ACL 2025 论文列表

Long-document QA presents challenges with large-scale text and long-distance dependencies. Recent advances in Large Language Models (LLMs) enable entire documents to be processed in a single pass. However, their computational cost is significantly high. Retrieval-Augmented Generation (RAG) methods split text into smaller chunks, but they often yield inferior results and may lose global context. Recent approaches that integrate LLMs into RAG via iterative summarization either underutilize LLM capabilities or still incur high computational costs. In this paper, we combine the high accuracy of LLMs with the efficiency of RAG and propose LLM-Guided Dynamic Progress Control with Attention-Based Hierarchical Weighted Graph (PECAN). Our method introduces two key improvements: (1) LLM-Guided Dynamic Progress Control: We leverage LLMs to dynamically control the retrieval process, adjusting the amount of retrieved information based on different queries to achieve a better balance of effectiveness and efficiency. (2) Attention-Guided Retrieval: We propose a novel retrieval method that constructs a hierarchical graph where edges are derived by LLM attention weights. Experimental results demonstrate that PECAN achieves LLM-level performance while maintaining computational complexity comparable to that of RAG methods on two single-document and two multi-document QA datasets.

Disinformation detection is a key aspect of media literacy. Psychological studies have shown that knowledge of persuasive fallacies helps individuals detect disinformation. Inspired by these findings, we experimented with large language models (LLMs) to test whether infusing persuasion knowledge enhances disinformation detection. As a result, we introduce the Persuasion-Augmented Chain of Thought (PCoT), a novel approach that leverages persuasion to improve disinformation detection in zero-shot classification. We extensively evaluate PCoT on online news and social media posts. Moreover, we publish two novel, up-to-date disinformation datasets: EUDisinfo and MultiDis. These datasets enable the evaluation of PCoT on content entirely unseen by the LLMs used in our experiments, as the content was published after the models’ knowledge cutoffs. We show that, on average, PCoT outperforms competitive methods by 15% across five LLMs and five datasets. These findings highlight the value of persuasion in strengthening zero-shot disinformation detection.

Aspect Sentiment Triplet Extraction (ASTE) is a subtask of Aspect-Based Sentiment Analysis (ABSA) that aims to extract aspect terms, corresponding opinion terms, and their associated sentiment polarities from text. Current end-to-end approaches, whether employing Large Language Models (LLMs) or complex neural network structures, struggle to effectively model the intricate latent relationships between aspects and opinions. Therefore, in this work, we propose Polarity-Aware Sentiment Triplet Extraction with LLM-as-a-judge (PASTEL), a novel pipeline that decomposes the ASTE task into structured subtasks. We employ finetuned LLMs to separately extract the aspect and opinion terms, incorporating a polarity-aware mechanism to enhance opinion extraction. After generating a candidate set through the Cartesian product of the extracted aspect and opinion-sentiment sets, we leverage an LLM-as-a-Judge to validate and prune these candidates. Experimental evaluations demonstrate that PASTEL outperforms existing baselines. Our findings highlight the necessity of modular decomposition in complex sentiment analysis tasks to fully exploit the capabilities of current LLMs.

Large language models (LLMs) have made significant strides in text-to-SQL tasks; however, small language models (SLMs) are crucial due to their low resource consumption and efficient inference for real-world deployment. Due to resource limitations, SLMs struggle to accurately interpret natural language questions and may overlook critical constraints, leading to challenges such as generating SQL with incorrect logic or incomplete conditions. To address these issues, we propose PARSQL, a novel framework that leverages SQL parsing and reasoning. Specifically, we design PARSer, an SQL parser that extracts constraints from SQL to generate sub-SQLs for data augmentation and producing step-by-step SQL explanations (reason) via both rule-based and LLM-based methods. We define a novel text-to-reason task and incorporate it into multi-task learning, thereby enhancing text-to-SQL performance. Additionally, we employ an efficient SQL selection strategy that conducts direct similarity computation between the generated SQLs and their corresponding reasons to derive the final SQL for post-correction. Extensive experiments show that our PARSQL outperforms models with the same model size on the BIRD and Spider benchmarks. Notably, PARSQL-3B achieves 56.98% execution accuracy on BIRD, rivaling 7B models with significantly fewer parameters, setting a new state-of-the-art performance. Code can be found [here](https://github.com/yaxundai/parsql).

The Middle East is characterized by remarkable linguistic diversity, with over 400 million inhabitants speaking more than 60 languages across multiple language families. This study presents a pioneering work in developing the first parallel corpora for eight severely under-resourced varieties in the region–PARME, addressing fundamental challenges in low-resource scenarios including non-standardized writing and dialectal complexity. Through an extensive community-driven initiative, volunteers contributed to the creation of over 36,000 translated sentences, marking a significant milestone in resource development. We evaluate machine translation capabilities through zero-shot approaches and fine-tuning experiments with pretrained machine translation models and provide a comprehensive analysis of limitations. Our findings reveal significant gaps in existing technologies for processing the selected languages, highlighting critical areas for improvement in language technology for Middle Eastern languages.

Dealing with long and highly complex technical text is a challenge for Large Language Models (LLMs), which still have to unfold their potential in supporting expensive and time intensive processes like patent drafting. Within patents, the description constitutes more than 90% of the document on average. Yet, its automatic generation remains understudied. When drafting patent applications, patent attorneys typically receive invention reports (IRs), which are usually confidential, hindering research on LLM-supported patent drafting.Often, pre-publication research papers serve as IRs. We leverage this duality to build PAP2PAT, an open and realistic benchmark for patent drafting consisting of 1.8k patent-paper pairs describing the same inventions. To address the complex long-document patent generation task, we propose chunk-based outline-guided generation using the research paper as invention specification. Our extensive evaluation using PAP2PAT and a human case study show that LLMs can effectively leverage information from the paper, but still struggle to provide the necessary level of detail. Fine-tuning leads to more patent-style language, but also to more hallucination. We release our data and code.

Paraphrasing is rooted in semantics, which makes evaluating paraphrase generation systems hard. Current paraphrase generators are typically evaluated using borrowed metrics from adjacent text-to-text tasks, like machine translation or text summarization. These metrics tend to have ties to the surface form of the reference text. This is not ideal for paraphrases as we typically want variation in the lexicon while persisting semantics. To address this problem, and inspired by learned similarity evaluation on plain text, we propose PAM, a Paraphrase AMR-Centric Evaluation Metric. This metric uses AMR graphs extracted from the input text, which consist of semantic structures agnostic to the text surface form, making the resulting evaluation metric more robust to variations in syntax or lexicon. Additionally, we evaluated PAM on different semantic textual similarity datasets and found that it improves the correlations with human semantic scores when compared to other AMR-based metrics.

Current large language model (LLM) applications often employ multi-component prompts, comprising both system and user prompts, to guide model behaviors. While recent advancements have demonstrated the efficacy of automatically optimizing either the system or user prompt to boost performance, such unilateral approaches often yield suboptimal outcomes due to the interdependent nature of these components. In this work, we introduce P3, a novel self-improvement framework that concurrently optimizes both system and user prompts through an iterative process. The offline optimized prompts are further leveraged to promote online prompting by performing query-dependent prompt optimization. Extensive experiments on general tasks (e.g., Arena-hard and Alpaca-eval) and reasoning tasks (e.g., GSM8K and GPQA) demonstrate that P3 achieves superior performance in the realm of automatic prompt optimization. Our results highlight the effectiveness of a holistic optimization strategy in enhancing LLM performance across diverse domains.

Pruning has become a widely adopted technique for reducing the hardware requirements of large language models (LLMs). To recover model performance after pruning, post-training is commonly employed to mitigate the resulting performance degradation. While post-training benefits from larger datasets, once the dataset size is already substantial, increasing the training data provides only limited performance gains. To balance post-training cost and model performance, it is necessary to explore the optimal amount of post-training data. Through extensive experiments on the Llama-3 and Qwen-2.5 series models, pruned using various common pruning methods, we uncover the scaling Law for Post-training after model Pruning, referred to as the P2 Law. This law identifies four key factors for predicting the pruned model’s post-training loss: the model size before pruning, the number of post-training tokens, the pruning rate, and the model’s loss before pruning. Moreover, P2 Law can generalize to larger dataset sizes, larger model sizes, and higher pruning rates, offering valuable insights for the post-training of pruned LLMs.

Personalized large language models (LLMs) have attracted great attention in many applications, such as emotional support and role-playing. However, existing works primarily focus on modeling explicit character profiles, while ignoring the underlying personality traits that truly shape behaviors and decision-making, hampering the development of more anthropomorphic and psychologically-grounded AI systems. In this paper, we explore the modeling of Big Five personality traits, which is the most widely used trait theory in psychology, and propose P-React, a mixture of experts (MoE)-based personalized LLM. Particularly, we integrate a Personality Specialization Loss (PSL) to better capture individual trait expressions, providing a more nuanced and psychologically grounded personality simulacrum. To facilitate research in this field, we curate OCEAN-Chat, a high-quality, human-verified dataset designed to train LLMs in expressing personality traits across diverse topics. Extensive experiments demonstrate the effectiveness of P-React in maintaining consistent and real personality.

This study explores the potential of phonological reasoning within text-based large language models (LLMs). Utilizing the PhonologyBench benchmark, we assess tasks like rhyme word generation, g2p conversion, and syllable counting. Our evaluations across 12 LLMs reveal that while few-shot learning offers inconsistent gains, the introduction of a novel Pedagogically-motivated Participatory Chain-of-Thought (P-CoT) prompt, which is anchored in educational theories like scaffolding and discovery learning, consistently enhances performance. This method leverages structured guidance to activate latent phonological abilities, achieving up to 52% improvement and even surpassing human baselines in certain tasks. Future work could aim to optimize P-CoT prompts for specific models or explore their application across different linguistic domains.

The rapid advancements in Large Language Models (LLMs) have revolutionized various natural language processing tasks. However, the substantial size of LLMs presents significant challenges in training or fine-tuning. While parameter-efficient approaches such as low-rank adaptation (LoRA) have gained popularity, they often compromise performance compared to full-rank fine-tuning. In this paper, we propose Outlier-weighed Layerwise Sampling (OWS), a new memory-efficient fine-tuning approach, inspired by the layerwise outlier distribution of LLMs. Unlike LoRA, which adds extra adapters to all layers, OWS strategically assigns higher sampling probabilities to layers with more outliers, selectively sampling only a few layers and fine-tuning their pre-trained weights. To further increase the number of fine-tuned layers without a proportional rise in memory costs, we incorporate gradient low-rank projection, further boosting the approach’s performance. Our extensive experiments across various architectures, including LLaMa2 and Mistral, demonstrate that OWS consistently outperforms baseline approaches, including full fine-tuning. Specifically, it achieves up to a 1.1% average accuracy gain on the Commonsense Reasoning benchmark, a 3.0% improvement on MMLU, and a notable 10% boost on MT-Bench, while being more memory efficient. OWS allows us to fine-tune 7B LLMs with only 21GB of memory. Our code is available at https://github.com/pixeli99/OWS.

Extreme activation outliers in Large Language Models (LLMs) critically degrade quantization performance, hindering efficient on-device deployment. While channel-wise operations and adaptive gradient scaling are recognized causes, practical mitigation remains challenging. We introduce **Outlier-Safe Pre-Training (OSP)**, a practical guideline that proactively prevents outlier formation, rather than relying on post-hoc mitigation. OSP combines three key innovations: (1) the Muon optimizer, eliminating privileged bases while maintaining training efficiency, (2) Single-Scale RMSNorm, preventing channel-wise amplification, and (3) a learnable embedding projection, redistributing activation magnitudes. We validate OSP by training a 1.4B-parameter model on 1 trillion tokens, which is the first production-scale LLM trained without such outliers. Under aggressive 4-bit quantization, our OSP model achieves a 35.7 average score across 10 benchmarks (versus 26.5 for an Adam-trained model), with only a 2% training overhead. Remarkably, OSP models exhibit near-zero excess kurtosis (0.04) compared to extreme values (1818.56) in standard models, fundamentally altering LLM quantization behavior. Our work demonstrates that outliers are not inherent to LLMs but are consequences of training strategies, paving the way for more efficient LLM deployment.

Text-Attributed Graphs (TAGs), which are characterized with text attributes, are widely used in the real world. When evaluating fully trained models designed for TAG predictions, they may perform significantly unsatisfactory on samples outside the In-Distribution (ID) data, which may raise serious security issues. To tackle it, Out-Of-Distribution (OOD) detection is introduced to the TAGs field, which aims to utilize a detector to classify OOD and ID samples. Recent studies attempt to introduce extra OOD datasets to regularize the detection model. However, due to the vastness of the OOD data space, high-quality OOD samples for training the detector are scarce and difficult to obtain in the real world. Thus, we utilize Large Language Models (LLMs) to generate the OOD training samples with high quality. There are two issues in this process: (1) LLMs tend to generate OOD-node samples significantly different from ID ones, with a limited learning value for OOD and ID relations. (2) Due to the inherent structure of TAGs, obtained OOD nodes need to be integrated with existing nodes by generating edges using LLMs. However, the large number of nodes makes reasoning over each node pair computationally unbearable. Toward these issues, we introduce LLMGuard with challenging OOD-node generation and lightweight edge predictors. Extensive experiments prove the effectiveness of LLMGuard. The source code is available.

Real-world instructions with multiple constraints pose a significant challenge to existing large language models (LLMs). An observation is that the LLMs exhibit dramatic performance fluctuation when disturbing the order of the incorporated constraints. Yet, none of the existing works has systematically investigated this position bias problem in the field of multi-constraint instruction following. To bridge this gap, we design a probing task where we quantitatively measure the difficulty distribution of the constraints by a novel Difficulty Distribution Index (CDDI). Through the experimental results, we find that LLMs are more performant when presented with the constraints in a “hard-to-easy” order. This preference can be generalized to LLMs with different architecture or different sizes of parameters. Additionally, we conduct an explanation study, providing an intuitive insight into the correlation between the LLM’s attention and constraint orders. Our code and dataset are publicly available at https://github.com/meowpass/PBIF.

Text-to-SQL demands precise reasoning to convert natural language questions into structured queries. While large language models (LLMs) excel in many reasoning tasks, their ability to leverage Chain-of-Thought (CoT) reasoning for text-to-SQL remains underexplored. We identify critical limitations: zero-shot CoT offers minimal gains, and Direct Preference Optimization (DPO) applied without CoT yields marginal improvements. We propose ExCoT-DPO, a novel framework that iteratively optimizes open-source LLMs by combining CoT reasoning with off-policy and on-policy DPO, relying solely on execution accuracy as feedback. This approach eliminates the need for reward models or human-annotated preferences. Our experimental results demonstrate significant performance gains: ExCoT-DPO improves execution accuracy on BIRD from 57.37% to 68.51% and on Spider from 78.81% to 86.59% for LLaMA-3 70B, with Qwen-2.5-Coder demonstrating similar improvements. Our best model achieves state-of-the-art performance in the single-model setting on both BIRD and Spider datasets.

Retrieval-augmented generation (RAG) is usually integrated into large language models (LLMs) to mitigate hallucinations and knowledge obsolescence. Whereas, conventional one-step retrieve-and-read methods are insufficient for multi-hop question answering, facing challenges of retrieval semantic mismatching and the high cost in handling interdependent subquestions. In this paper, we propose Optimizing Question Semantic Space for Dynamic Retrieval-Augmented Multi-hop Question Answering (Q-DREAM). Q-DREAM consists of three key modules: (1) the Question Decomposition Module (QDM), which decomposes multi-hop questions into fine-grained subquestions; (2) the Subquestion Dependency Optimizer Module (SDOM), which models the interdependent relations of subquestions for better understanding; and (3) the Dynamic Passage Retrieval Module (DPRM), which aligns subquestions with relevant passages by optimizing the semantic embeddings.Experimental results across various benchmarks demonstrate that Q-DREAM significantly outperforms existing RAG methods, achieving state-of-the-art performance in both in-domain and out-of-domain settings. Notably, Q-DREAM also improves retrieval efficiency while maintaining high accuracy compared with recent baselines.

We propose a method to optimize language model pre-training data mixtures through efficient approximation of the cross-entropy loss corresponding to each candidate mixture via a Mixture of Data Experts (MDE). We use this approximation as a source of additional features in a regression model, trained from observations of model loss for a small number of mixtures. Experiments with Transformer decoder-only language models in the range of 70M to 10B parameters on the SlimPajama dataset show that our method achieves significantly better performance than approaches that train regression models using only the mixture rates as input features. Combining this improved optimization method with an objective that takes into account cross-entropy on end task data leads to superior performance on few-shot downstream evaluations. We also provide theoretical insights on why aggregation of data expert predictions can provide good approximations to model losses for data mixtures.

Retrieval-augmented generation (RAG) encounters challenges when addressing complex queries, particularly multi-hop questions. While several methods tackle multi-hop queries by iteratively generating internal queries and retrieving external documents, these approaches are computationally expensive. In this paper, we identify a three-stage information processing pattern in LLMs during layer-by-layer reasoning, consisting of extraction, processing, and subsequent extraction steps. This observation suggests that the representations in intermediate layers contain richer information compared to those in other layers. Building on this insight, we propose Layer-wise RAG (L-RAG). Unlike prior methods that focus on generating new internal queries, L-RAG leverages intermediate representations from the middle layers, which capture next-hop information, to retrieve external knowledge. L-RAG achieves performance comparable to multi-step approaches while maintaining inference overhead similar to that of standard RAG. Experimental results show that L-RAG outperforms existing RAG methods on open-domain multi-hop question-answering datasets, including MuSiQue, HotpotQA, and 2WikiMultiHopQA. The code is available in https://anonymous.4open.science/r/L-RAG-ADD5/.