NAACL 2025 Papers

Stance detection is critical for understanding the underlying position or attitude expressed toward a topic. Large language models (LLMs) have demonstrated significant advancements across various natural language processing tasks including stance detection, however, their performance in stance detection is limited by biases and spurious correlations inherent due to their data-driven nature. Our statistical experiment reveals that LLMs are prone to generate biased stances due to sentiment-stance spurious correlations and preference towards certain individuals and topics. Furthermore, the results demonstrate a strong negative correlation between stance bias and stance detection performance, underscoring the importance of mitigating bias to enhance the utility of LLMs in stance detection. Therefore, in this paper, we propose a Counterfactual Augmented Calibration Network (FACTUAL), which a novel calibration network is devised to calibrate potential bias in the stance prediction of LLMs. Further, to address the challenge of effectively learning bias representations and the difficulty in the generalizability of debiasing, we construct counterfactual augmented data. This approach enhances the calibration network, facilitating the debiasing and out-of-domain generalization. Experimental results on in-target and zero-shot stance detection tasks show that the proposed FACTUAL can effectively mitigate biases of LLMs, achieving state-of-the-art results.

In education, high-quality exams must cover broad specifications across diverse difficulty levels during the assembly and calibration of test items to effectively measure examinees’ competence. However, balancing the trade-off of selecting relevant test items while fulfilling exam specifications without bias is challenging, particularly when manual item selection and exam assembly rely on a pre-validated item base. To address this limitation, we propose a new mixed-integer programming re-ranking approach to improve relevance, while mitigating bias on an industry-grade exam assembly platform. We evaluate our approach by comparing it against nine bias mitigation re-ranking methods in 225 experiments on a real-world benchmark data set from vocational education services. Experimental results demonstrate a 17% relevance improvement with a 9% bias reduction when integrating sequential optimization techniques with improved contextual relevance augmentation and scoring using a large language model. Our approach bridges information retrieval and exam assembly, enhancing the human-in-the-loop exam assembly process while promoting unbiased exam design

Efficient finetuning of large language models (LLMs) aims to adapt the LLMs with reduced computational and memory costs. Previous LoRA-based approaches initialize the low-rank matrices with Gaussian distribution and zero values while keeping the original weight matrices frozen. However, the trainable model parameters optimized in an unguided subspace might interfere with the well-learned subspace of the pretrained weight matrices. In this paper, we propose MiLoRA, a simple yet effective LLM finetuning approach that only updates the minor singular components of the weight matrix while keeping the principal singular components frozen. It is observed that the minor matrix corresponds to the noisy or long-tail information, while the principal matrix contains important knowledge. The MiLoRA initializes the low-rank matrices within a subspace that is orthogonal to the principal matrix, thus the pretrained knowledge is expected to be well preserved. During finetuning, MiLoRA makes the most use of the less-optimized subspace for learning the labeled dataset. Extensive experiments on commonsense reasoning, math reasoning, instruction following and visual instruction following benchmarks present the superior performance of our method.

**Multimodal Chain of Thought (MCoT)** is a popular prompting strategy for improving the performance of multimodal large language models (MLLMs) across a range of complex reasoning tasks. Despite its popularity, there is a notable absence of automated methods for evaluating the quality of reasoning steps in MCoT. To address this gap, we propose **Multimodal Chain-of-Thought Evaluation (MiCEval)**, a framework designed to assess the correctness of reasoning chains by evaluating the quality of both the description and each reasoning step. The evaluation of the description component focuses on the accuracy of the image descriptions, while the reasoning step evaluates the quality of each step as it is conditionally generated based on the preceding steps. MiCEval is built upon a fine-grained dataset with annotations that rate each step according to correctness, relevance, and informativeness. Extensive experiments on four state-of-the-art MLLMs show that step-wise evaluations using MiCEval align more closely with human judgments compared to existing methods based on cosine similarity or fine-tuning approaches. MiCEval datasets and code can be found at: [https://anonymous_github/MicEval](https://anonymous.4open.science/r/MiCEval-847F/README.md).

The discovery of novel mechanical metamaterials, whose properties are dominated by their engineered structures rather than chemical composition, is a knowledge-intensive and resource-demanding process. To accelerate the design of novel metamaterials, we present MetaScientist, a human-in-the-loop system that integrates advanced AI capabilities with expert oversight with two primary phases: (1) hypothesis generation, where the system performs complex reasoning to generate novel and scientifically sound hypotheses, supported with domain-specific foundation models and inductive biases retrieved from existing literature; (2) 3D structure synthesis, where a 3D structure is synthesized with a novel 3D diffusion model based on the textual hypothesis and refined it with a LLM-based refinement model to achieve better structure properties. At each phase, domain experts iteratively validate the system outputs, and provide feedback and supplementary materials to ensure the alignment of the outputs with scientific principles and human preferences. Through extensive evaluation from human scientists, MetaScientist is able to deliver novel and valid mechanical metamaterial designs that have the potential to be highly impactful in the metamaterial field.

Large Language Models (LLMs) acquire extensive knowledge and remarkable abilities from extensive text corpora, making them powerful tools for various applications. To make LLMs more usable, aligning them with human preferences is essential. Existing alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), typically embed predefined preferences directly within the model’s parameters. These methods, however, often result in a static alignment that can not account for the diversity of human preferences in practical applications.In response to this challenge, we propose an effective method, MetaAlign, which aims to help LLMs dynamically align with various explicit or implicit preferences specified at inference time. Experimental results show that LLMs optimized on our meticulously constructed MetaAlign Dataset can effectively align with any preferences specified at the inference stage, validating the feasibility of MetaAlign. We hope that our work can provide some insights into the alignment of language models.

External tools help large language models succeed at tasks where they would otherwise typically fail. In existing frameworks, choosing tools at test time relies on naive greedy decoding, regardless of whether the model has been fine-tuned on tool-annotated data or prompted with in-context examples. In contrast, we find that gathering and choosing among a suitable set of candidate tools has greater potential to lead to an optimal selection. We present Tool selECTion via meta-reasONing (TECTON), a two-phase system that first *reasons* over a task and outputs candidate tools using a custom fine-tuned language modelling head. Then, with the custom head disabled, it *meta-reasons* (i.e., it reasons over the previous reasoning process) to make a final choice. We show that TECTON results in substantial gains—both in-distribution and out-of-distribution—on a range of math reasoning datasets.

Numerous recent studies have shown that Large Language Models (LLMs) are biased towards a Western and Anglo-centric worldview, which compromises their usefulness in non-Western cultural settings. However, “culture” is a complex, multifaceted topic, and its awareness, representation, and modeling in LLMs and LLM-based applications can be defined and measured in numerous ways. In this position paper, we ask what does it mean for an LLM to possess “cultural awareness”, and through a thought experiment, which is an extension of the Octopus test proposed by Bender and Koller (2020), we argue that it is not cultural awareness or knowledge, rather meta-cultural competence, which is required of an LLM and LLM-based AI system that will make it useful across various, including completely unseen, cultures. We lay out the principles of meta-cultural competence AI systems, and discuss ways to measure and model those.

The recent success of specialized Large Language Models (LLMs) in domains such as mathematical reasoning and coding has led to growing interest in methods for merging these expert LLMs into a unified Mixture-of-Experts (MoE) model, with the goal of enhancing performance in each domain while retaining effectiveness on general tasks. However, effective merging of expert models remains an open challenge, especially for models with highly divergent weight parameters or different architectures. State-of-the-art MoE merging methods only work with homogeneous model architectures and rely on simple unweighted averaging to merge expert layers, which does not address parameter interference and requires extensive fine-tuning of the merged MoE to restore performance. To address these limitations, this paper introduces new MoE merging techniques, including strategies to mitigate parameter interference, routing heuristics to reduce the need for MoE fine-tuning, and a novel method for merging experts with different architectures. Extensive experiments across multiple domains demonstrate the effectiveness of our proposed methods, reducing fine-tuning costs, improving performance over state-of-the-art methods, and expanding the applicability of MoE merging.

Spoken Named Entity Recognition (NER) aims to extract named entities from speech and categorise them into types like person, location, organization, etc. In this work, we present *VietMed-NER* - the first spoken NER dataset in the medical domain. To our knowledge, our Vietnamese real-world dataset is the largest spoken NER dataset in the world regarding the number of entity types, featuring 18 distinct types. Furthermore, we present baseline results using various state-of-the-art pre-trained models: encoder-only and sequence-to-sequence; and conduct quantitative and qualitative error analysis. We found that pre-trained multilingual models generally outperform monolingual models on reference text and ASR output and encoders outperform sequence-to-sequence models in NER tasks. By translating the transcripts, the dataset can also be utilised for text NER in the medical domain in other languages than Vietnamese. All code, data and models are publicly available.

In the human activity of word translation, two languages face each other, mutually searching their own language system for the semantic place of words in the other language. We discover the huge network formed by the chain of these mutual translations as *Word Translation Network*, a network where words are nodes, and translation volume is represented as edges, and propose *Word Translation Map*, a novel interface for exploring this network. *Word Translation Map* points to the semantic configurations of many words in multiple languages at once, containing the information of existing dictionaries such as bilingual and synonym dictionaries. We have also implemented and published this interface as a web application, focusing on seven language pairs. This paper first defines the *Word Translation Network* and describes how to actually construct the network from bilingual corpora, followed by an analysis of the properties of the network. Next, we explain how to design a *Word Translation Map* using this network, and finally, we analyze the features of the *Word Translation Map* as a dictionary. The web application is publicly accessible at www.media-of-langue.org.

Medical Visual Question Answering (Med-VQA), which offers language responses to image-based medical inquiries, represents a challenging task and significant advancement in healthcare. It assists medical experts to swiftly interpret medical images, thereby enabling faster and more accurate diagnoses. However, the model interpretability and transparency of existing Med-VQA solutions are often limited, posing challenges in understanding their decision-making processes. To address this issue, we devise a semi-automated annotation process to streamline data preparation and build new benchmark Med-VQA datasets R-RAD, R-SLAKE and R-Path. These datasets provide intermediate medical decision-making rationales generated by multimodal large language models and human annotations for question-answering pairs in existing Med-VQA datasets, i.e., VQA-RAD, SLAKE and PathVQA. Moreover, we design a novel framework, MedThink, which finetunes lightweight pretrained generative models by incorporating medical decision-making rationales. MedThink includes three distinct strategies to generate decision outcomes and corresponding rationales, clearly showcasing the medical decision-making process during reasoning. Our comprehensive experiments show that our method achieves an accuracy of 83.5% on R-RAD, 86.3% on R-SLAKE and 87.2% on R-Path. These results significantly exceed those of existing state-of-the-art models with comparable parameters. Datasets and code are available at https://github.com/Tang-xiaoxiao/Medthink.

Numerous advanced Large Language Models (LLMs) now support context lengths up to 128K, and some extend to 200K. Some benchmarks in the generic domain have also followed up on evaluating long-context capabilities. In the medical domain, tasks are distinctive due to the unique contexts and need for domain expertise, necessitating further evaluation. However, despite the frequent presence of long texts in medical scenarios, evaluation benchmarks of long-context capabilities for LLMs in this field are still rare. In this paper, we propose MedOdyssey, the first medical long-context benchmark with seven length levels ranging from 4K to 200K tokens. MedOdyssey consists of two primary components: the medical-context “needles in a haystack” task and a series of tasks specific to medical applications, together comprising 10 datasets. The first component includes challenges such as counter-intuitive reasoning and novel (unknown) facts injection to mitigate knowledge leakage and data contamination of LLMs. The second component confronts the challenge of requiring professional medical expertise. Especially, we design the ‘“Maximum Identical Context” principle to improve fairness by guaranteeing that different LLMs observe as many identical contexts as possible. Our experiment evaluates advanced proprietary and open-source LLMs tailored for processing long contexts and presents detailed performance analyses. This highlights that LLMs still face challenges and need for further research in this area. Our code and data are released in the repository: https://github.com/JOHNNY-fans/MedOdyssey.

Embedding-based retrieval (EBR), the mainstream approach in information retrieval (IR), aims to help users obtain relevant information and plays a crucial role in retrieval-augmented generation (RAG) techniques of large language models (LLMs). Numerous methods have been proposed to significantly improve the quality of retrieved content and many generic benchmarks are proposed to evaluate the retrieval abilities of embedding models. However, texts in the medical domain present unique contexts, structures, and language patterns, such as terminology, doctor-patient dialogue, and electronic health records (EHRs). Despite these unique features, specific benchmarks for medical context retrieval are still lacking. In this paper, we propose MedEureka, an enriched benchmark designed to evaluate medical-context retrieval capabilities of embedding models with multi-granularity and multi-data types. MedEureka includes four levels of granularity and six types of medical texts, encompassing 18 datasets, incorporating granularity and data type description to prompt instruction-fine-tuned text embedding models for embedding generation. We also provide the MedEureka Toolkit to support evaluation on the MedEureka test set. Our experiments evaluate state-of-the-art open-source and proprietary embedding models, and fine-tuned classical baselines, providing a detailed performance analysis. This underscores the challenges of using embedding models for medical domain retrieval and the need for further research. Our code and data are released in the repository: https://github.com/JOHNNY-fans/MedEureka.

Large language models (LLMs) demonstrate significant potential in advancing medical applications, yet their capabilities in addressing medical ethics challenges remain underexplored. This paper introduces MedEthicEval, a novel benchmark designed to systematically evaluate LLMs in the domain of medical ethics. Our framework encompasses two key components: knowledge, assessing the models’ grasp of medical ethics principles, and application, focusing on their ability to apply these principles across diverse scenarios. To support this benchmark, we consulted with medical ethics researchers and developed three datasets addressing distinct ethical challenges: blatant violations of medical ethics, priority dilemmas with clear inclinations, and equilibrium dilemmas without obvious resolutions. MedEthicEval serves as a critical tool for understanding LLMs’ ethical reasoning in healthcare, paving the way for their responsible and effective use in medical contexts.

Medical coding standardizes clinical data but is both time-consuming and error-prone. Traditional Natural Language Processing (NLP) methods struggle with automating coding due to the large label space, lengthy text inputs, and the absence of supporting evidence annotations that justify code selection. Recent advancements in Generative Artificial Intelligence (AI) offer promising solutions to these challenges. In this work, we introduce MedCodER, an emerging Generative AI framework for automatic medical coding that leverages extraction, retrieval, and re-ranking techniques as core components. MedCodER achieves a micro-F1 score of 0.62 on International Classification of Diseases (ICD) code prediction, significantly outperforming state-of-the-art methods. Additionally, we present a new dataset containing medical records annotated with disease diagnoses, ICD codes, and supporting evidence texts (https://doi.org/10.5281/zenodo.13308316). Ablation tests confirm that MedCodER’s performance depends on the integration of each of its aforementioned components, as performance declines when these components are evaluated in isolation.

Transformer-based language models have achieved significant success; however, their internal mechanisms remain largely opaque due to the complexity of non-linear interactions and high-dimensional operations. While previous studies have demonstrated that these models implicitly embed reasoning trees, humans typically employ various distinct logical reasoning mechanisms to complete the same task. It is still unclear which multi-step reasoning mechanisms are used by language models to solve such tasks. In this paper, we aim to address this question by investigating the mechanistic interpretability of language models, particularly in the context of multi-step reasoning tasks. Specifically, we employ circuit analysis and self-influence functions to evaluate the changing importance of each token throughout the reasoning process, allowing us to map the reasoning paths adopted by the model. We apply this methodology to the GPT-2 model on a prediction task (IOI) and demonstrate that the underlying circuits reveal a human-interpretable reasoning process used by the model.

Large language models (LLMs) are susceptible to memorizing training data, raising concerns about the potential extraction of sensitive information at generation time. Discoverable extraction is the most common method for measuring this issue: split a training example into a prefix and suffix, then prompt the LLM with the prefix, and deem the example extractable if the LLM generates the matching suffix using greedy sampling. This definition yields a yes-or-no determination of whether extraction was successful with respect to a single query. Though efficient to compute, we show that this definition is unreliable because it does not account for non-determinism present in more realistic (non-greedy) sampling schemes, for which LLMs produce a range of outputs for the same prompt. We introduce probabilistic discoverable extraction, which, without additional cost, relaxes discoverable extraction by considering multiple queries to quantify the probability of extracting a target sequence. We evaluate our probabilistic measure across different models, sampling schemes, and training-data repetitions, and find that this measure provides more nuanced information about extraction risk compared to traditional discoverable extraction.

We are exposed to much information trying to influence us, such as teaser messages, debates, politically framed news, and propaganda — all of which use persuasive language. With the recent interest in Large Language Models (LLMs), we study the ability of LLMs to produce persuasive text. As opposed to prior work which focuses on particular domains or types of persuasion, we conduct a general study across various domains to measure and benchmark to what degree LLMs produce persuasive language - both when explicitly instructed to rewrite text to be more or less persuasive and when only instructed to paraphrase. We construct the new dataset Persuasive-Pairs of pairs of a short text and its rewrite by an LLM to amplify or diminish persuasive language. We multi-annotate the pairs on a relative scale for persuasive language: a valuable resource in itself, and for training a regression model to score and benchmark persuasive language, including for new LLMs across domains. In our analysis, we find that different ‘personas’ in LLaMA3’s system prompt change persuasive language substantially, even when only instructed to paraphrase.