ACL 2025 论文列表

Language is an intricately structured system, and a key goal of NLP interpretability is to provide methodological insights for understanding how language models internally represent this structure. In this paper, we use Shapley Taylor interaction indices (STII) in order to examine how language and speech models internally relate and structure their inputs. Pairwise Shapley interactions give us an attribution measure of how much two inputs work together to influence model outputs beyond if we linearly added their independent influences, providing a view into how models encode structural interactions between inputs. We relate the interaction patterns in models to three underlying linguistic structures: syntactic structure, non-compositional semantics, and phonetic interaction. We find that autoregressive text models encode interactions that correlate with the syntactic proximity of inputs, and that both autoregressive and masked models encode nonlinear interactions in idiomatic phrases with non-compositional semantics. Our speech results show that inputs are more entangled for pairs where a neighboring consonant is likely to influence a vowel or approximant, showing that models encode the phonetic interaction needed for extracting discrete phonemic representations.

This paper argues that the relationship between lexical identity and prosody—one well-studied parameter of linguistic variation—can be characterized using information theory. We predict that languages that use prosody to make lexical distinctions should exhibit a higher mutual information between word identity and prosody, compared to languages that don’t. We test this hypothesis in the domain of pitch, which is used to make lexical distinctions in tonal languages, like Cantonese. We use a dataset of speakers reading sentences aloud in ten languages across five language families to estimate the mutual information between the text and their pitch curves. We find that, across languages, pitch curves display similar amounts of entropy. However, these curves are easier to predict given their associated text in the tonal languages, compared to pitch- and stress-accent languages, and thus the mutual information is higher in these languages, supporting our hypothesis. Our results support perspectives that view linguistic typology as gradient, rather than categorical.

With the continuous advancement in the performance of open-source large language models (LLMs), their inference services have attracted a substantial user base by offering quality comparable to closed-source models at a significantly lower cost. However, it has also given rise to trust issues regarding model consistency between users and third-party service providers. Specifically, service providers can effortlessly degrade a model’s parameter scale or precision for more margin profits, and although users may perceptibly experience differences in text quality, they often lack a reliable method for concrete monitoring. To address this problem, we propose a paradigm for model consistency monitoring on the user side. It constructs metrics based on the logits produced by LLMs to differentiate sequences generated by degraded models. Furthermore, by leveraging model offloading techniques, we demonstrate that the proposed method is implementable on consumer-grade devices. Metric evaluations conducted on three widely used LLMs series (OPT, Llama 3.1 and Qwen 2.5) along with system prototype efficiency tests on a consumer device (RTX 3080 TI) confirm both the effectiveness and feasibility of the proposed approach.

Measuring the generalization ability of Large Language Models (LLMs) is challenging due to data contamination. As models grow and computation becomes cheaper, ensuring tasks and test cases are unseen during training phases will become nearly impossible. We argue that knowledge-retrieval and reasoning tasks are not ideal for measuring generalization, as LLMs are not trained for specific tasks. Instead, we propose user behavior prediction, also a key aspect of personalization, as a theoretically sound, scalable, and robust alternative. We introduce a novel framework for this approach and test it on movie and music recommendation datasets for GPT-4o, GPT-4o-mini, and Llama-3.1-8B-Instruct. Results align with our framework’s predictions, showing GPT-4o outperforms GPT-4o-mini and Llama, though all models have much room for improvement, especially Llama.

Large multimodal models exhibit remarkable intelligence, yet their embodied cognitive abilities during motion in open-ended urban aerial spaces remain to be explored. We introduce a benchmark to evaluate whether video-large language models (Video-LLMs) can naturally process continuous first-person visual observations like humans, enabling recall, perception, reasoning, and navigation. We have manually control drones to collect 3D embodied motion video data from real-world cities and simulated environments, resulting in 1.5k video clips. Then we design a pipeline to generate 5.2k multiple-choice questions. Evaluations of 17 widely-used Video-LLMs reveal current limitations in urban embodied cognition. Correlation analysis provides insight into the relationships between different tasks, showing that causal reasoning has a strong correlation with recall, perception, and navigation, while the abilities for counterfactual and associative reasoning exhibit lower correlation with other tasks. We also validate the potential for Sim-to-Real transfer in urban embodiment through fine-tuning.

Mixture-of-Experts (MoE) shines brightly in large language models (LLMs) and demonstrates outstanding performance in plentiful natural language processing tasks. However, existing methods transforming LLMs from dense to MoE face significant data requirements and typically rely on large-scale post-training. In this paper, we propose Upcycling Instruction Tuning (UpIT), a data-efficient approach for tuning a dense pre-trained model into a MoE instruction model. Specifically, we first point out that intermediate checkpoints during instruction tuning of the dense model are naturally suitable for specialized experts, and then propose an expert expansion stage to flexibly achieve models with flexible numbers of experts, where genetic algorithm and parameter merging are introduced to ensure sufficient diversity of new extended experts. To ensure that each specialized expert in the MoE model works as expected, we select a small amount of seed data that each expert excels to pre-optimize the router. Extensive experiments with various data scales and upcycling settings demonstrate the outstanding performance and data efficiency of UpIT, as well as stable improvement in expert or data scaling. Further analysis reveals the importance of ensuring expert diversity in upcycling.

Maximum a posteriori decoding, a commonly used method for neural machine translation (NMT), aims to maximize the estimated posterior probability. However, high estimated probability does not always lead to high translation quality. Minimum Bayes Risk (MBR) decoding offers an alternative by seeking hypotheses with the highest expected utility.Inspired by Quality Estimation (QE) reranking which uses the QE model as a ranker, we propose source-based MBR (sMBR) decoding, a novel approach that utilizes quasi-sources (generated via paraphrasing or back-translation) as “support hypotheses” and a reference-free quality estimation metric as the utility function, marking the first work to solely use sources in MBR decoding. Experiments show that sMBR outperforms QE reranking and the standard MBR decoding. Our findings suggest that sMBR is a promising approach for NMT decoding.

BERT-family have been increasingly explored for adaptation to scenarios beyond language understanding tasks, with more recent efforts focused on enabling them to become good instruction followers. These explorations have endowed BERT-family with new roles and human expectations, showcasing their potential on par with current state-of-the-art (SOTA) large language models (LLMs). However, several certain shortcomings in previous BERT-family, such as the relatively sub-optimal training corpora, learning procedure, and model architecture, all impede the further advancement of these models for serving as general and competitive LLMs. Therefore, we aim to address these deficiencies in this paper. Our study not only introduces a more suitable pre-training task that helps BERT-family excel in wider applications to realize generality but also explores the integration of cutting-edge technologies into our model to further enhance their capabilities. Our final models, termed **Bi**directional **G**eneral **L**anguage **M**odels (**BiGLM**), exhibit performance levels comparable to current SOTA LLMs across a spectrum of tasks. Moreover, we conduct detailed analyses to study the effects of scaling and training corpora for BiGLM. To the best of our knowledge, our work represents the early attempt to offer a recipe for building novel types of scalable, general, and competitive LLMs that diverge from current autoregressive modeling methodology. Our codes and models are available on Github.

Large Vision Language Models (LVLMs) have shown impressive performance on various vision-language tasks. However, while objects in natural scenes inevitably exhibit visual variations in position, scale, orientation, and context due to changes in viewpoint and environment, the robustness of LVLMs to these fundamental visual variations remains largely unexplored. To address this gap, we introduce V²R-Bench, a comprehensive benchmark framework for evaluating Visual Variation Robustness of LVLMs, which encompasses automated evaluation dataset generation and principled metrics for thorough robustness assessment. Through extensive evaluation of 13 LVLMs, we reveal a surprising vulnerability to visual variations, affecting even advanced models that excel at complex vision-language tasks yet significantly underperform on simple tasks like object recognition. Interestingly, these models exhibit a distinct visual position bias that contradicts theories of effective receptive fields and demonstrate a human-like visual acuity threshold. To identify the source of these vulnerabilities, we propose a systematic framework for component-level analysis, featuring a novel visualization approach for aligned visual features. Results show that these vulnerabilities stem from error accumulation in the pipeline architecture and inadequate multimodal alignment. Complementary experiments with synthetic data further demonstrate that these limitations are fundamentally architectural challenges, underscoring the need for architectural innovations in future LVLM designs.

Large Language Models (LLMs) excel in reasoning tasks through Chain-of-Thought (CoT) prompting. However, CoT prompting greatly increases computational demands, which has prompted growing interest in distilling CoT capabilities into Small Language Models (SLMs). This study systematically examines the factors influencing CoT distillation, including the choice of granularity, format and teacher model. Through experiments involving four teacher models and seven student models across seven mathematical and commonsense reasoning datasets, we uncover three key findings: (1) Unlike LLMs, SLMs exhibit a *non-monotonic* relationship with granularity, with stronger models benefiting from finer-grained reasoning and weaker models performing better with simpler CoT supervision; (2) CoT format significantly impacts LLMs but has *minimal* effect on SLMs, likely due to their reliance on supervised fine-tuning rather than pretraining preferences; (3) Stronger teacher models do *NOT* always produce better student models, as diversity and complexity in CoT supervision can outweigh accuracy alone. These findings emphasize the need to tailor CoT strategies to specific student model, offering actionable insights for optimizing CoT distillation in SLMs.

Although substantial efforts have been made to mitigate catastrophic forgetting in continual learning, the intrinsic mechanisms are not well understood. In this work, we demonstrate the existence of “pseudo forgetting”: the performance degradation in previous tasks is not attributed to a loss of capabilities, but rather to the failure of the instructions to activate the appropriate model capabilities. We show that the model’s performance on previous tasks can be restored through two simple interventions: (1) providing partial external correct rationale, and (2) appending semantically meaningless suffixes to the original instructions, to guide the generation of correct rationales. Through empirical analysis of the internal mechanisms governing rationale generation, we reveal that models exhibiting pseudo forgetting show reduced instruction dependence during rationale generation, leading to suboptimal activation of their inherent capabilities. Based on this insight, we propose Rationale-Guidance Difficulty based Replay (RGD-R) framework that dynamically allocates replay data based on the model’s ability to correctly leverage the intrinsic capabilities. Experimental results demonstrate that RGD-R effectively mitigates pseudo forgetting while maintaining model plasticity.

Privacy risks in text-only Large Language Models (LLMs) are well studied, particularly their tendency to memorize and leak sensitive information. However, Multi-modal Large Language Models (MLLMs), which process both text and images, introduce unique privacy challenges that remain underexplored. Compared to text-only models, MLLMs can extract and expose sensitive information embedded in images, posing new privacy risks. We reveal that some MLLMs are susceptible to privacy breaches, leaking sensitive data embedded in images or stored in memory. Specifically, in this paper, we (1) introduce MM-Privacy, a comprehensive dataset designed to assess privacy risks across various multi-modal tasks and scenarios, where we define Disclosure Risks and Retention Risks. (2) systematically evaluate different MLLMs using MM-Privacy and demonstrate how models leak sensitive data across various tasks, and (3) provide additional insights into the role of task inconsistency in privacy risks, emphasizing the urgent need for mitigation strategies. Our findings highlight privacy concerns in MLLMs, underscoring the necessity of safeguards to prevent data exposure. Part of our dataset and code can be found here.

Large Language Model (LLM) agents have become increasingly prevalent across various real-world applications. They enhance decision-making by storing private user-agent interactions in the memory module for demonstrations, introducing new privacy risks for LLM agents. In this work, we systematically investigate the vulnerability of LLM agents to our proposed Memory EXTRaction Attack (MEXTRA) under a black-box setting. To extract private information from memory, we propose an effective attacking prompt design and an automated prompt generation method based on different levels of knowledge about the LLM agent. Experiments on two representative agents demonstrate the effectiveness of MEXTRA. Moreover, we explore key factors influencing memory leakage from both the agent designer’s and the attacker’s perspectives. Our findings highlight the urgent need for effective memory safeguards in LLM agent design and deployment.

The mechanisms behind multilingual capabilities in Large Language Models (LLMs) have been examined using neuron-based or internal-activation-based methods. However, these methods often face challenges such as superposition and layer-wise activation variance, which limit their reliability. Sparse Autoencoders (SAEs) offer a more nuanced analysis by decomposing the activations of LLMs into a sparse linear combination of SAE features. We introduce a novel metric to assess the monolinguality of features obtained from SAEs, discovering that some features are strongly related to specific languages. Additionally, we show that ablating these SAE features only significantly reduces abilities in one language of LLMs, leaving others almost unaffected. Interestingly, we find some languages have multiple synergistic SAE features, and ablating them together yields greater improvement than ablating individually. Moreover, we leverage these SAE-derived language-specific features to enhance steering vectors, achieving control over the language generated by LLMs. The code is publicly available at https://github.com/Aatrox103/multilingual-llm-features.

Large language models (LLMs) offer powerful capabilities but come with significant environmental impact, particularly in carbon emissions. Existing studies benchmark carbon emissions but lack a standardized basis for comparison across different model configurations. To address this, we introduce the concept of functional unit (FU) as a standardized basis and develop FUEL, the first FU-based framework for evaluating LLM serving’s environmental impact. Through three case studies, we uncover key insights and trade-offs in reducing carbon emissions by optimizing model size, quantization strategy, and hardware choice, paving the way for more sustainable LLM serving. The code is available at https://github.com/jojacola/FUEL.

Despite the impressive performance of multilingual large language models (mLLMs) in various natural language processing tasks, their ability to understand procedural texts, particularly those with culture-specific content, remains largely unexplored. Texts describing cultural procedures, including rituals, traditional craftsmanship, and social etiquette, require an inherent understanding of cultural context, presenting a significant challenge for mLLMs. In this work, we introduce CAPTex, a benchmark designed to evaluate mLLMs’ ability to process and reason over culturally diverse procedural texts in multiple languages. Using a range of evaluation methods, we find that (1) mLLMs struggle with culturally contextualized procedural content, particularly in low-resource languages; (2) performance varies across cultural domains, with some proving more difficult than others; and (3) models perform better on multiple-choice tasks presented in conversational formats than on direct questions. These results highlight the current limitations of mLLMs and emphasize the need for culturally informed benchmarks like CAPTex to support more accurate and inclusive language understanding.

Chain-of-thought (CoT) prompting has been widely adopted to enhance the reasoning capabilities of large language models (LLMs). However, the effectiveness of CoT reasoning is inconsistent across tasks with different reasoning types. This work presents a novel perspective to understand CoT behavior through the lens of confirmation bias in cognitive psychology. Specifically, we examine how model internal beliefs, approximated by direct question-answering probabilities, affect both reasoning generation (Q → R) and reasoning-guided answer prediction (QR → A) in CoT. By decomposing CoT into a two-stage process, we conduct a thorough correlation analysis in model beliefs, rationale attributes, and stage-wise performance. Our results provide strong evidence of confirmation bias in LLMs, such that model beliefs not only skew the reasoning process but also influence how rationales are utilized for answer prediction. Furthermore, the interplay between task vulnerability to confirmation bias and the strength of beliefs also provides explanations for CoT effectiveness across reasoning tasks and models. Overall, this study provides a valuable insight for the needs of better prompting strategies that mitigate confirmation bias to enhance reasoning performance. Code is available at https://github.com/yuewan2/biasedcot.

Dynamical systems theory provides a framework for analyzing iterative processes and evolution over time. Within such systems, repetitive transformations can lead to stable configurations, known as attractors, including fixed points and limit cycles. Applying this perspective to large language models (LLMs), which iteratively map input text to output text, provides a principled approach to characterizing long-term behaviors. Successive paraphrasing serves as a compelling testbed for exploring such dynamics, as paraphrases re-express the same underlying meaning with linguistic variation. Although LLMs are expected to explore a diverse set of paraphrases in the text space, our study reveals that successive paraphrasing converges to stable periodic states, such as 2-period attractor cycles, limiting linguistic diversity. This phenomenon is attributed to the self-reinforcing nature of LLMs, as they iteratively favour and amplify certain textual forms over others. This pattern persists with increasing generation randomness or alternating prompts and LLMs. These findings underscore inherent constraints in LLM generative capability, while offering a novel dynamical systems perspective for studying their expressive potential.

Document retrieval in real-world scenarios faces significant challenges due to diverse document formats and modalities. Traditional text-based approaches rely on tailored parsing techniques that disregard layout information and are prone to errors, while recent parsing-free visual methods often struggle to capture fine-grained textual semantics in text-rich scenarios. To address these limitations, we propose Unveil, a novel visual-textual embedding framework that effectively integrates textual and visual features for robust document representation. Through knowledge distillation, we transfer the semantic understanding capabilities from the visual-textual embedding model to a purely visual model, enabling efficient parsing-free retrieval while preserving semantic fidelity. Experimental results demonstrate that our visual-textual embedding method surpasses existing approaches, while knowledge distillation successfully bridges the performance gap between visual-textual and visual-only methods, improving both retrieval accuracy and efficiency.