AACL 2025 Papers

Unstructured Knowledge Editing (UKE) is crucial for updating the relevant knowledge of large language models (LLMs). It focuses on unstructured inputs, such as long or free-form texts, which are common forms of real-world knowledge. Although previous studies have proposed effective methods and tested them, some issues exist: (1) Lack of Locality evaluation for UKE, and (2) Abnormal failure of fine-tuning (FT) based methods for UKE.To address these issues, we first construct two datasets, UnKEBench-Loc and AKEW-Loc (CF), by extending two existing UKE datasets with locality test data from the unstructured and structured views. This enables a systematic evaluation of the Locality of post-edited models. Furthermore, we identify four factors that may affect the performance of FT-based methods. Based on these factors, we conduct experiments to determine how the well-performing FT-based methods should be trained for the UKE task, providing a training recipe for future research. Our experimental results indicate that the FT-based method with the optimal setting (FT-UKE) is surprisingly strong, outperforming the existing state-of-the-art (SOTA).In batch editing scenarios, FT-UKE shows strong performance as well, with its advantage over SOTA methods increasing as the batch size grows, expanding the average metric lead from +6.78% to +10.80%. Our code and data will be released on Github.

In recent years, there has been growing interest in agents that collaborate with humans on creative tasks, and research has begun to explore such collaboration within Minecraft. However, most existing studies on agents in Minecraft focus on scenarios where an agent constructs objects independently on the basis of given instructions, making it difficult to achieve joint construction through dialogue-based cooperation with humans. Prior work, such as the Action-Utterance Model, used small-scale large language models (LLMs), which resulted in limited accuracy. In this study, we attempt to build an agent capable of collaborative construction using LLMs by integrating the framework of the Action-Utterance Model with that of Creative Agents, which leverages more recent and powerful LLMs for more accurate and flexible building. We had two agents conduct the Collaborative Garden Task through simulations and evaluate both the generated gardens and the dialogue content. Through this evaluation, we confirm that the agents are capable of producing gardens with a certain level of quality and can actively offer suggestions and assert their opinions. Furthermore, we conduct a comparative analysis with human-human collaboration to identify current challenges faced by agents and to discuss future directions for improvement toward achieving more human-like cooperative behavior.

We present JPharmatron, a Japanese domain-specific large language model (LLM) for the pharmaceutical field, developed through continual pre-training on two billion Japanese pharmaceutical tokens and eight billion English biomedical tokens. For rigorous evaluation, we introduce JPharmaBench, a benchmark suite consisting of three new benchmarks: YakugakuQA, based on national pharmacist licensing exams; NayoseQA, which tests cross-lingual synonym and terminology normalization; and SogoCheck, a novel task involving cross-document consistency checking.We evaluate our model against open-source medical LLMs and commercial models, including GPT-4o. Experimental results show that JPharmatron outperforms existing open models and achieves competitive performance with commercial ones.Interestingly, even GPT-4o performs poorly on SogoCheck, suggesting that cross-sentence consistency reasoning remains an open challenge.JPharmatron enables secure and local model deployment for pharmaceutical tasks, where privacy and legal constraints limit the use of closed models. Besides, JPharmaBench offers a reproducible framework for evaluating Japanese pharmaceutical natural language processing. Together, they demonstrate the feasibility of practical and cost-efficient language models for Japanese healthcare and pharmaceutical sectors.Our model, codes, and datasets are available on HuggingFace: https://huggingface.co/collections/EQUES/jpharmatron and https://huggingface.co/collections/EQUES/jpharmabench.

In the era of rapid Internet and social media development, individuals readily share their investment opinions online. The overwhelming volume of such opinions makes comprehensive evaluation impractical, highlighting the need for an effective recommendation system that can identify valuable insights. To address this challenge, we propose an argument-based sentiment analysis framework that incorporates a new perspective on opinion strength. Our approach introduces the concept of a Fuzzy Strength Degree (FSD), derived from the difference between analysts’ target and closing prices, to quantify the intensity of opinions. By integrating argument mining techniques, we further decompose each opinion into claims and premises, examine their relationships, and use these structures to evaluate the persuasive strength of the arguments. This dual strategy allows us to rank both professional and amateur investor opinions without relying on user history or social signals. Experiments show that our method works best for analyst reports, while on social media, simpler approaches based on wording and professionalism features perform better. Moreover, our analysis of professional analysts’ and traders’ behaviors reveals that top-ranked opinions are more likely to influence subsequent market actions. These findings demonstrate that argument structure and quantified opinion strength provide a novel and reliable foundation for investment opinion recommendation.

Large Language Models (LLMs) are increasingly used to support scientific writing, but their tendency to produce hallucinated content threatens academic reliability. Existing benchmarks have addressed hallucination detection in general-domain tasks, such as fact-checking or question answering, but they do not reflect the fine-grained, domain-specific needs of scientific communication. We introduce SciHallu, a dataset for identifying hallucinations in academic text at three levels of granularity: token, sentence, and paragraph. To establish a reliable ground truth, we select source passages from research papers published prior to the widespread adoption of LLMs. Our dataset includes both hallucinated and non-hallucinated paragraph instances, constructed through controlled perturbations at varying levels of noise and validated by human annotators. A rationale is paired with each instance, explaining the nature of the modification. SciHallu covers multiple academic fields, such as Computer Science, Health Sciences, and Humanities and Social Sciences. It is built using a model-guided annotation pipeline, followed by expert human validation. We evaluate state-of-the-art LLMs on both binary and fine-grained classification tasks, revealing challenges in detecting subtle hallucinations. SciHallu supports the development of context-aware systems for more trustworthy scientific content generation.

Thomas C. Schelling, awarded the 2005 Nobel Memorial Prize in Economic Sciences, pointed out that ”individuals decisions (micromotives), while often personal and localized, can lead to societal outcomes (macrobehavior) that are far more complex and different from what the individuals intended.” The current research related to large language models’ (LLMs’) micromotives, such as preferences or biases, assumes that users will make more appropriate decisions once LLMs are devoid of preferences or biases. Consequently, a series of studies has focused on removing bias from LLMs. In the NLP community, while there are many discussions on LLMs’ micromotives, previous studies have seldom conducted a systematic examination of how LLMs may influence society’s macrobehavior. In this paper, we follow the design of Schelling’s model of segregation to observe the relationship between the micromotives and macrobehavior of LLMs. Our results indicate that, regardless of the level of bias in LLMs, a highly segregated society will emerge as more people follow LLMs’ suggestions. We hope our discussion will spark further consideration of the fundamental assumption regarding the mitigation of LLMs’ micromotives and encourage a reevaluation of how LLMs may influence users and society.

In the deployment of Large Language Models (LLMs), “spurious correctness”—where answers are correct but reasoning contains errors—poses a critical risk by creating an illusion of reliability. While prior work on LLM confidence estimation focuses on answer-level or entire reasoning path confidence, these coarse-grained approaches fail to identify which specific parts of the reasoning contain errors. We propose a fine-grained confidence estimation framework that computes confidence scores for individual evidence triplets within reasoning chains, enabling precise localization of errors. Using carefully designed prompts, we generate answers, evidence in triplet format, and their respective confidence scores simultaneously, allowing automatic detection of spurious correctness patterns where partial evidence contains factual errors. Evaluated on both Japanese and English multi-hop QA benchmarks across multiple models from three model families representing different architectures and training approaches, we show that our approach exhibits superior calibration performance for evidence confidence and demonstrates effective ability to detect spurious correct answers (up to 0.84 on our primary discrimination metric). The consistent improvements across languages demonstrate the generalizability of our method. As a secondary benefit, joint generation of confidence scores improves answer confidence calibration by up to 43%. This prompt-based approach requires no model retraining and is immediately applicable to existing LLMs.

Language is, as commonly theorized, largely arbitrary. Yet, systematic relationships between phonetics and semantics have been observed in many specific cases. To what degree could those systematic relationships manifest themselves in large scale, quantitative investigations–both in previously identified and unidentified phenomena? This work undertakes a distributional approach to quantifying phonosemantic iconicity at scale across 6 diverse languages (English, Spanish, Hindi, Finnish, Turkish, and Tamil). In each language, we analyze the alignment of morphemes’ phonetic and semantic similarity spaces with a suite of statistical measures, and discover an array of interpretable phonosemantic alignments not previously identified in the literature, along with crosslinguistic patterns. We also analyze 5 previously hypothesized phonosemantic alignments, finding support for some such alignments and mixed results for others.

Sign language understanding remains a significant challenge, particularly for low-resource sign languages with limited annotated data. Motivated by the success of large-scale pretraining in deep learning, we propose Multi-Stream Masked Autoencoder (MS-MAE) — a simple yet effective framework for learning sign language representations from skeleton-based video data. We pretrained a model with MS-MAE on the YouTube-ASL dataset, and then adapted it to multiple downstream tasks across different sign languages. Experimental results show that MS-MAE achieves competitive or superior performance on a range of isolated sign language recognition benchmarks and gloss-free sign language translation tasks across several sign languages. These findings highlight the potential of leveraging large-scale, high-resource sign language data to boost performance in low-resource sign language scenarios. Additionally, visualization of the model’s attention maps reveals its ability to cluster adjacent pose sequences within a sentence, some of which align with individual signs, offering insights into the mechanisms underlying successful transfer learning.

The web-scale of pretraining data has created an important evaluation challenge: to disentangle linguistic competence on cases well-represented in pretraining data from generalization to out-of-domain language, specifically the dynamic, real-world instances less common in pretraining data. To this end, we construct a diagnostic evaluation to systematically assess natural language understanding in LLMs by leveraging Construction Grammar (CxG). CxG provides a psycholinguistically grounded framework for testing generalization, as it explicitly links syntactic forms to abstract, non-lexical meanings. Our novel inference evaluation dataset consists of English phrasal constructions, for which speakers are known to be able to abstract over commonplace instantiations in order to understand and produce creative instantiations. Our evaluation dataset uses CxG to evaluate two central questions: first, if models can “understand” the semantics of sentences for instances that are likely to appear in pretraining data less often, but are intuitive and easy for people to understand. Second, if LLMs can deploy the appropriate constructional semantics given constructions that are syntactically identical but with divergent meanings. Our results demonstrate that state-of-the-art models, including GPT-o1, exhibit a performance drop of over 40% on our second task, revealing a failure to generalize over syntactically identical forms to arrive at distinct constructional meanings in the way humans do. We make our novel dataset and associated experimental data, including prompts and model responses, publicly available.

Vision-language Models (VLMs) have made significant strides in visual understanding and query response generation, but often face challenges of high computational cost and inference latency due to autoregressive decoding. In this work, we introduce an imitation-learning-based Self-Speculative Decoding (SSD) framework, named FastVLM, to address these limitations. Our approach employs a lightweight draft model for token generation in an autoregressive manner, while a full model verifies these tokens non-autoregressively. Accepted tokens proceed seamlessly, while rejected tokens are corrected by the full model and used to guide the draft model’s refinement. Through an imitation network, FastVLM enhances the draft model by integrating deeper-level insights from the full model’s architecture. Also, it maintains the performance integrity of the full model while training the draft model, achieving a balance between efficiency and accuracy. Our method speeds up the inference process by 1.55-1.85\times as compared to the final layer with minimal loss in performance. The source code is available at https://github.com/Div290/SSD.

Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.

Audio Large Language Models (AudioLLMs) have achieved strong results in semantic tasks like speech recognition and translation, but remain limited in modeling paralinguistic cues such as emotion. Existing approaches often treat emotion understanding as a classification problem, offering little insight into the underlying rationale behind predictions. In this work, we explore emotion reasoning, a strategy that leverages the generative capabilities of AudioLLMs to enhance emotion recognition by producing semantically aligned, evidence-grounded explanations. To support this in multitask AudioLLMs, we introduce a unified framework combining reasoning-augmented data supervision, dual-encoder architecture, and task-alternating training. This approach enables AudioLLMs to effectively learn different tasks while incorporating emotional reasoning. Experiments on IEMOCAP and MELD show that our approach not only improves emotion prediction accuracy but also enhances the coherence and evidential grounding of the generated responses. Experiments on two out-of-domain datasets demonstrate the generalization capabilities of the resulting model.

This paper introduces EmplifAI, a Japanese empathetic dialogue dataset designed to support patients coping with chronic medical conditions. They often experience a wide range of positive and negative emotions (e.g., hope and despair) that shift across different stages of disease management. EmplifAI addresses this complexity by providing situation-based dialogues grounded in 28 fine-grained emotion categories, adapted and validated from the GoEmotions taxonomy. The dataset includes 280 medically contextualized situations and 4,125 two-turn dialogues, collected through crowdsourcing and expert review.To evaluate emotional alignment in empathetic dialogues, we assessed model predictions on situation–dialogue pairs using BERTScore across multiple large language models (LLMs), achieving F1 scores of ≤ 0.83. Fine-tuning a baseline Japanese LLM (LLM-jp-3.1-13b-instruct4) with EmplifAI resulted in notable improvements in fluency, general empathy, and emotion-specific empathy. Furthermore, we compared the scores assigned by LLM-as-a-Judge and human raters on dialogues generated by multiple LLMs to validate our evaluation pipeline and discuss the insights and potential risks derived from the correlation analysis.

Although behavioral studies have documented numerical reasoning errors in large language models (LLMs), the underlying representational mechanisms remain unclear. We hypothesize that numerical attributes occupy shared latent subspaces and investigate two questions: (1) How do LLMs internally integrate multiple numerical attributes of a single entity? (2) How does irrelevant numerical context perturb these representations and their downstream outputs? To address these questions, we combine linear probing with partial correlation analysis and prompt-based vulnerability tests across models of varying sizes. Our results show that LLMs encode real-world numerical correlations but tend to systematically amplify them. Moreover, irrelevant context induces consistent shifts in magnitude representations, with downstream effects that vary by model size. These findings reveal a vulnerability in LLM decision-making and lay the groundwork for fairer, representation-aware control under multi-attribute entanglement.

Chemical molecules can be represented as graphs or as language descriptions. Training unimodal models on graphs results in different encodings than training them on language. Therefore, the existing literature force-aligns the unimodal models during training to use them in downstream applications such as drug discovery. But to what extent are graph and language unimodal model representations inherently aligned, i.e., aligned prior to any force-alignment training? Knowing this is useful for a more expedient and effective forced-alignment. For the first time, we explore methods to gauge the alignment of graph and language unimodal models. We find compelling differences between models and their ability to represent slight structural differences without force-alignment. We also present an unified unimodal alignment (U2A) benchmark for gauging the inherent alignment between graph and language encoders which we make available with this paper.

Journalistic interviews are creative, dynamic processes where success hinges on insightful, real-time questioning. While Large Language Models (LLMs) can assist, their tendency to generate coherent but uninspired questions optimizes for probable, not insightful, continuations. This paper investigates whether a structured, multi-agent approach can overcome this limitation to act as a more effective creative partner for journalists. We introduce MAJI, a system designed for this purpose, which employs a divergent-convergent architecture: a committee of specialized agents generates a diverse set of questions, and a convergent agent selects the optimal one. We evaluated MAJI against a suite of strong LLM baselines. Our results demonstrate that our multi-agent framework produces questions that are more coherent, elaborate, and original (+36.9% for our best model vs. a standard LLM baseline), exceeded strong LLM baselines on key measures of creative question quality. Most critically, in a blind survey, professional journalists preferred MAJI’s selected questions over those from the baseline by a margin of more than two to one. We present the system’s evolution, highlighting the architectural trade-offs that enable MAJI to augment, rather than simply automate, journalistic inquiry. We will release the code upon publication.

Large language models (LLMs) often benefit from verbalized reasoning at inference time, but it remains unclear which aspects of task difficulty these extra reasoning tokens address. To investigate this question, we construct a controlled setting where task complexity can be precisely manipulated to study its effect on reasoning length. Deterministic finite automata (DFAs) offer a formalism through which we can characterize task complexity through measurable properties such as run length (number of reasoning steps required) and state-space size (decision complexity). We first show that across different tasks and models of different sizes and training paradigms, there exists an optimal amount of reasoning tokens such that the probability of producing a correct solution is maximized. We then investigate which properties of complexity govern this critical length: we find that task instances with longer corresponding underlying DFA runs (i.e. demand greater latent state-tracking requirements) correlate with longer reasoning lengths, but, surprisingly, that DFA size (i.e. state-space complexity) does not. We then demonstrate an implication of these findings: being able to predict the optimal number of reasoning tokens for new problems and filtering out non-optimal length answers results in consistent accuracy improvements.

Retrieval-Augmented Generation (RAG) enhances LLMs by grounding answers in retrieved passages, which is key in factual Question Answering. However, generated answers may still be unfaithful to the passages, either due to retrieval or generation errors. Many RAG downstream applications rely on assessing answer faithfulness for applying fallback strategies, yet address it implicitly, without a consistent evaluation methodology. We introduce the task of Answering with Faithfulness (AwF), which brings faithfulness prediction to the forefront, explicitly coupling it with answer generation. We define variants of the precision and recall metrics tailored to this task, facilitating direct evaluation and comparison of different AwF methods.We then demonstrate, both theoretically and empirically, that for RAG applications using AwF as a sub-procedure, an improvement to the AwF metrics translates to an improvement to the downstream performance. This results in improved performance for recently published results.