ACL 2025 论文列表

Detecting AI-generated text is a difficult problem to begin with; detecting AI-generated text on social media is made even more difficult due to the short text length and informal, idiosyncratic language of the internet. It is nonetheless important to tackle this problem, as social media represents a significant attack vector in online influence campaigns, which may be bolstered through the use of mass-produced AI-generated posts supporting (or opposing) particular policies, decisions, or events. We approach this problem with the mindset and resources of a reasonably sophisticated threat actor, and create a dataset of 505,159 AI-generated social media posts from a combination of open-source, closed-source, and fine-tuned LLMs, covering 11 different controversial topics. We show that while the posts can be detected under typical research assumptions about knowledge of and access to the generating models, under the more realistic assumption that an attacker will not release their fine-tuned model to the public, detectability drops dramatically. This result is confirmed with a human study. Ablation experiments highlight the vulnerability of various detection algorithms to fine-tuned LLMs. This result has implications across all detection domains, since fine-tuning is a generally applicable and realistic LLM use case.

Online misinformation remains a critical challenge, and fact-checkers increasingly rely on claim matching systems that use sentence embedding models to retrieve relevant fact-checks. However, as users interact with claims online, they often introduce edits, and it remains unclear whether current embedding models used in retrieval are robust to such edits. To investigate this, we introduce a perturbation framework that generates valid and natural claim variations, enabling us to assess the robustness of a wide-range of sentence embedding models in a multi-stage retrieval pipeline and evaluate the effectiveness of various mitigation approaches. Our evaluation reveals that standard embedding models exhibit notable performance drops on edited claims, while LLM-distilled embedding models offer improved robustness at a higher computational cost. Although a strong reranker helps to reduce the performance drop, it cannot fully compensate for first-stage retrieval gaps. To address these retrieval gaps, we evaluate train- and inference-time mitigation approaches, demonstrating that they can improve in-domain robustness by up to 17 percentage points and boost out-of-domain generalization by 10 percentage points. Overall, our findings provide practical improvements to claim-matching systems, enabling more reliable fact-checking of evolving misinformation.

Programming with a coding assistant is a fundamentally interactive process, yet existing static benchmarks fail to capture key features of model-user collaboration. We introduce an interactive evaluation pipeline to examine how LLMs incorporate different types of feedback in a collaborative setting, in which we obfuscate the input of static coding benchmarks so that the code model must interact with a simulated user. Across 10 models and 3 datasets, the relative rankings of models often permute greatly between static and interactive settings, despite models being fairly robust to feedback that contains errors. We also observe that similarly effective feedback types differ in terms of how models respond to higher- vs. lower-quality feedback. Moreover, feedback type impacts the degree to which the models make aesthetic or behavioral edits to their output. Our work aims to “re-evaluate” model coding capabilities through an interactive lens toward bridging the gap between existing evaluations and real-world usage.

Large Language Models (LLMs) are known to be vulnerable to backdoor attacks, where triggers embedded in poisoned samples can maliciously alter LLMs’ behaviors. In this paper, we move beyond attacking LLMs and instead examine backdoor attacks through the novel lens of natural language explanations. Specifically, we leverage LLMs’ generative capabilities to produce human-readable explanations for their decisions, enabling direct comparisons between explanations for clean and poisoned samples. Our results show that backdoored models produce coherent explanations for clean inputs but diverse and logically flawed explanations for poisoned data, a pattern consistent across classification and generation tasks for different backdoor attacks. Further analysis reveals key insights into the explanation generation process. At the token level, explanation tokens associated with poisoned samples only appear in the final few transformer layers. At the sentence level, attention dynamics indicate that poisoned inputs shift attention away from the original input context during explanation generation. These findings enhance our understanding of backdoor mechanisms in LLMs and present a promising framework for detecting vulnerabilities through explainability.

Prompt engineering for large language models is challenging, as even small prompt perturbations or model changes can significantly impact the generated output texts. Existing evaluation methods of LLM outputs, either automated metrics or human evaluation, have limitations, such as providing limited insights or being labor-intensive. We propose Spotlight, a new approach that combines both automation and human analysis. Based on data mining techniques, we automatically distinguish between random (decoding) variations and systematic differences in language model outputs. This process provides token patterns that describe the systematic differences and guide the user in manually analyzing the effects of their prompts and changes in models efficiently. We create three benchmarks to quantitatively test the reliability of token pattern extraction methods and demonstrate that our approach provides new insights into established prompt data. From a human-centric perspective, through demonstration studies and a user study, we show that our token pattern approach helps users understand the systematic differences of language model outputs. We are further able to discover relevant differences caused by prompt and model changes (e.g. related to gender or culture), thus supporting the prompt engineering process and human-centric model behavior research.

Text-embedding models often exhibit biases arising from the data on which they are trained. In this paper, we examine a hitherto unexplored bias in text-embeddings: bias arising from the presence of names such as persons, locations, organizations etc. in the text. Our study shows how the presence of name-bias in text-embedding models can potentially lead to erroneous conclusions in the assessment of thematic similarity. Text-embeddings can mistakenly indicate similarity between texts based on names in the text, even when their actual semantic content has no similarity or indicate dissimilarity simply because of the names in the text even when the texts match semantically. We first demonstrate the presence of name bias in different text-embedding models and then propose text-anonymization during inference which involves removing references to names, while preserving the core theme of the text. The efficacy of the anonymization approach is demonstrated on three downstream NLP tasks involving embedding similarities, achieving significant performance gains. Our simple and training-optimization-free approach offers a practical and easily implementable solution to mitigate name bias.

Mental-health stigma remains a pervasive social problem that hampers treatment-seeking and recovery. Existing resources for training neural models to finely classify such stigma are limited, relying primarily on social-media or synthetic data without theoretical underpinnings. To remedy this gap, we present an expert-annotated, theory-informed corpus of human-chatbot interviews, comprising 4,141 snippets from 684 participants with documented socio-cultural backgrounds. Our experiments benchmark state-of-the-art neural models and empirically unpack the challenges of stigma detection. This dataset can facilitate research on computationally detecting, neutralizing, and counteracting mental-health stigma. Our corpus is openly available at https://github.com/HanMeng2004/Mental-Health-Stigma-Interview-Corpus.

Recent advancements in large language models (LLMs) with extended context windows have significantly improved various tasks. To improve long-context capabilities, much work focuses on augmenting LLM’s capabilities with synthetic data. Existing methods often leverage the Self-Instruct framework to generate long-context instruction-tuning data. However, our preliminary experiments show that fewer than 35% of samples generated by Qwen-2-72B are multi-hop, and over 40% exhibit poor quality, limiting comprehensive understanding and further research. To address this, we propose the Multi-agent Interactive Multi-hop Generation (MIMG) framework, which integrates a quality verification agent, a single-hop question generation agent, a multiple question sampling strategy, and a multi-hop question merger agent. This framework significantly improves data quality, with high-quality, multi-hop, and diverse data. Furthermore, we conduct a thorough analysis of document selection, question merging, and validation techniques through extensive experiments across various models. Our results demonstrate that synthetic high-quality long-context instruction data can enhance model performance, surpassing even models trained on larger amounts of human-annotated data.

We investigate long-context vulnerabilities in Large Language Models (LLMs) through Many-Shot Jailbreaking (MSJ). Our experiments utilize context length of up to 128K tokens. Through comprehensive analysis with various many-shot attack settings with different instruction styles, shot density, topic, and format, we reveal that context length is the primary factor determining attack effectiveness. Critically, we find that successful attacks do not require carefully crafted harmful content. Even repetitive shots or random dummy text can circumvent model safety measures, suggesting fundamental limitations in long-context processing capabilities of LLMs. The safety behavior of well-aligned models becomes increasingly inconsistent with longer contexts. These findings highlight significant safety gaps in context expansion capabilities of LLMs, emphasizing the need for new safety mechanisms.

In this work, we conduct systematic research in a challenging area: the automatic evaluation of book-length stories (>100K tokens). Our study focuses on two key questions: (1) understanding which evaluation aspects matter most to readers, and (2) exploring effective methods for evaluating lengthy stories. We introduce the first large-scale benchmark, **LongStoryEval**, comprising 600 newly published books with an average length of 121K tokens (maximum 397K). Each book includes its average rating and multiple reader reviews, presented as critiques organized by evaluation aspects. By analyzing all user-mentioned aspects, we propose an *evaluation criteria structure* and conduct experiments to identify the most significant aspects among the 8 top-level criteria. For evaluation methods, we compare the effectiveness of three types: *aggregation-based*, *incremental-updated*, and *summary-based* evaluations. Our findings reveal that aggregation- and summary-based evaluations perform better, with the former excelling in detail assessment and the latter offering greater efficiency. Building on these insights, we further propose **NovelCritique**, an 8B model that leverages the efficient summary-based framework to review and score stories across specified aspects. NovelCritique outperforms commercial models like GPT-4o in aligning with human evaluations. All our datasets and codes will be released to foster further research.

As large language models (LLMs) have progressed towards more human-like and human–AI communications prevalent, prompting has emerged as a decisive component. However, there is limited conceptual consensus on what exactly quantifies natural language prompts. We attempt to address this question by conducting a meta-analysis surveying 150+ prompting-related papers from leading NLP and AI conferences (2022–2024), and blogs. We propose a property- and human-centric framework for evaluating prompt quality, encompassing 21 properties categorized into six dimensions. We then examine how existing studies assess their impact on LLMs, revealing their imbalanced support across models and tasks, and substantial research gaps. Further, we analyze correlations among properties in high-quality natural language prompts, deriving prompting recommendations. Finally, we explore multi-property prompt enhancements in reasoning tasks, observing that single-property enhancements often have the greatest impact. Our findings establish a foundation for property-centric prompt evaluation and optimization, bridging the gaps between human–AI communication and opening new prompting research directions.

In this study, we investigate whether non-English-centric large language models, ‘think’ in their specialized language. Specifically, we analyze how intermediate layer representations, when projected into the vocabulary space, favor certain languages during generation—termed as latent languages. We categorize non-English-centric models into two groups: CPMs, which are English-centric models with continued pre-training on its specialized language, and BLMs, which are pre-trained on a balanced mix of multiple languages from scratch. Our findings reveal that while English-centric models rely exclusively on English as their latent language, non-English-centric models activate multiple latent languages, dynamically selecting the most similar one based on both the source and target languages. This also influences responses to culture difference questions, reducing English-centric biases in non-English models. This study deepens our understanding of language representation in non-English-centric LLMs, shedding light on the intricate dynamics of multilingual processing at the representational level.

Transforming recorded videos into concise and accurate textual summaries is a growing challenge in multimodal learning. This paper introduces VISTA, a dataset specifically designed for video-to-text summarization in scientific domains. VISTA contains 18,599 recorded AI conference presentations paired with their corresponding paper abstracts. We benchmark the performance of state-of-the-art large models and apply a plan-based framework to better capture the structured nature of abstracts. Both human and automated evaluations confirm that explicit planning enhances summary quality and factual consistency. However, a considerable gap remains between models and human performance, highlighting the challenges of our dataset. This study aims to pave the way for future research on scientific video-to-text summarization.

What makes a difference in the post-training of LLMs? We investigate the training patterns of different layers in large language models (LLMs) through the lens of the gradient. We are specifically interested in how fast vs. slow thinking affects the layer-wise gradients, given the recent popularity of training LLMs on reasoning paths such as chain-of-thoughts (CoT) and process rewards. In our study, fast thinking without CoT leads to larger gradients and larger differences of gradients across layers than slow thinking (Detailed CoT), indicating the learning stability brought by the latter. Additionally, we study whether the gradient patterns can reflect the correctness of responses when training different LLMs using slow vs. fast thinking paths. The results show that the gradients of slow thinking can distinguish correct and irrelevant reasoning paths. As a comparison, we conduct similar gradient analyses on non-reasoning knowledge learning tasks, on which, however, trivially increasing the response length does not lead to similar behaviors of slow thinking. Our study strengthens fundamental understandings of LLM training and sheds novel insights on its efficiency and stability, which pave the way towards building a generalizable System-2 agent.

Large Language Models (LLMs) require alignment with human preferences to avoid generating offensive, false, or meaningless content. Recently, low-resource methods for LLM alignment have been popular, while still facing challenges in obtaining both high-quality and aligned content. Motivated by the observation that the difficulty of generating aligned responses is concentrated at the beginning of decoding, we propose a novel framework, Weak-to-Strong Decoding (WSD), to enhance the alignment ability of base models by the guidance of a small aligned model. The small model first drafts well-aligned beginnings, followed by the large base model to continue the rest, controlled by a well-designed auto-switch mechanism. We also collect a new dataset, GenerAlign, to fine-tune a small-sized Pilot-3B as the draft model, which effectively enhances different base models under the WSD framework to outperform all baseline methods, while avoiding degradation on downstream tasks, termed as the alignment tax. Extensive experiments are further conducted to examine the impact of different settings and time efficiency, as well as analyses on the intrinsic mechanisms of WSD in depth.

Parameter-efficient fine-tuning (PEFT) large language models (LLMs) have shown impressive performance in various downstream tasks. However, in many real-world scenarios, the collected training data inevitably contains noisy labels. To learn from noisy labels, most solutions select samples with small losses for model training. However, the selected samples, in turn, impact the loss computation in the next iteration. An inaccurate initial selection can create a vicious cycle, leading to suboptimal performance. To break this cycle, we propose Delora, a novel framework that decouples the sample selection from model training. For sample selection, Delora establishes a noisy label detector by introducing clean and noisy LoRA. Benefiting from the memory effect, the clean LoRA is encouraged to memorize clean data, while the noisy LoRA is constrained to memorize mislabeled data, which serves as a learnable threshold for selecting clean and noisy samples. For model training, Delora can use carefully selected samples to fine-tune language models seamlessly. Experimental results on synthetic and real-world noisy datasets demonstrate the effectiveness of Delora in noisy label detection and text classification.

Retrieval-augmented generation (RAG) demonstrates remarkable performance across tasks in open-domain question-answering. However, traditional search engines may retrieve shallow content, limiting the ability of LLMs to handle complex, multi-layered information. To address this, we introduce WebWalkerQA, a benchmark designed to assess the ability of LLMs to perform web traversal. It evaluates the capacity of LLMs to traverse a website’s subpages to extract high-quality data systematically. We propose WebWalker, which is a multi-agent framework that mimics human-like web navigation through an explore-critic paradigm. Extensive experimental results show that WebWalkerQA is challenging and demonstrates the effectiveness of RAG combined with WebWalker, through this horizontal and vertical integration in real-world scenarios.

With the rapid advancement of Generative AI technology, Multimodal Large Language Models(MLLMs) have the potential to act as AI software engineers capable of executing complex web application development. Considering that the model requires a confluence of multidimensional sub-capabilities to address the challenges of various development phases, constructing a multi-view evaluation framework is crucial for accurately guiding the enhancement of development efficiency. However, existing benchmarks usually fail to provide an assessment of sub-capabilities and focus solely on webpage generation outcomes. In this work, we draw inspiration from the principles of software engineering and further propose WebUIBench, a benchmark systematically designed to evaluate MLLMs in four key areas: WebUI Perception, HTML Programming, WebUI-HTML Understanding, and WebUI-to-Code. WebUIBench comprises 21K high-quality question-answer pairs derived from over 0.7K real-world websites. The extensive evaluation of 29 mainstream MLLMs uncovers the skill characteristics and various weakness that models encountered during the development process.

The main goal of this work is the creation of the Italian version of the WebNLG corpus through the application of Neural Machine Translation (NMT) and post-editing with hand-written rules. To achieve this goal, in a first step, several existing NMT models were analysed and compared in order to identify the system with the highest performance on the original corpus. In a second step, after using the best NMT system, we semi-automatically designed and applied a number of rules to refine and improve the quality of the produced resource, creating a new corpus named WebNLG-IT. We used this resource for fine-tuning several LLMs for RDF-to-text tasks. In this way, comparing the performance of LLM-based generators on both Italian and English, we have (1) evaluated the quality of WebNLG-IT with respect to the original English version, (2) released the first fine-tuned LLM-based system for generating Italian from semantic web triples and (3) introduced an Italian version of a modular generation pipeline for RDF-to-text.