ACL 2025 论文列表

As the capabilities of large language models (LLMs) continue to expand, aligning these models with human values remains a significant challenge. Recent studies show that reasoning abilities contribute significantly to model safety, while integrating Mixture-of-Experts (MoE) architectures can further enhance alignment.In this work, we address a fundamental question:How to effectively incorporate reasoning abilitiesand MoE architectures into self-alignment processin LLMs?We propose Mixture of insighTful Experts (MoTE), a novel framework that synergistically combines reasoning chains and expert mixtures to improve self-alignments.From a data perspective, MoTE employs a structured reasoning chain comprising four key stages: Question Analysis, Answer Guidance, Safe Answer, and Safety Checking. This approach enhances safety through multi-step reasoning and proves effective even for smaller and less powerful LLMs (e.g., 7B models). From an architectural perspective, MoTE adopts a multi-LoRA framework with step-level routing, where each expert is dedicated to a specific reasoning step. This design eliminates the need for balance losses, ensures stable training, and supports adaptive inference lengths. Experimental results demonstrate that MoTE significantly improves model safety, jailbreak resistance, and over-refusal capabilities, achieving performance comparable to OpenAI’s state-of-the-art o1 model.

Text-rich Graph Knowledge Bases (TG-KBs) have become increasingly crucial for answering queries by providing textual and structural knowledge. However, current retrieval methods often retrieve these two types of knowledge in isolation without considering their mutual reinforcement and existing hybrid methods even bypass structural retrieval entirely. To fill this gap, we propose a Mixture of Structural-and-Textual Retrieval (MoR) to retrieve these two types of knowledge via a Planning-Reasoning-Organizing framework. In the Planning stage, MoR generates textual planning graphs delineating the logic for answering queries. Following planning graphs, in the Reasoning stage, MoR interweaves structural traversal and textual matching to obtain candidates from TG-KBs. In the Organizing stage, MoR further reranks fetched candidates based on their structural trajectory. Extensive experiments demonstrate the superiority of MoR in harmonizing structural and textual retrieval with inspiring insights, including imbalanced retrieving performance across different query logics and the benefits of integrating structural trajectories for candidate reranking.

In the era of large language models (LLMs), the Chinese Spelling Check (CSC) task has seen various LLM methods developed, yet their performance remains unsatisfactory. In contrast, fine-tuned BERT-based models, relying on high-quality in-domain data, show excellent performance but suffer from edit pattern overfitting. This paper proposes a novel dynamic mixture approach that effectively combines the probability distributions of small models and LLMs during the beam search decoding phase, achieving a balanced enhancement of precise corrections from small models and the fluency of LLMs. This approach also eliminates the need for fine-tuning LLMs, saving significant time and resources, and facilitating domain adaptation. Comprehensive experiments demonstrate that our mixture approach significantly boosts error correction capabilities, achieving state-of-the-art results across multiple datasets. Our code is available at https://github.com/zhqiao-nlp/MSLLM.

Automated Essay Scoring (AES) plays a crucial role in language assessment. In particular, cross-prompt essay trait scoring provides learners with valuable feedback to improve their writing skills. However, due to the scarcity of prompts, most existing methods overlook critical information, such as content from prompts or essays, resulting in incomplete assessment perspectives. In this paper, we propose a robust AES framework, the Mixture of Ordered Scoring Experts (MOOSE), which integrates information from both prompts and essays. MOOSE employs three specialized experts to evaluate (1) the overall quality of an essay, (2) the relative quality across multiple essays, and (3) the relevance between an essay and its prompt. MOOSE introduces the ordered aggregation of assessment results from these experts along with effective feature learning techniques. Experimental results demonstrate that MOOSE achieves exceptionally stable and state-of-the-art performance in both cross-prompt scoring and multi-trait scoring on the ASAP++ dataset. The source code is released at https://github.com/antslabtw/MOOSE-AES.

Large Vision-Language Models (LVLMs) have exhibited impressive capabilities across various visual tasks, yet they remain hindered by the persistent challenge of hallucinations. To address this critical issue, we propose Mixture of Decoding (MoD), a novel approach for hallucination mitigation that dynamically adapts decoding strategies by evaluating the correctness of the model’s attention on image tokens. Specifically, MoD measures the consistency between outputs generated from the original image tokens and those derived from the model’s attended image tokens, to distinguish the correctness aforementioned. If the outputs are consistent, indicating correct attention, MoD employs a complementary strategy to amplify critical information. Conversely, if the outputs are inconsistent, suggesting erroneous attention, MoD utilizes a contrastive strategy to suppress misleading information. Extensive experiments demonstrate that MoD significantly outperforms existing decoding methods across multiple mainstream benchmarks, effectively mitigating hallucinations in LVLMs. Code is available at https://github.com/xlchen0205/MoD.

Recent advancements in Large Language Models (LLMs) have demonstrated enhanced reasoning capabilities, evolving from Chain-of-Thought (CoT) prompting to advanced, product-oriented solutions like OpenAI o1. During our re-implementation of this model, we noticed that in multimodal tasks requiring visual input (e.g., geometry problems), Multimodal LLMs (MLLMs) struggle to maintain focus on the visual information, in other words, MLLMs suffer from a gradual decline in attention to visual information as reasoning progresses, causing text-over-relied outputs. To investigate this, we ablate image inputs during long-chain reasoning. Concretely, we truncate the reasoning process midway, then re-complete the reasoning process with the input image removed. We observe only a ~2 accuracy drop on MathVista’s test-hard subset, revealing the model’s textual outputs dominate the following reasoning process. Motivated by this, we propose Take-along Visual Conditioning (TVC), a strategy that shifts image input to critical reasoning stages and compresses redundant visual tokens via dynamic pruning. This methodology helps the model retain attention to the visual components throughout the reasoning. Our approach achieves state-of-the-art performance on average across five mathematical reasoning benchmarks (+3.4% vs previous sota), demonstrating the effectiveness of TVC in enhancing multimodal reasoning systems. The project page is available at https://sun-hailong.github.io/projects/TVC.

Empirical risk minimization (ERM) incentivizes models to exploit shortcuts, i.e., spurious correlations between input attributes and labels that are prevalent in the majority of the training data but unrelated to the task at hand. This reliance hinders generalization on minority examples, where such correlations do not hold. Existing shortcut mitigation approaches are model-specific, difficult to tune, computationally expensive, and fail to improve learned representations. To address these issues, we propose InterpoLated Learning (InterpoLL) which interpolates the representations of majority examples to include features from intra-class minority examples with shortcut-mitigating patterns. This weakens shortcut influence, enabling models to acquire features predictive across both minority and majority examples. Experimental results on multiple natural language understanding tasks demonstrate that InterpoLL improves minority generalization over both ERM and state-of-the-art mitigation methods, without compromising accuracy on majority examples. Notably, these gains persist across encoder, encoder-decoder, and decoder-only architectures, demonstrating the method’s broad applicability.

Large language models (LLMs) often exhibit systematic preferences for certain answer choices when responding to multiple-choice questions—a behavior known as selection bias. This bias reduces the accuracy and reliability of LLM outputs, limiting their usefulness in decision-critical applications. While prior work has focused on adjusting model inputs or outputs to mitigate this issue, our work takes a fundamentally different approach by identifying and removing the internal sources of bias. We introduce two methods: Bias Node Pruning (BNP), which prunes parameters that contribute to selection bias, and Auxiliary Option Injection (AOI), which introduces an additional answer choice to reduce bias in both white-box and black-box settings. To address the shortcomings of existing evaluation metrics, we propose Choice Kullback-Leibler Divergence (CKLD), a new metric that captures distributional imbalances in model predictions. Experiments on three LLMs across multiple datasets demonstrate that our methods consistently improve answer accuracy while reducing selection bias, providing a robust solution for both open- and closed-source models.

While Large Language Models (LLMs) support long contexts, they struggle with performance degradation within the context window. Current solutions incur prohibitive training costs, leaving statistical behaviors and cost-effective approaches underexplored. From the decoding perspective, we identify the Posterior Salience Attenuation (PSA) phenomenon, where the salience ratio correlates with long-text performance degradation. Notably, despite the attenuation, gold tokens still occupy high-ranking positions in the decoding space. Motivated by it, we propose the training-free Positional Contrastive Decoding (PCD) that contrasts the logits derived from long-aware attention with those from designed local-aware attention, enabling the model to focus on the gains introduced by large-scale short-to-long training. Through the analysis of long-term decay simulation, we demonstrate that PCD effectively alleviates attention score degradation. Experimental results show that PCD achieves state-of-the-art performance on long-context benchmarks.

High-quality paraphrases are easy to produce using instruction-tuned language models or specialized paraphrasing models. Although this capability has a variety of benign applications, paraphrasing attacks—paraphrases applied to machine-generated texts—are known to significantly degrade the performance of machine-text detectors. This motivates us to consider the novel problem of paraphrase inversion, where, given paraphrased text, the objective is to recover an approximation of the original text. The closer the approximation is to the original text, the better machine-text detectors will perform. We propose an approach which frames the problem as translation from paraphrased text back to the original text, which requires examples of texts and corresponding paraphrases to train the inversion model. Fortunately, such training data can easily be generated, given a corpus of original texts and one or more paraphrasing models. We find that language models such as GPT-4 and Llama-3 exhibit biases when paraphrasing which an inversion model can learn with a modest amount of data. Perhaps surprisingly, we also find that such models generalize well, including to paraphrase models unseen at training time. Finally, we show that when combined with a paraphrased-text detector, our inversion models provide an effective defense against paraphrasing attacks, and overall our approach yields an average improvement of +22% AUROC across seven machine-text detectors and three different domains.

To address the phenomenon of similar classes, existing methods in few-shot continual relation extraction (FCRE) face two main challenges: non-representative prototypes and representation bias, especially when the number of available samples is limited. In our work, we propose Minion to address these challenges. Firstly, we leverage the General Orthogonal Frame (GOF) structure, based on the concept of Neural Collapse, to create robust class prototypes with clear separation, even between analogous classes. Secondly, we utilize label description representations as global class representatives within the fast-slow contrastive learning paradigm. These representations consistently encapsulate the essential attributes of each relation, acting as global information that helps mitigate overfitting and reduces representation bias caused by the limited local few-shot examples within a class. Extensive experiments on well-known FCRE benchmarks show that our method outperforms state-of-the-art approaches, demonstrating its effectiveness for advancing RE system.

Efficiently updating multilingual knowledge in large language models (LLMs) without disrupting coherent factual representations across languages remains a significant challenge. While deploying separate editing systems for each language might seem viable, this approach incurs substantial costs due to the need to manage multiple models. A more efficient solution involves integrating knowledge updates across all languages into a unified model. However, sequential edits across languages often lead to destructive parameter interference, significantly degrading multilingual generalization and the accuracy of injected knowledge. To address this issue, we propose LangEdit, a novel null-space constrained framework designed to precisely isolate language-specific knowledge updates. The core innovation of LangEdit lies in its ability to project parameter updates for each language onto the orthogonal complement of other languages’ subspaces. This approach mathematically guarantees update independence while preserving multilingual generalization capabilities. We conduct a comprehensive evaluation across three model architectures, six languages, and four downstream tasks, demonstrating that LangEdit effectively mitigates parameter interference and outperforms existing state-of-the-art editing methods. Our results highlight its potential for enabling efficient and accurate multilingual knowledge updates in LLMs.

In this paper, we identify a critical problem, “lost-in-retrieval”, in retrieval-augmented multi-hop question answering (QA): the key entities are missed in LLMs’ sub-question decomposition. “Lost-in-retrieval” significantly degrades the retrieval performance, which disrupts the reasoning chain and leads to the incorrect answers. To resolve this problem, we propose a progressive retrieval and rewriting method, namely ChainRAG, which sequentially handles each sub-question by completing missing key entities and retrieving relevant sentences from a sentence graph for answer generation. Each step in our retrieval and rewriting process builds upon the previous one, creating a seamless chain that leads to accurate retrieval and answers. Finally, all retrieved sentences and sub-question answers are integrated to generate a comprehensive answer to the original question. We evaluate ChainRAG on three multi-hop QA datasets—MuSiQue, 2Wiki, and HotpotQA—using three large language models: GPT4o-mini, Qwen2.5-72B, and GLM-4-Plus. Empirical results demonstrate that ChainRAG consistently outperforms baselines in both effectiveness and efficiency.

Multimodal Large Language Models (MLLMs) are known to hallucinate, which limits their practical applications. Recent works have attempted to apply Direct Preference Optimization (DPO) to enhance the performance of MLLMs, but have shown inconsistent improvements in mitigating hallucinations. To address this issue more effectively, we introduce Hallucination-targeted Direct Preference Optimization (HDPO) to reduce hallucinations in MLLMs. Unlike previous approaches, our method tackles hallucinations from their diverse forms and causes. Specifically, we develop three types of preference pair data targeting the following causes of MLLM hallucinations: (1) insufficient visual capabilities, (2) long context generation, and (3) multimodal conflicts. Experimental results demonstrate that our method achieves superior performance across multiple hallucination evaluation datasets, surpassing most state-of-the-art (SOTA) methods and highlighting the potential of our approach. Ablation studies and in-depth analyses further confirm the effectiveness of our method and suggest the potential for further improvements through scaling up.

Recent advances in large language models (LLMs) have demonstrated the effectiveness of chain-of-thought (CoT) prompting. Few-Shot-CoT relies on task-specific, manually labeled demonstrations, limiting its generalization to unseen tasks. While Zero-Shot-CoT eliminates this reliance, it often underperforms. To address this, existing methods aim to automatically generate demonstrations in zero-shot settings. However, these generated demonstrations face challenges due to demonstration bias: 1) selected demonstrations may contain errors, and 2) they may not be suitable or representative enough for all questions. To mitigate these biases, we propose Global Coevolutionary Reasoning (GCR). The method first applies Zero-Shot-CoT to answer all questions, then clusters the results. For each cluster, a random sample is selected, and these selected samples serve as demonstrations for each other. The model then iteratively re-answers the questions and updates their rationales based on these demonstrations, enabling coevolutionary reasoning to progressively improve the quality of the answers. This process of random sampling and coevolutionary reasoning is repeated until all questions have been re-answered. Experimental results on ten datasets using GPT-3.5-turbo and GPT-4o-mini show that GCR outperforms baseline methods without any performance degradation caused by demonstration bias. Additionally, GCR is orthogonal to existing methods and can be seamlessly integrated with them. The code is available at: https://github.com/GouChuan/GCR.

Deep transformer models have been used to detect linguistic anomalies in patient transcripts for early Alzheimer’s disease (AD) screening. While pre-trained neural language models (LMs) fine-tuned on AD transcripts perform well, little research has explored the effects of the gender of the speakers represented by these transcripts. This work addresses gender confounding in dementia detection and proposes two methods: the Extended Confounding Filter and the Dual Filter, which isolate and ablate weights associated with gender. We evaluate these methods on dementia datasets with first-person narratives from patients with cognitive impairment and healthy controls. Our results show transformer models tend to overfit to training data distributions. Disrupting gender-related weights results in a deconfounded dementia classifier, with the trade-off of slightly reduced dementia detection performance.

In retrieval augmented generation (RAG) systems, each individual component—the LLM, embedder, and corpus—could introduce biases in the form of skews towards certain genders or political leanings. In this work, we study the conflict between biases of each component and their relationship to the overall bias of the RAG system, which we call bias conflict. Examining both gender and political biases as case studies, we show that bias conflict can be characterized through a linear relationship among components despite its complexity. Through fine-tuning, we demonstrate how to control the bias of the embedder while maintaining utility and reveal the importance of reverse-biasing the embedder to mitigate bias in the overall system, Additionally, we find that LLMs and tasks exhibit varying sensitivities to bias, a crucial factor to consider for debiasing. Our results underscore that a fair RAG system can be better achieved by carefully controlling the bias of the embedder rather than increasing its fairness.

Long-context language models (LCLMs) can process long context, but still exhibit position bias, also known as “lost in the middle”, which indicates placing key information in the middle of the context will significantly affect performance. To mitigating this, we first explore the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias. Then we identify that, in addition to position embeddings, positional information in hidden states also contributes to position bias, and it manifests itself in specific channels of hidden states, called positional hidden states. Based on these, we propose a method to mitigate position bias by scaling positional hidden states. Experiments on NaturalQuestions Multi-document QA, KV retrieval and LongBench, using various models including RoPE models, context window-extended models, and Alibi models, demonstrate the effectiveness and generalizability of our approach. Our method can improve performance by up to 15.2% in “lost in the middle” benchmark by modifying just one channel of hidden states. Our code is available at https://aka.ms/PositionalHidden.

Many applications of Large Language Models (LLMs) require them to either simulate people or offer personalized functionality, making the demographic representativeness of LLMs crucial for equitable utility. At the same time, we know little about the extent to which these models actually reflect the demographic attributes and behaviors of certain groups or populations, with conflicting findings in empirical research. To shed light on this debate, we review 211 papers on the demographic representativeness of LLMs. We find that while 29% of the studies report positive conclusions on the representativeness of LLMs, 30% of these do not evaluate LLMs across multiple demographic categories or within demographic subcategories. Another 35% and 47% of the papers concluding positively fail to specify these subcategories altogether for gender and race, respectively. Of the articles that do report subcategories, fewer than half include marginalized groups in their study. Finally, more than a third of the papers do not define the target population to whom their findings apply; of those that do define it either implicitly or explicitly, a large majority study only the U.S. Taken together, our findings suggest an inflated perception of LLM representativeness in the broader community. We recommend more precise evaluation methods and comprehensive documentation of demographic attributes to ensure the responsible use of LLMs for social applications.