NAACL 2025 论文列表

Decoding and expressing brain activity in a comprehensible form is a challenging frontier in AI. This paper presents *Thought2Text*, which uses instruction-tuned Large Language Models (LLMs) fine-tuned with EEG data to achieve this goal. The approach involves three stages: (1) training an EEG encoder for visual feature extraction, (2) fine-tuning LLMs on image and text data, enabling multimodal description generation, and (3) further fine-tuning on EEG embeddings to generate text directly from EEG during inference. Experiments on a public EEG dataset collected for six subjects with image stimuli and text captions demonstrate the efficacy of multimodal LLMs (*LLaMA-v3*, *Mistral-v0.3*, *Qwen2.5*), validated using traditional language generation evaluation metrics, as well as *fluency* and *adequacy* measures. This approach marks a significant advancement towards portable, low-cost “thoughts-to-text” technology with potential applications in both neuroscience and natural language processing.

In a systematic way, we investigate a widely asked question: Do LLMs really understand what they say?, which relates to the more familiar term Stochastic Parrot. To this end, we propose a summative assessment over a carefully designed physical concept understanding task, P HYSI C O. Our task alleviates the memorization issue via the usage of grid-format inputs that abstractly describe physical phenomena. The grids represents varying levels of understanding, from the core phenomenon, application examples to analogies to other abstract patterns in the grid world. A comprehensive study on our task demonstrates: (1) state-of-the-art LLMs, including GPT-4o, o1 and Gemini 2.0 flash thinking, lag behind humans by ∼40%; (2) the stochastic parrot phenomenon is present in LLMs, as they fail on our grid task but can describe and recognize the same concepts well in natural language; (3) our task challenges the LLMs due to intrinsic difficulties rather than the unfamiliar grid format, as in-context learning and fine-tuning on same formatted data added little to their performance.

Automatic summarization has consistently attracted attention due to its versatility and wide application in various downstream tasks. Despite its popularity, we find that annotation efforts have largely been disjointed, and have lacked common terminology. Consequently, it is challenging to discover existing resources or identify coherent research directions. To address this, we survey a large body of work spanning 133 datasets in over 100 languages, creating a novel ontology covering sample properties, collection methods and distribution. With this ontology we make key observations, including the lack of accessible high-quality datasets for low-resource languages, and the field’s overreliance on the news domain and on automatically collected distant supervision. Finally, we make available a web interface that allows users to interact and explore our ontology and dataset collection, as well as a template for a summarization data card, which can be used to streamline future research into a more coherent body of work.

Embedding models play a crucial role in Natural Language Processing (NLP) by creating text embeddings used in various tasks such as information retrieval and assessing semantic text similarity. This paper focuses on research related to embedding models in the Russian language. It introduces a new Russian-focused embedding model called ru-en-RoSBERTa and the ruMTEB benchmark, the Russian version extending the Massive Text Embedding Benchmark (MTEB). Our benchmark includes seven categories of tasks, such as semantic textual similarity, text classification, reranking, and retrieval.The research also assesses a representative set of Russian and multilingual models on the proposed benchmark. The findings indicate that the new model achieves results that are on par with state-of-the-art models in Russian. We release the model ru-en-RoSBERTa, and the ruMTEB framework comes with open-source code, integration into the original framework and a public leaderboard.

Spoken language understanding research to date has generally carried a heavy text perspective. Most datasets are derived from text, which is then subsequently synthesized into speech, and most models typically rely on automatic transcriptions of speech. This is to the detriment of prosody–additional information carried by the speech signal beyond the phonetics of the words themselves and difficult to recover from text alone. In this work, we investigate the role of prosody in Spoken Question Answering. By isolating prosodic and lexical information on the SLUE-SQA-5 dataset, which consists of natural speech, we demonstrate that models trained on prosodic information alone can perform reasonably well by utilizing prosodic cues. However, we find that when lexical information is available, models tend to predominantly rely on it. Our findings suggest that while prosodic cues provide valuable supplementary information, more effective integration methods are required to ensure prosody contributes more significantly alongside lexical features.

High annotation costs from hiring or crowdsourcing complicate the creation of large, high-quality datasets needed for training reliable text classifiers. Recent research suggests using Large Language Models (LLMs) to automate the annotation process, reducing these costs while maintaining data quality. LLMs have shown promising results in annotating downstream tasks like hate speech detection and political framing. Building on the success in these areas, this study investigates whether LLMs are viable for annotating a complex task of media bias detection and whether a downstream media bias classifier can be trained on such data. We create Annolexical, the first large-scale dataset for media bias classification with over 48k synthetically annotated examples. Our classifier fine-tuned on it surpasses all of the annotator LLMs by 5-9% in Mathew’s Correlation Coefficient (MCC) and performs close to or outperforms the model trained on human-labeled data when evaluated on two media bias benchmark datasets (BABE and BASIL). This study demonstrates how our approach significantly reduces the cost of dataset creation in the media bias domain and, by extension - the development of the classifiers, while our subsequent behavioral stress-testing reveals some of its current limitations and trade-offs.

Large Multimodal Models (LMMs) exhibit impressive performance across various multimodal tasks. However, their effectiveness in cross-cultural contexts remains limited due to the predominantly Western-centric nature of most data and models. Conversely, multi-agent models have shown significant capability in solving complex tasks. Our study evaluates the collective performance of LMMs in a multi-agent interaction setting for the novel task of cultural image captioning. Our contributions are as follows: (1) We introduce MosAIC, a Multi-Agent framework to enhance cross-cultural Image Captioning using LMMs with distinct cultural personas; (2) We provide a dataset of culturally enriched image captions in English for images from China, India, and Romania across three datasets: GeoDE, GD-VCR, CVQA; (3) We propose a culture-adaptable metric for evaluating cultural information within image captions; and (4) We show that the multi-agent interaction outperforms single-agent models across different metrics, and offer valuable insights for future research.

While large language models (LLMs) are dominating the field of natural language processing, it remains an open question how well these models can perform spatial reasoning. Contrary to recent studies suggesting that LLMs struggle with spatial reasoning tasks, we demonstrate in this paper that a novel prompting technique, termed Patient Visualization of Thought (Patient-VoT), can boost LLMs’ spatial reasoning abilities. The core idea behind Patient-VoT is to explicitly integrate *bullet lists, coordinates, and visualizations* into the reasoning process. By applying Patient-VoT, we achieve a significant boost in spatial reasoning performance compared to prior prompting techniques. We also show that integrating bullet lists into reasoning is effective in planning tasks, highlighting its general effectiveness across different applications.

Large language models (LLMs) exhibit remarkable capabilities on not just language tasks, but also various tasks that are not linguistic in nature, such as logical reasoning and social inference. In the human brain, neuroscience has identified a core language system that selectively and causally supports language processing. We here ask whether similar specialization for language emerges in LLMs. We identify language-selective units within 18 popular LLMs, using the same localization approach that is used in neuroscience. We then establish the causal role of these units by demonstrating that ablating LLM language-selective units – but not random units – leads to drastic deficits in language tasks. Correspondingly, language-selective LLM units are more aligned to brain recordings from the human language system than random units. Finally, we investigate whether our localization method extends to other cognitive domains: while we find specialized networks in some LLMs for reasoning and social capabilities, there are substantial differences among models. These findings provide functional and causal evidence for specialization in large language models, and highlight parallels with the functional organization in the brain.

The glyphic writing system of Chinese incorporates information-rich visual features in each character, such as radicals that provide hints about meaning or pronunciation. However, there has been no investigation into whether contemporary Large Language Models (LLMs) and Vision-Language Models (VLMs) can harness these sub-character features in Chinese through prompting. In this study, we establish a benchmark to evaluate LLMs’ and VLMs’ understanding of visual elements in Chinese characters, including radicals, composition structures, strokes, and stroke counts. Our results reveal that models surprisingly exhibit some, but still limited, knowledge of the visual information, regardless of whether images of characters are provided. To incite models’ ability to use radicals, we further experiment with incorporating radicals into the prompts for Chinese language processing (CLP) tasks. We observe consistent improvement in Part-Of-Speech tagging when providing additional information about radicals, suggesting the potential to enhance CLP by integrating sub-character information.

Last few years have seen unprecedented advances in capabilities of Large Language Models (LLMs). These advancements promise to benefit a vast array of application domains. However, due to their immense size, performing inference with LLMs is both costly and slow. Consequently, a plethora of recent work has proposed strategies to enhance inference efficiency, e.g., quantization, pruning, and caching. These acceleration strategies reduce the inference cost and latency, often by several factors, while maintaining much of the predictive performance measured via common benchmarks. In this work, we explore another critical aspect of LLM performance: demographic bias in model generations due to inference acceleration optimizations. Using a wide range of metrics, we probe bias in model outputs from a number of angles. Analysis of outputs before and after inference acceleration shows significant change in bias. Worryingly, these bias effects are complex and unpredictable. A combination of an acceleration strategy and bias type may show little bias change in one model but may lead to a large effect in another. Our results highlight a need for in-depth and case-by-case evaluation of model bias after it has been modified to accelerate inference.This paper contains prompts and outputs which may be deemed offensive.

Domain-specific machine translation (MT) poses significant challenges due to specialized terminology, particularly when translating across multiple languages with scarce resources. In this study, we present the first impact analysis of domain-specific terminology on multilingual MT for finance, focusing on European languages within the subdomain of macroeconomics. To this end, we construct a multi-parallel corpus from the European Central Bank, aligned across 22 languages. Using this resource, we compare open-source multilingual MT systems with large language models (LLMs) that possess multilingual capabilities. Furthermore, by developing and curating an English financial glossary, we propose a methodology to analyze the relationship between translation performance (into English) and the accuracy of financial term matching, obtaining significant correlation results. Finally, using the multi-parallel corpus and the English glossary, we automatically align a multilingual financial terminology, validating the English-Spanish alignments and incorporating them into our discussion. Our findings provide valuable insights into the current state of financial MT for European languages and offer resources for future research and system improvements.

Current evaluations of large language models (LLMs) often overlook non-determinism, typically focusing on a single output per example. This limits our understanding of LLM performance variability in real-world applications. Our study addresses this issue by exploring key questions about the performance differences between greedy decoding and sampling, identifying benchmarks’ consistency regarding non-determinism, and examining unique model behaviors. Through extensive experiments, we observe that greedy decoding generally outperforms sampling methods for most evaluated tasks. We also observe consistent performance across different LLM sizes and alignment methods, noting that alignment can reduce sampling variance. Moreover, our best-of-N sampling approach demonstrates that smaller LLMs can match or surpass larger models such as GPT-4-Turbo, highlighting the untapped potential of smaller LLMs. This research shows the importance of considering non-determinism in LLM evaluations and provides insights for future LLM development and evaluation.

Decoder-only language models have the ability to dynamically switch between various computational tasks based on input prompts. Despite many successful applications of prompting, there is very limited understanding of the internal mechanism behind such flexibility. In this work, we investigate how different prompting methods affect the geometry of representations in these models. Employing a framework grounded in statistical physics, we reveal that various prompting techniques, while achieving similar performance, operate through distinct representational mechanisms for task adaptation. Our analysis highlights critical geometric effects of input distribution samples and label semantics in few-shot in-context learning. We also demonstrate evidence of synergistic and interfering interactions between different tasks on the representational level. Our work contributes to the theoretical understanding of large language models and lays the groundwork for developing more effective, representation-aware prompting strategies.

This paper investigates whether large language models (LLMs) utilize numerical attributes encoded in a low-dimensional subspace of theembedding space when answering questions involving numeric comparisons, e.g., Was Cristiano born before Messi? We first identified,using partial least squares regression, these subspaces, which effectively encode the numerical attributes associated with the entities in comparison prompts. Further, we demonstrate causality, by intervening in these subspaces to manipulate hidden states, thereby altering the LLM’s comparison outcomes. Experiments conducted on three different LLMs showed that our results hold across different numerical attributes, indicating that LLMs utilize the linearly encoded information for numerical reasoning.

As language models (LMs) become capable of handling a wide range of tasks, their evaluation is becoming as challenging as their development. Most generation benchmarks currently assess LMs using abstract evaluation criteria-like helpfulness and harmlessness-which often lack the flexibility and granularity of human assessment. Additionally, these benchmarks tend to focus disproportionately on specific capabilities such as instruction following, leading to coverage bias. To overcome these limitations, we introduce the BiGGen Bench, a principled generation benchmark designed to thoroughly evaluate nine distinct capabilities of LMs across 77 diverse tasks. A key feature of the BiGGen Bench is its use of instance-specific evaluation criteria, closely mirroring the nuanced discernment of human evaluation. We apply this benchmark to assess 100 frontier LMs using five evaluator LMs. Our code, data, and evaluation results are all publicly available at https://github.com/prometheus-eval/prometheus-eval.

Sign language models could make modern language technologies more accessible to those who sign, but the supply of accurately labeled data struggles to meet the demand associated with training large, end-to-end neural models. As an alternative to this approach, we explore how knowledge about the linguistic structure of signs may be used as inductive priors for learning sign recognition and comprehension tasks. We first construct the American Sign Language Knowledge Graph (ASLKG) from 11 sources of linguistic knowledge, with emphasis on features related to signs’ phonological and lexical-semantic properties. Then, we use the ASLKG to train neuro-symbolic models on ASL video input tasks, achieving accuracies of 91% for isolated sign recognition, 14% for predicting the semantic features of unseen signs, and 36% for classifying the topic of Youtube-ASL videos.

Multilingual large language models (LLMs) have gained prominence, but concerns arise regarding their reliability beyond English. This study addresses the gap in cross-lingual semantic evaluation by introducing a novel benchmark for cross-lingual sense disambiguation, StingrayBench. In this paper, we demonstrate using false friends—words that are orthographically similar but have completely different meanings in two languages— as a possible approach to pinpoint the limitation of cross-lingual sense disambiguation in LLMs. We collect false friends in four language pairs, namely Indonesian-Malay, Indonesian-Tagalog, Chinese-Japanese, and English-German; and challenge LLMs to distinguish the use of them in context. In our analysis of various models, we observe they tend to be biased toward higher-resource languages. We also propose new metrics for quantifying the cross-lingual sense bias and comprehension based on our benchmark. Our work contributes to developing more diverse and inclusive language modeling, promoting fairer access for the wider multilingual community.

Text2Sql is a task that converts natural language questions into SQL queries. In previous research on LLM fine-tuning, researchers typically input both the entire database schema and the natural language question into the model. This approach has two issues: 1) the model’s context is limited when dealing with a large number of database tables; 2) the question is often related to only a few tables, leading to excessive irrelevant information that distracts the model. To address these issues, we employed pure fine-tuning strategy to reduce redundancy. The model fine-tuned with pure prompts, using prompts that are only 53% of the baseline length, outperforms the baseline (fine-tuned with all tables in the prompt) by 8.2% and 8.6% in Test-suite accuracy (TS) and exact-set-match accuracy (EM), respectively, on the Spider dev set. Under the most refined Spider dev set of prompts, the model achieves TS and EM scores of 73.5% and 75.4%, respectively, approaching state-of-the-art (SOTA) levels. To leverage the capabilities of the model with pure prompts, we applied pure knowledge distillation strategy to transfer its abilities. The distilled student model achieved a 1.9% improvement in TS, while the teacher model’s prompt length was only 23% of that of the student model.