NAACL 2025 论文列表

The success of Transformer-based Language Models (LMs) stems from their attention mechanism. While this mechanism has been extensively studied in explainability research, particularly through the attention values obtained during the forward pass of LMs, the backward pass of attention has been largely overlooked.In this work, we study the mathematics of the backward pass of attention, revealing that it implicitly calculates an attention matrix we refer to as “Reversed Attention”.We visualized Reversed Attention and examine its properties, demonstrating its ability to elucidate the models’ behavior and edit dynamics.In an experimental setup, we showcase the ability of Reversed Attention to directly alter the forward pass of attention, without modifying the model’s weights, using a novel method called “attention patching”.In addition to enhancing the comprehension of how LMs configure attention layers during backpropagation, Reversed Attention maps contribute to a more interpretable backward pass.

Reverse thinking plays a crucial role in human reasoning. Humans can reason not only from a problem to a solution but also in reverse, i.e., start from the solution and reason towards the problem. This often enhances overall reasoning performance as it enables consistency checks between their forward and backward thinking. To enable Large Language Models (LLMs) to perform reverse thinking, we introduce Reverse-Enhanced Thinking (RevThink), a framework composed of data augmentation and learning objectives. In RevThink, we augment the dataset by collecting structured forward-backward reasoning from a teacher model, consisting of: (1) the original question, (2) forward reasoning, (3) backward question, and (4) backward reasoning. We then employ three objectives to train a smaller student model in a multi-task learning fashion: (a) generate forward reasoning from a question, (b) generate a backward question from a question, and (c) generate backward reasoning from the backward question. Experiments across 12 datasets covering commonsense, math, and logical reasoning show an average 13.53% improvement over the student model’s zero-shot performance and a 6.84% improvement over the strongest knowledge distillation baselines. Moreover, our method demonstrates sample efficiency – using only 10% of the correct forward reasoning from the training data, it outperforms a standard fine-tuning method trained on 10x more forward reasoning. RevThink also exhibits strong generalization to out-of-distribution held-out datasets.

Question answering (QA)—giving correct answers to questions—is a popular task, but we test **reverse question answering (RQA)**: for an input answer, give a question with that answer. Past work tests QA and RQA separately, but we test them jointly, comparing their difficulty, aiding benchmark design, and checking reasoning consistency. We run 16 LLMs on QA and RQA with trivia questions/answers, revealing: 1) Versus RQA, LLMs are much less accurate in RQA for numerical answers, but slightly more accurate in RQA for textual answers; 2) LLMs often answer their own invalid questions from RQA accurately in QA, so RQA errors are not just from knowledge gaps; 3) RQA errors correlate with question difficulty and inversely correlate with answer frequencies in the Dolma corpus; and 4) LLMs struggle to give valid multi-hop questions. By finding question and answer types that lead to RQA errors, we suggest improvements for LLM reasoning.

Humans are accustomed to reading and writing in a forward manner, and this natural bias extends to text understanding in auto-regressive large language models (LLMs). This paper investigates whether LLMs, like humans, struggle with reverse modeling, specifically with reversed text inputs. We found that publicly available pre-trained LLMs cannot understand such inputs. However, LLMs trained from scratch with both forward and reverse texts can understand them equally well during inference across multiple languages.Our case study shows that different-content texts result in different losses if input (to LLMs) in different directions—some get lower losses for forward while some for reverse. This leads us to a simple and nice solution for data selection based on the loss differences between forward and reverse directions. Using our selected data in continued pretraining can boost LLMs’ performance by a large margin across different language understanding benchmarks.

Autonomous planning has been an ongoing pursuit since the inception of artificial intelligence. Based on curated problem solvers, early planning agents could deliver precise solutions for specific tasks but lacked generalization. The emergence of large language models (LLMs) and their powerful reasoning capabilities has reignited interest in autonomous planning by automatically generating reasonable solutions for given tasks. However, prior research and our experiments show that current language agents still lack human-level planning abilities. Even the state-of-the-art reasoning model, OpenAI o1, achieves only 15.6% on one of the complex real-world planning benchmarks. This highlights a critical question: What hinders language agents from achieving human-level planning? Although existing studies have highlighted weak performance in agent planning, the deeper underlying issues and the mechanisms and limitations of the strategies proposed to address them remain insufficiently understood. In this work, we apply the feature attribution study and identify two key factors that hinder agent planning: the limited role of constraints and the diminishing influence of questions. We also find that although current strategies help mitigate these challenges, they do not fully resolve them, indicating that agents still have a long way to go before reaching human-level intelligence.

Large language models (LLMs) have demonstrated impressive capabilities in generating human-like text and have been shown to store factual knowledge within their extensive parameters. However, models like ChatGPT can still actively or passively generate false or misleading information, increasing the challenge of distinguishing between human-created and machine-generated content. This poses significant risks to the authenticity and reliability of digital communication. This work aims to enhance retrieval models’ ability to identify the authenticity of texts generated by large language models, with the goal of improving the truthfulness of retrieved texts and reducing the harm of false information in the era of large models. Our contributions include: (1) we construct a diverse dataset of authentic human-authored texts and highly deceptive AI-generated texts from various domains; (2) we propose a self-supervised training method, RetrieverGuard, that enables the model to capture textual rules and styles of false information from the corpus without human-labelled data, achieving higher accuracy and robustness in identifying misleading and highly deceptive AI-generated content.

The Knowledge Graph Completion (KGC) task aims to infer the missing entity from an incomplete triple. Existing embedding-based methods rely solely on triples in the KG, which is vulnerable to specious relation patterns and long-tail entities. On the other hand, text-based methods struggle with the semantic gap between KG triples and natural language. Apart from triples, entity contexts (e.g., labels, descriptions, aliases) also play a significant role in augmenting KGs. To address these limitations, we propose KGR3, a context-enriched framework for KGC. KGR3 is composed of three modules. Firstly, the Retrieval module gathers supporting triples from the KG, collects plausible candidate answers from a base embedding model, and retrieves context for each related entity. Then, the Reasoning module employs a large language model to generate potential answers for each query triple. Finally, the Re-ranking module combines candidate answers from the two modules mentioned above, and fine-tunes an LLM to provide the best answer. Extensive experiments on widely used datasets demonstrate that KGR3 consistently improves various KGC methods. Specifically, the best variant of KGR3 achieves absolute Hits@1 improvements of 12.3% and 5.6% on the FB15k237 and WN18RR datasets.

Document-level relation extraction (DocRE) provides a broad context for extracting one or more relations for each entity pair. Large language models (LLMs) have made great progress in relation extraction tasks. However, one of the main challenges we face is that LLMs have difficulty in multi-label relation prediction tasks. Additionally, another noteworthy challenge and discovery we reveal: the small language models (SLMs) for DocRE tend to classify existing relations as ”no relation” (NA), while LLMs tend to predict existing relations for all entity pairs. To address these challenges, we propose a novel method that utilizes LLMs as a refiner, employing task distribution and probability fusion. The task distribution we carefully designed aims to distinguish hard and easy tasks, and feed hard tasks to our LLMs-based framework to reevaluate and refine. Further, in order to effectively solve the multi-label relation prediction problem in the refinement process, we propose a probability fusion method, ensuring and enhancing fusion predictions by maintaining a balance between SLMs and LLMs. Extensive experiments on widely-used datasets demonstrate that our method outperforms existing LLMbased methods without fine-tuning by an average of 25.2% F1. Refining SLMs using our method consistently boosts the performance of the SLMs, achieving new state-of-the-art results compared to existing SLMs and LLMs. Our code: https://github.com/Drasick/Drell.

Word similarity has many applications to social science and cultural analytics tasks like measuring meaning change over time and making sense of contested terms. Yet traditional similarity methods based on cosine similarity between word embeddings cannot capture the context-dependent, asymmetrical, polysemous nature of semantic similarity. We propose a new measure of similarity, Word Confusion, that reframes semantic similarity in terms of feature-based classification confusion. Word Confusion is inspired by Tversky (1977)’s suggestion that similarity features be chosen dynamically. Here we train a classifier to map contextual embeddings to word identities and use the classifier confusion (the probability of choosing a confounding word c instead of the correct target word t) as a measure of the similarity of c and t. The set of potential confounding words acts as the chosen features. Our method is comparable to cosine similarity in matching human similarity judgments across several datasets (MEN, WirdSim353, and SimLex), and can measure similarity using predetermined features of interest. We demonstrate our model’s ability to make use of dynamic features by applying it to test a hypothesis about changes in the 18th C. meaning of the French word “révolution” from popular to state action during the French Revolution. We hope this reimagining of semantic similarity will inspire the development of new tools that better capture the multi-faceted and dynamic nature of language, advancing the fields of computational social science and cultural analytics and beyond.

Entity alignment (EA) is crucial for integrating multi-source knowledge graphs (KGs), aiming to identify equivalent entities across different graphs. However, most existing EA decoding methods rely on both entity and relation embeddings, limiting their generalizability and efficiency, especially in GNN-based models. To address these challenges, we propose Triple Feature Propagation (TFP), an adaptable and fast EA decoding framework that only utilizes entity embeddings. TFP reconstructs KG representation by maximizing the smoothness of entity embeddings. The discretized smoothness-maximization process yields the explicit Euler solution of TFP. We also generalize multi-view matrices: entity-to-entity, entity-to-relation, relation-to-entity, and relation-to-triple, to capture structural diversity. Extensive experiments on public datasets demonstrate that TFP is fast and adaptable to various encoders, achieving comparable results to state-of-the-art methods in under 6 seconds, and surpassing them in many cases.

Large language models (LLMs) have shown remarkable effectiveness across various domains, with data augmentation methods utilizing GPT for synthetic data generation becoming prevalent. However, the quality and utility of augmented data remain questionable, and current methods lack clear metrics for evaluating data characteristics. To address these challenges, we propose ResoFilter, a novel method that integrates models, data, and tasks to refine datasets. ResoFilter leverages the fine-tuning process to obtain Data-Parameter features for data selection, offering improved interpretability by representing data characteristics through model weights. Our experiments demonstrate that ResoFilter achieves comparable results to full-scale fine-tuning using only half the data in mathematical tasks and exhibits strong generalization across different models and domains. This method provides valuable insights for constructing synthetic datasets and evaluating high-quality data, offering a promising solution for enhancing data augmentation techniques and improving training dataset quality for LLMs. For reproducibility, we will release our code and data upon acceptance.

The pace of scientific research, vital for improving human life, is complex, slow, and needs specialized expertise. Meanwhile, novel, impactful research often stems from both a deep understanding of prior work, and a cross-pollination of ideas across domains and fields. To enhance the productivity of researchers, we propose ResearchAgent, which leverages the encyclopedic knowledge and linguistic reasoning capabilities of Large Language Models (LLMs) to assist them in their work. This system automatically defines novel problems, proposes methods and designs experiments, while iteratively refining them based on the feedback from collaborative LLM-powered reviewing agents. Specifically, starting with a core scientific paper, ResearchAgent is augmented not only with relevant publications by connecting information over an academic graph but also entities retrieved from a knowledge store derived from shared underlying concepts mined across numerous papers. Then, mimicking a scientific approach to improving ideas with peer discussions, we leverage multiple LLM-based ReviewingAgents that provide reviews and feedback via iterative revision processes. These reviewing agents are instantiated with human preference-aligned LLMs whose criteria for evaluation are elicited from actual human judgments via LLM prompting. We experimentally validate our ResearchAgent on scientific publications across multiple disciplines, showing its effectiveness in generating novel, clear, and valid ideas based on both human and model-based evaluation results. Our initial foray into AI-mediated scientific research has important implications for the development of future systems aimed at supporting researchers in their ideation and operationalization of novel work.

What kind of internal mechanisms might Transformers use to conduct fluid, natural-sounding conversations? Prior work has illustrated by construction how Transformers can solve various synthetic tasks, such as sorting a list or recognizing formal languages, but it remains unclear how to extend this approach to a conversational setting. In this work, we propose using ELIZA, a classic rule-based chatbot, as a setting for formal, mechanistic analysis of Transformer-based chatbots. ELIZA allows us to formally model key aspects of conversation, including local pattern matching and long-term dialogue state tracking. We first present a theoretical construction of a Transformer that implements the ELIZA chatbot. Building on prior constructions, particularly those for simulating finite-state automata, we show how simpler mechanisms can be composed and extended to produce more sophisticated behavior. Next, we conduct a set of empirical analyses of Transformers trained on synthetically generated ELIZA conversations. Our analysis illustrates the kinds of mechanisms these models tend to prefer—for example, models favor an induction head mechanism over a more precise, position-based copying mechanism; and using intermediate generations to simulate recurrent data structures, akin to an implicit scratchpad or Chain-of-Thought.Overall, by drawing an explicit connection between neural chatbots and interpretable, symbolic mechanisms, our results provide a new framework for the mechanistic analysis of conversational agents.

Despite active research on Automated Essay Scoring (AES), there is a noticeable scarcity of studies focusing on predicting creativity scores for essays. In this study, we develop a new essay rubric specifically designed for assessing creativity in essays. Leveraging this rubric, we construct ground truth data consisting of 5,048 essays. Furthermore, we propose a novel self-supervised learning model that recognizes cluster patterns within the essay embedding space and leverages them for creativity scoring. This approach aims to automatically generate a high-quality training set, thereby facilitating the training of diverse language models. Our experimental findings indicated a substantial enhancement in the assessment of essay creativity, demonstrating an increase in F1-score up to 58% compared to the primary state-of-the-art models across the ASAP and AIHUB datasets.

This paper introduces repetition neurons, which can be regarded as “skill neurons” responsible for the repetition problem in text generation tasks. These neurons are progressively activated more strongly as repetition continues, indicating that they perceive repetition as a task to copy the previous context repeatedly, similar to in-context learning. We identify these repetition neurons by comparing activation values before and after the onset of repetition in texts generated by recent pre-trained language models. We analyze the repetition neurons in three English and one Japanese pre-trained language models and observe similar patterns across them.

Topic models allow researchers to extract latent factors from text data and use those variables in downstream statistical analyses. However, these methodologies can vary significantly due to initialization differences, randomness in sampling procedures, or noisy data. Reliability of these methods is of particular concern as many researchers treat learned topic models as ground truth for subsequent analyses. In this work, we show that the standard practice for quantifying topic model reliability fails to capture essential aspects of the variation in two widely-used topic models. Drawing from a extensive literature on measurement theory, we provide empirical and theoretical analyses of three other metrics for evaluating the reliability of topic models. On synthetic and real-world data, we show that McDonald’s 𝜔 provides the best encapsulation of reliability. This metric provides an essential tool for validation of topic model methodologies that should be a standard component of any topic model-based research.

Preference alignment via reward models helps build safe, helpful, and reliable large language models (LLMs). However, subjectivity in preference judgments and the lack of representative sampling in preference data collection can introduce new biases, hindering reward models’ fairness and equity. In this work, we introduce a framework for evaluating dialect biases in reward models and conduct a case study on biases against African American Language (AAL) through several experiments comparing reward model preferences and behavior on paired White Mainstream English (WME) and both machine-translated and human-written AAL corpora. We show that reward models are less aligned with human preferences when processing AAL texts vs. WME ones (-4% accuracy on average), frequently disprefer AAL-aligned texts vs. WME-aligned ones, and steer conversations toward WME, even when prompted with AAL texts. Our findings provide a targeted analysis of anti-AAL biases at a relatively understudied stage in LLM development, highlighting representational harms and ethical questions about the desired behavior of LLMs concerning AAL.

Reinforcement learning (RL) is a promising approach for aligning large language models (LLMs) knowledge with sequential decision-making tasks. However, few studies have thoroughly investigated the impact on LLM agents capabilities of fine-tuning them with RL in a specific environment. In this paper, we propose a novel framework to analyze the sensitivity of LLMs to prompt formulations following RL training in a textual environment. Our findings reveal that the performance of LLMs degrades when faced with prompt formulations different from those used during the RL training phase. Besides, we analyze the source of this sensitivity by examining the model’s internal representations and salient tokens. Finally, we propose to use a contrastive loss to mitigate this sensitivity and improve the robustness and generalization capabilities of LLMs.

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking when the accuracy of the reward model is not high enough. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. A common solution to prevent reward hacking in preference learning techniques is to optimize a reward using proximity regularization, which ensures that the language model remains close to the reference model. In this research, we propose MBR-BoN, a variant of BoN that aims to mitigate reward hacking at inference time by incorporating the Minimum Bayes Risk (MBR) objective as a proximity regularization term. We show empirically and analytically that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer for BoN sampling. We evaluate MBR-BoN on the AlpacaFarm and Anthropic’s hh-rlhf datasets and show that it outperforms both BoN sampling and MBR decoding. As an application of MBR-BoN, we use it to generate a pairwise preference learning dataset. Experimental results show that DPO models trained on a dataset generated with MBR-BoN outperform a DPO model generated with vanilla BoN.