantically consistent response using learnable token embeddings. LSC's lightweight forward processing of summary tokens only introduces negligible runtime overhead (at most 0.9%) on top of standard decoding of the base LLM, and requires no changes to the model architecture. Across 6 short-form and 5 long-form reasoning benchmarks (e.g., MATH, MMLU, TruthfulQA), LSC surpasses SC, USC, and WUCS on both short-form and long-form on average performance, while adding negligible computational overhead on vanilla inference. These results position LSC as a reliable consistency-selection method that works effectively across various answer formats. Additionally, LSC provides well-calibrated confidence estimates, maintaining low expected calibration error across both answer formats.
Large language model (LLM) services now answer billions of queries per day, and industry reports show that inference, not training, accounts for more than 90% of total power consumption. However, existing benchmarks focus on either training/fine-tuning or performance of inference and provide little support for power consumption measurement and analysis of inference. We introduce TokenPowerBench, the first lightweight and extensible benchmark designed for LLM-inference power consumption studies. The benchmark combines a declarative configuration interface covering model choice, prompt set, and inference engine, a measurement layer that captures GPU-, node-, and system-level power without specialized power meters, and a phase-aligned metrics pipeline that attributes energy to the prefill and decode stages of every request. These elements make it straightforward to explore the power consumed by an LLM inference run; furthermore, by varying batch size, context length, parallelism strategy and quantization, users can quickly assess how each setting affects joules per token and other energy-efficiency metrics. We evaluate TokenPowerBench on four of the most widely used model series (Llama, Falcon, Qwen, and Mistral). Our experiments cover from 1 billion parameters up to the frontier-scale Llama3-405B model. Furthermore, we release TokenPowerBench as open source to help users to measure power consumption, forecast operating expenses, and meet sustainability targets when deploying LLM services.
Large Language Models (LLMs) often suffer from hallucinations and outdated or incomplete knowledge. Retrieval-Augmented Generation (RAG) is proposed to address these issues by integrating external knowledge like that in knowledge graphs (KGs) into LLMs. However, leveraging private KGs in RAG systems poses significant privacy risks due to the black-box nature of LLMs and potential insecure data transmission. In this paper, we investigate the privacy-protected RAG scenario for the first time, where entities in KGs are anonymous for LLMs, thus preventing them from accessing entity semantics. Due to the loss of semantics of entities, previous RAG systems cannot retrieve question-relevant knowledge from KGs by matching questions with the meaningless identifiers of anonymous entities. To realize an effective RAG system in this scenario, two key challenges must be addressed: (1) How can anonymous entities be converted into retrievable information? (2) How to retrieve question-relevant anonymous entities? To address these challenges, we propose a novel Abstraction Reasoning on Graph (ARoG) framework including relation-centric abstraction and structure-oriented abstraction strategies. For challenge (1), the first strategy abstracts entities into high-level concepts by dynamically capturing the semantics of their adjacent relations. Hence, it supplements meaningful semantics which can further support the retrieval process. For challenge (2), the second strategy transforms unstructured natural language questions into structured abstract concept paths. These paths can be more effectively aligned with the abstracted concepts in KGs, thereby improving retrieval performance. In addition to guiding LLMs to effectively retrieve knowledge from KGs, these abstraction strategies also strictly protect privacy from being exposed to LLMs. Experiments on three datasets demonstrate that ARoG achieves strong performance and privacy-robustness, establishing a new practical direction for privacy-protected RAG systems.
GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging
PDF ↗Beyond scratch coding, exploiting large-scale code repositories (e.g., GitHub) for practical tasks is vital in real-world software development, yet current benchmarks rarely evaluate code agents in such authentic, workflow-driven scenarios. To bridge this gap, we introduce GitTaskBench, a benchmark designed to systematically assess this capability via 54 realistic tasks across 7 modalities and 7 domains. Each task pairs a relevant repository with an automated, human-curated evaluation harness specifying practical success criteria. Beyond measuring execution and task success, we also propose the alpha-value metric to quantify the economic benefit of agent performance, which integrates task success rates, token cost, and average developer salaries. Experiments across three state-of-the-art agent frameworks with multiple advanced LLMs show that leveraging code repositories for complex task solving remains challenging: even the best-performing system, OpenHands+Claude 3.7, solves only 48.15% of tasks. Error analysis attributes over half of failures to seemingly mundane yet critical steps like environment setup and dependency resolution, highlighting the need for more robust workflow management and increased timeout preparedness. By releasing GitTaskBench, we aim to drive progress and attention toward repository-aware code reasoning, execution, and deployment---moving agents closer to solving complex, end-to-end real-world tasks.
Large Language Models (LLMs) often produce fluent but factually incorrect responses, a phenomenon known as hallucination. Abstention, where the model chooses not to answer and instead outputs phrases such as "I don't know", is a common safeguard. However, existing abstention methods typically rely on post-generation signals, such as generation variations or feedback, which limits their ability to prevent unreliable responses in advance. In this paper, we introduce Aspect-Based Causal Abstention (ABCA), a new framework that enables early abstention by analysing the internal diversity of LLM knowledge through causal inference. This diversity reflects the multifaceted nature of parametric knowledge acquired from various sources, representing diverse aspects such as disciplines, legal contexts, or temporal frames. ABCA estimates causal effects conditioned on these aspects to assess the reliability of knowledge relevant to a given query. Based on these estimates, we enable two types of abstention: Type-1, where aspect effects are inconsistent (knowledge conflict), and Type-2, where aspect effects consistently support abstention (knowledge insufficiency). Experiments on standard benchmarks demonstrate that ABCA improves abstention reliability, achieves state-of-the-art performance, and enhances the interpretability of abstention decisions.
Probabilities Are All You Need: A Probability-Only Approach to Uncertainty Estimation in Large Language Models
PDF ↗Large Language Models (LLMs) exhibit strong performance across various natural language processing (NLP) tasks but remain vulnerable to hallucinations, generating factually incorrect or misleading outputs. Uncertainty estimation, often using predictive entropy estimation, is key to addressing this issue. However, existing methods often require multiple samples or extra computation to assess semantic entropy. This paper proposes an efficient, training-free uncertainty estimation method that approximates predictive entropy using the responses' top-K probabilities. Moreover, we employ an adaptive mechanism to determine K to enhance flexibility and filter out low-confidence probabilities. Experimental results on three free-form question-answering datasets across several LLMs demonstrate that our method outperforms expensive state-of-the-art baselines, contributing to the broader goal of enhancing LLM trustworthiness.
Language models (LMs) are often adapted through supervised fine-tuning (SFT) to specialize their capabilities for downstream tasks. However, in typical scenarios where the fine-tuning data is limited, e.g., compared to pre-training, SFT can lead LMs to overfit, causing them to rely on spurious patterns within the target task or to compromise other broadly useful capabilities as a side effect of narrow specialization. In this paper, we propose Learning-from-the-Undesirable (LfU), a simple yet effective regularization scheme for SFT to mitigate overfitting issues when fine-tuning LMs with limited data. Specifically, we aim to regularize the fine-tuning process to favor solutions that are resilient to “undesirable” model updates, e.g., gradient ascent steps that steer the model toward undesirable behaviors. To this end, we propose a novel form of consistency regularization that directly aligns internal representations of the model with those after an undesirable update. By leveraging representation-level data augmentation through undesirable updates, LfU effectively promotes generalization under limited data. Our experiments on diverse LM downstream tasks show that LfU serves as an effective prior that enhances adaptability while preserving pretrained knowledge. For example, our LM from LfU achieves a 16.8% average improvement on math tasks compared to vanilla SFT on the same dataset, where the latter even leads to degraded performance on those tasks. Furthermore, LfU exhibits improved robustness to prompt variations, e.g., yielding a 92.1% lower standard deviation in output performances compared to SFT, highlighting its versatile effects.
A Multi-Agent LLM Framework for Multi-Domain Low-Resource In-Context NER via Knowledge Retrieval, Disambiguation and Reflective Analysis
PDF ↗In-context learning (ICL) with large language models (LLMs) has emerged as a promising paradigm for named entity recognition (NER) in low-resource scenarios. However, existing ICL-based NER methods suffer from three key limitations: (1) reliance on dynamic retrieval of annotated examples, which is problematic when annotated data is scarce; (2) limited generalization to unseen domains due to the LLM's insufficient internal domain knowledge; and (3) failure to incorporate external knowledge or resolve entity ambiguities. To address these challenges, we propose KDR-Agent, a novel multi-agent framework for multi-domain low-resource in-context NER that integrates Knowledge retrieval, Disambiguation, and Reflective analysis. KDR-Agent leverages natural-language type definitions and a static set of entity-level contrastive demonstrations to reduce dependency on large annotated corpora. A central planner coordinates specialized agents to (i) retrieve factual knowledge from Wikipedia for domain-specific mentions, (ii) resolve ambiguous entities via contextualized reasoning, and (iii) reflect on and correct model predictions through structured self-assessment. Experiments across ten datasets from five domains demonstrate that KDR-Agent significantly outperforms existing zero-shot and few-shot ICL baselines across multiple LLM backbones.
Hearing More with Less: Multi-Modal Retrieval-and-Selection Augmented Conversational LLM-Based ASR
PDF ↗Automatic Speech Recognition (ASR) aims to convert human speech content into corresponding text. In conversational scenarios, effectively utilizing context can enhance its accuracy. Large Language Models' (LLMs) exceptional long-context understanding and reasoning abilities enable LLM-based ASR (LLM-ASR) to leverage historical context for recognizing conversational speech, which has a high degree of contextual relevance. However, existing conversational LLM-ASR methods use a fixed number of preceding utterances or the entire conversation history as context, resulting in significant ASR confusion and computational costs due to massive irrelevant and redundant information. This paper proposes a multi-modal retrieval-and-selection method named MARS that augments conversational LLM-ASR by enabling it to retrieve and select the most relevant acoustic and textual historical context for the current utterance. Specifically, multi-modal retrieval obtains a set of candidate historical contexts, each exhibiting high acoustic or textual similarity to the current utterance. Multi-modal selection calculates the acoustic and textual similarities for each retrieved candidate historical context and, by employing our proposed near-ideal ranking method to consider both similarities, selects the best historical context. Evaluations on the Interspeech 2025 Multilingual Conversational Speech Language Model Challenge dataset show that the LLM-ASR, when trained on only 1.5K hours of data and equipped with the MARS, outperforms the state-of-the-art top-ranking system trained on 179K hours of data.