NAACL 2025 Papers

Self-consistency mitigates hallucinations in Large Language Models (LLMs) by sampling multiple reasoning paths, but it lacks a systematic approach to determine the optimal number of samples or select the most faithful rationale. To address this limitation, we introduce Reasoning-Aware Self-Consistency (RASC), a novel framework that enhances sampling efficiency and reasoning faithfulness by dynamically evaluating both outputs and rationales. RASC assesses the quality of reasoning and the consistency of answers for each generated sample, using these assessments to guide early stopping decisions and rationale selection. The framework employs criteria-based stopping and weighted majority voting, enabling more informed choices on when to halt sampling and which rationale to select. Our comprehensive experiments across diverse question-answering datasets demonstrate that RASC outperforms existing methods, reducing sample usage by approximately 70% while maintaining accuracy. Moreover, RASC facilitates the selection of high-fidelity rationales, thereby improving the faithfulness of LLM outputs. Our approach effectively addresses the efficiency-accuracy trade-off in LLM reasoning tasks, offering a new perspective for more nuanced, faithful, and effective utilization of LLMs in resource-constrained environments.

In a rapidly globalizing and digital world, content such as book and product reviews created by people from diverse cultures are read and consumed by others from different corners of the world. In this paper, we investigate the extent and patterns of gaps in understandability of book reviews due to the presence of culturally-specific items and elements that might be alien to users from another culture. Our user-study on 57 book reviews from Goodreads reveal that 83% of the reviews had at least one culture-specific difficult-to-understand element. We also evaluate the efficacy of GPT-4o in identifying such items, given the cultural background of the reader; the results are mixed, implying a significant scope for improvement. Our datasets are available here: https://github.com/sougata-ub/reading_between_lines.

Recently, mobile AI agents have gained increasing attention. Given a task, mobile AI agents can interact with mobile devices in multiple steps and finally form a GUI flow that solves the task. However, existing agents tend to focus on most task-relevant elements at each step, leading to local optimal solutions and ignoring the overall GUI flow. To address this issue, we constructed a training dataset called MobileReach, which breaks the task into page reaching and operation subtasks. Furthermore, we propose ReachAgent, a two-stage framework that focuses on improving its task-completion abilities. It utilizes the page reaching and page operation subtasks, along with reward-based preference GUI flows, to further enhance the agent. Experimental results show that ReachAgent significantly improves the Intersection over Union (IoU) Accuracy and Text Accuracy by 7.12% and 7.69% on the step-level and 4.72% and 4.63% on the task-level compared to the SOTA agent. Our data and code will be released upon acceptance.

In this study, we introduce RePD, an innovative attack Retrieval-based Prompt Decomposition framework designed to mitigate the risk of jailbreak attacks on large language models (LLMs). Despite rigorous pre-training and fine-tuning focused on ethical alignment, LLMs are still susceptible to jailbreak exploits. RePD operates on a one-shot learning model, wherein it accesses a database of pre-collected jailbreak prompt templates to identify and decompose harmful inquiries embedded within user prompts. This process involves integrating the decomposition of the jailbreak prompt into the user’s original query into a one-shot learning example to effectively teach the LLM to discern and separate malicious components. Consequently, the LLM is equipped to first neutralize any potentially harmful elements before addressing the user’s prompt in a manner that aligns with its ethical guidelines. RePD is versatile and compatible with a variety of open-source LLMs acting as agents. Through comprehensive experimentation with both harmful and benign prompts, we have demonstrated the efficacy of our proposed RePD in enhancing the resilience of LLMs against jailbreak attacks, without compromising their performance in responding to typical user requests.

The automatic evaluation of instruction following typically involves using large language models (LLMs) to assess response quality. However, there is a lack of comprehensive evaluation of these LLM-based evaluators across two dimensions: the base LLMs and the evaluation protocols. Therefore, we present a thorough meta-evaluation of instruction following, including 25 base LLMs and 15 recently proposed evaluation protocols, on 4 human-annotated datasets, assessing the evaluation accuracy of the LLM-evaluators. Our evaluation allows us to identify the best-performing base LLMs and evaluation protocols with a high degree of robustness. Moreover, our evaluation reveals key findings: (1) Base LLM performance ranking remains largely consistent across evaluation protocols, with less capable LLMs showing greater improvement from protocol enhancements; (2) Robust evaluation of evaluation protocols requires many base LLMs with varying capability levels, as protocol effectiveness depends on the base LLM used; (3) Evaluation results on different datasets are not always consistent, so a rigorous evaluation requires multiple datasets with distinctive features. We release our meta-evaluation suite ReIFE, which provides the codebase and evaluation result collection for over 500 LLM-evaluators, laying groundwork for future research in instruction-following evaluation.

Recent advancements in Large Language Models (LLMs) have set themselves apart with their exceptional performance in complex language modelling tasks. However, these models are also known for their significant computational and storage requirements, primarily due to the quadratic computation complexity of softmax attention. To mitigate this issue, linear attention has been designed to reduce the quadratic space-time complexity that is inherent in standard transformers. In this work, we embarked on a comprehensive exploration of three key components that substantially impact the performance of the Gated Linear Attention module: feature maps, normalization, and the gating mechanism. We developed a feature mapping function to address some crucial issues that previous suggestions overlooked. Then we offered further rationale for the integration of normalization layers to stabilize the training process. Moreover, we explored the saturation phenomenon of the gating mechanism and augmented it with a refining module. We conducted extensive experiments and showed our architecture outperforms previous Gated Linear Attention mechanisms in extensive tasks including training from scratch and post-linearization with continual pre-training.

Evaluating and ranking the capabilities of different LLMs is crucial for understanding their performance and alignment with human preferences. Due to the high cost and time-consuming nature of human evaluations, an automatic LLM bencher (i.e., an automatic evaluation framework that aims to rank LLMs based on their alignment with human preferences) is indispensable. An automatic LLM bencher consists of four components: the input set (e.g., a user instruction), the evaluation model (e.g., an LLM), the evaluation type (e.g., pairwise comparison), and the aggregation method (e.g., the ELO rating system). However, previous work has not thoroughly explored how to select these components or how their different combinations influence the results. In this work, through controlled experiments, we provide a series of recommendations on how to choose each component to better automate the evaluation of LLMs. Furthermore, we discovered that when evaluating LLMs with similar performance, the performance of the automatic LLM bencher declines sharply, underscoring the limitations of current benchers and calling for future work. Lastly, we found that the evaluation models’ performance at the instance level (e.g., the accuracy of selecting the best output) does not always align with their effectiveness when used as a component of a bencher, highlighting the importance of dedicated system-level evaluation of benchers.

Large language models (LLM) hold significant potential for applications in biomedicine, but they struggle with hallucinations and outdated knowledge.While retrieval-augmented generation (RAG) is generally employed to address these issues, it also has its own set of challenges: (1) LLMs are vulnerable to irrelevant or unhelpful context, (2) medical queries are often not well-targeted for helpful information, and (3) retrievers are prone to bias toward the specific source corpus they were trained on. In this study, we present RAG2 (RAtionale-Guided RAG), a new framework for enhancing the reliability of RAG in biomedical contexts. RAG2 incorporates three key innovations: a small filtering model trained on perplexity-based labels of rationales, which selectively augments informative snippets of documents while filtering out distractors; LLM-generated rationales as queries to improve the utility of retrieved snippets; a structure designed to retrieve snippets evenly from a comprehensive set of four biomedical corpora, effectively mitigating retriever bias. Our experiments demonstrate that RAG2 improves the state-of-the-art LLMs of varying sizes, with improvements of up to 6.1%, and it outperforms the previous best medical RAG model by up to 5.6% across three medical question-answering benchmarks. Our code is available at https://github.com/dmis-lab/RAG2

Existing automated essay scoring (AES) has solely relied on essay text without using explanatory rationales for the scores, thereby forgoing an opportunity to capture the specific aspects evaluated by rubric indicators in a fine-grained manner. This paper introduces Rationale-based Multiple Trait Scoring (RMTS), a novel approach for multi-trait essay scoring that integrates prompt-engineering-based large language models (LLMs) with a fine-tuning-based essay scoring model using a smaller large language model (S-LLM). RMTS uses an LLM-based trait-wise rationale generation system where a separate LLM agent generates trait-specific rationales based on rubric guidelines, which the scoring model uses to accurately predict multi-trait scores. Extensive experiments on benchmark datasets, including ASAP, ASAP++, and Feedback Prize, show that RMTS significantly outperforms state-of-the-art models and vanilla S-LLMs in trait-specific scoring. By assisting quantitative assessment with fine-grained qualitative rationales, RMTS enhances the trait-wise reliability, providing partial explanations about essays. The code is available at https://github.com/BBeeChu/RMTS.git.

The efficient compression of large language models (LLMs) has become increasingly popular. However, recovering the performance of compressed LLMs remains a major challenge. The current practice in LLM compression entails the implementation of structural pruning, complemented by a recovery phase that leverages the Low-Rank Adaptation (LoRA) algorithm. Structural pruning’s uneven modification of model architecture, coupled with standard LoRA’s fixed configuration allocation across layers in an online pipeline, leads to suboptimal performance in various downstream tasks for pruned models. To address this challenge, we introduce RankAdaptor, a hierarchical rank allocation method that enables efficient fine-tuning of pruned LLMs according to layerwise specific recovery requirements. We employ a performance model that conducts offline meta-learning and online incremental learning to explore optimal rank values for each layer. Comprehensive experiments on popular benchmarks show that RankAdaptor consistently outperforms state-of-the-art methods across a variety of pruning settings and LLM architectures, with improvements ranging from 0.7% to 5.5%.

The profound success of transformer-based language models can largely be attributed to their ability to integrate relevant contextual information from an input sequence in order to generate a response or complete a task. However, we know very little about the algorithms that a model employs to implement this capability, nor do we understand their failure modes. For example, given the prompt “John is going fishing, so he walks over to the bank. Can he make an ATM transaction?”, a model may incorrectly respond “Yes” if it has not properly contextualized “bank” as a geographical feature, rather than a financial institution. We propose the LLM Race Conditions Hypothesis as an explanation of contextualization errors of this form. This hypothesis identifies dependencies between tokens (e.g., “bank” must be properly contextualized before the final token, "?", integrates information from “bank”), and claims that contextualization errors are a result of violating these dependencies. Using a variety of techniques from mechanistic interpretability, we provide correlational and causal evidence in support of the hypothesis and suggest inference-time interventions to address it.

Semantic matching plays a pivotal role in e-commerce by facilitating better product discovery and driving sales within online stores. Transformer models have proven exceptionally effective in mapping queries to an embedding space, positioning semantically related entities (queries or products) in close proximity. Despite their effectiveness, the high computational demands of large transformer models pose challenges for their deployment in real-time scenarios. This paper presents RTSM, an advanced knowledge distillation framework designed for Real-Time Semantic Matching. Our approach develops accurate, low-latency student models by leveraging both soft labels from a teacher model and ground truth generated from pairwise query-product and query-query signals. These signals are sourced from direct audits, synthetic examples created by LLMs, user interaction data, and taxonomy-based datasets, with custom loss functions enhancing learning efficiency. Experimental evaluations on internal and external e-commerce datasets demonstrate a 2-2.5% increase in ROC-AUC compared to directly trained student models, outperforming both the teacher model and state-of-the-art knowledge distillation benchmarks.

This paper addresses the task of legal summarization, which involves distilling complex legal documents into concise, coherent summaries. Current approaches often struggle with content theme deviation and inconsistent writing styles due to their reliance solely on source documents. We propose RELexED, a retrieval-augmented framework that utilizes exemplar summaries along with the source document to guide the model. RELexED employs a two-stage exemplar selection strategy, leveraging a determinantal point process to balance the trade-off between similarity of exemplars to the query and diversity among exemplars, with scores computed via influence functions. Experimental results on two legal summarization datasets demonstrate that RELexED significantly outperforms models that do not utilize exemplars and those that rely solely on similarity-based exemplar selection.

The ability to communicate uncertainty and knowledge limitations is crucial for the safety of large language models (LLMs). Current evaluations of these abilities typically examine the correspondence between model accuracy and its internal probabilities or linguistic outputs. However, evaluation of the uncertainty of LLM communication should also focus on the behaviors of their human interlocutors: how much do users rely on what the LLM says? We introduce an interaction-centered evaluation approach called Rel-A.I. (pronounced “rely”) that quantifies whether and how humans rely on LLMs’ responses, complementing existing calibration evaluations. Through nine user studies with 450 participants, we investigate three crucial aspects that influence user reliance. We show that emphatic expressions of politeness (e.g., “I’m happy to help!”) that precede LLM answers will cause participants to perceive these models as more competent, and in turn, rely 30% more on their generations. Additionally, the context of the interaction, such as the knowledge domain and nature of previous interactions with the LLM, substantially influences user reliance (e.g., users will rely 10% more on LLMs when responding to questions involving calculations). Our results show that calibration and language quality alone are insufficient in informing which LLMs are safely calibrated, and illustrate the need to consider features of the interactional context.

Automatic melody-to-lyric (M2L) generation aims to create lyrics that align with a given melody. While most previous approaches generate lyrics from scratch, revision—editing plain text draft to fit it into the melody—offers a much more flexible and practical alternative. This enables broad applications, such as generating lyrics from flexible inputs (keywords, themes, or full text that needs refining to be singable), song translation (preserving meaning across languages while keeping the melody intact), or style transfer (adapting lyrics to different genres). This paper introduces REFFLY (REvision Framework For LYrics), the first revision framework for editing and generating melody-aligned lyrics. We train the lyric revision module using our curated synthesized melody-aligned lyrics dataset, enabling it to transform plain text into lyrics that align with a given melody. To further enhance the revision ability, we propose training-free heuristics aimed at preserving both semantic meaning and musical consistency throughout the editing process. Experimental results demonstrate the effectiveness of REFFLY across various tasks (e.g. song translation), showing that our model outperforms strong baselines, including Lyra (CITATION) and GPT-4, by 25% in both musicality and text quality.

Social media provides a valuable lens for assessing public perceptions and opinions. This paper focuses on the concept of truthfulness stance, which evaluates whether a textual utterance affirms, disputes, or remains neutral or indifferent toward a factual claim. Our systematic analysis fills a gap in the existing literature by offering the first in-depth conceptual framework encompassing various definitions of stance. We introduce RATSD (Retrieval Augmented Truthfulness Stance Detection), a novel method that leverages large language models (LLMs) with retrieval-augmented generation (RAG) to enhance the contextual understanding of tweets in relation to claims. RATSD is evaluated on TSD-CT, our newly developed dataset containing 3,105 claim-tweet pairs, along with existing benchmark datasets. Our experiment results demonstrate that RATSD outperforms state-of-the-art methods, achieving a significant increase in Macro-F1 score on TSD-CT. Our contributions establish a foundation for advancing research in misinformation analysis and provide valuable tools for understanding public perceptions in digital discourse.

Large Language Models (LLMs) have significantly advanced natural language processing, but content repetition in open-source LLMs remains a critical challenge that adversely affects user experience. The repetition penalty parameter (RPP) aims to mitigate this issue by preventing repeated content generation, but excessive use of RPP can compromise the overall quality. In this paper, we propose Repetition-Aware Performance (RAP), a novel evaluation metric that quantifies and integrates repetition penalty into the assessment of model performance, enabling tuning of RPP. We evaluate our approach using twelve open-source LLMs, ranging from 2 billion to 70 billion parameters, tested on question answering and machine translation tasks across three datasets with varying prompting techniques. Experimental results show that RAP effectively tunes RPP, helping to identify a trade-off value that significantly reduces repetition while minimizing performance loss. Upon acceptance, we will release the code and the dataset of generated text, providing a valuable resource for further research on repetition detection and LLMs evaluation.

Multi-modal retrieval-augmented Question Answering (MRAQA), integrating text and images, has gained significant attention in information retrieval (IR) and natural language processing (NLP). Traditional ranking methods rely on small encoder-based language models, which are incompatible with modern decoder-based generative large language models (LLMs) that have advanced various NLP tasks. To bridge this gap, we propose RAMQA, a unified framework combining learning-to-rank methods with generative permutation-enhanced ranking techniques. We first train a pointwise multi-modal ranker using LLaVA as the backbone. Then, we apply instruction tuning to train a LLaMA model for re-ranking the top-k documents using an innovative autoregressive multi-task learning approach. Our generative ranking model generates re-ranked document IDs and specific answers from document candidates in various permutations. Experiments on two MRAQA benchmarks, WebQA and MultiModalQA, show significant improvements over strong baselines, highlighting the effectiveness of our approach. Data and code will be made public once the paper is accepted.

Existing large language models (LLMs) show exceptional problem-solving capabilities but might struggle with complex reasoning tasks. Despite the successes of chain-of-thought and tree-based search methods, they mainly depend on the internal knowledge of LLMs to search over intermediate reasoning steps, limited to dealing with simple tasks involving fewer reasoning steps. In this paper, we propose RAG-Star, a novel RAG approach that integrates the retrieved information to guide the tree-based deliberative reasoning process that relies on the inherent knowledge of LLMs. By leveraging Monte Carlo Tree Search, RAG-Star iteratively plans intermediate sub-queries and answers for reasoning based on the LLM itself. To consolidate internal and external knowledge, we propose a retrieval-augmented verification that utilizes query- and answer-aware reward modeling to provide feedback for the inherent reasoning of LLMs. Our experiments involving Llama-3.1-8B-Instruct and GPT-4o demonstrate that RAG-Star significantly outperforms previous RAG and reasoning methods. Our codes and data are publicly available at https://github.com/RUCAIBox/RAG-Star.