AACL 2025 论文列表

In this work, we address the problem of text anonymization where the goal is to prevent adversaries from correctly inferring private attributes of the author, while keeping the text utility, i.e., meaning and semantics. We propose IncogniText, a technique that anonymizes the text to mislead a potential adversary into predicting a wrong private attribute value. Our empirical evaluation shows a reduction of private attribute leakage by more than across 8 different private attributes. Finally, we demonstrate the maturity of IncogniText for real-world applications by distilling its anonymization capability into a set of LoRA parameters associated with an on-device model. Our results show the possibility of reducing privacy leakage by more than half with limited impact on utility.

The present article introduces IndicClaimBuster, a novel multilingual claim verification dataset comprising \approx 9K claims and their corresponding evidence in English, Hindi, Bengali, and Hindi-English CodeMixed texts. The data set covers three key domains: politics, law and order, and health, to address the challenges of verifiable facts. Each claim was sourced from reputable Indian news portals and is accompanied by three pieces of evidence, two LLM-generated and one manually curated. Additionally, a separate attempt was conducted to generate refuted claims by employing an LLM. We further develop two frameworks: an unsupervised baseline and a two-stage pipeline that comprises evidence retrieval and veracity prediction modules. For retrieval, we fine-tuned SBERT models, with e5-base demonstrating superior average performance across languages, whereas for veracity prediction, multilingual transformers (mBERT, XLM-R, MuRIL, IndicBERTv2) were fine-tuned. Results indicate MuRIL and IndicBERTv2 excel in Indian languages, while XLM-R performs the best for CodeMix. Our work contributes a high-quality multilingual dataset and strong baseline methodologies, offering valuable resources for advancing automated claim verification in linguistically diverse and low-resource settings for Indian languages. The IndicClaimBuster dataset is available at: https://github.com/pritampal98/indic-claim-buster

The ability to accurately interpret implied meanings plays a crucial role in human communication and language use, and language models are also expected to possess this capability. This study demonstrates that providing language models with pragmatic theories as prompts is an effective in-context learning approach for tasks to understand implied meanings. Specifically, we propose an approach in which an overview of pragmatic theories, such as Gricean pragmatics and Relevance Theory, is presented as a prompt to the language model, guiding it through a step-by-step reasoning process to derive a final interpretation. Experimental results showed that, compared to the baseline, which prompts intermediate reasoning without presenting pragmatic theories (0-shot Chain-of-Thought), our methods enabled language models to achieve up to 9.6% higher scores on pragmatic reasoning tasks. Furthermore, we show that even without explaining the details of pragmatic theories, merely mentioning their names in the prompt leads to a certain performance improvement (around 1-3%) in larger models compared to the baseline.

We present a culturally-grounded multimodal dataset of 1,060 traditional recipes crowdsourced from rural communities across remote regions of Eastern India, spanning 10 endangered languages. These recipes, rich in linguistic and cultural nuance, were collected using a mobile interface designed for contributors with low digital literacy. Endangered Language Recipes (ELR)-1000—captures not only culinary practices but also the socio-cultural context embedded in indigenous food traditions. We evaluate the performance of several state-of-the-art large language models (LLMs) on translating these recipes into English and find the following: despite the models’ capabilities, they struggle with low-resource, culturally-specific language. However, we observe that providing targeted context—including background information about the languages, translation examples, and guidelines for cultural preservation—leads to significant improvements in translation quality. Our results underscore the need for benchmarks that cater to underrepresented languages and domains to advance equitable and culturally-aware language technologies. As part of this work, we release the ELR-1000 dataset to the NLP community, hoping it motivates the development of language technologies for endangered languages.

The potential of Large Language Models (LLMs) as instruments for measuring social phenomena is constrained by the methodological limitations of current probing techniques. Prevailing methods rely on static, handcrafted probe sets whose quality is highly dependent on their authors’ subjective expertise. This results in measurement tools with inconsistent statistical reliability that defy systematic optimization. Such an “artisanal” approach, akin to using an “uneven ruler,” undermines the scientific rigor of its findings and severely limits the applicability of LLMs in the social sciences. To elevate bias measurement from a craft to a science, we introduce the Psychometric-driven Probe Optimization (PMPO) framework. This framework treats a probe set as an optimizable scientific instrument and, for the first time, utilizes a Neural Genetic Algorithm that leverages a powerful LLM as a “neural genetic operator.” Through a hybrid strategy of gradient-guided mutation and creative rephrasing, PMPO automatically enhances the probe set’s reliability, sensitivity, and diversity. We first establish the external validity of our foundational measurement method (PLC), demonstrating a high correlation between its measurement of occupational gender bias and real-world U.S. Bureau of Labor Statistics data (average Pearson’s r=0.83, p<.001). Building on this, we show that the PMPO framework can elevate a standard probe set’s internal consistency (Cronbach’s Alpha) from 0.90 to an exceptional 0.96 within 10 generations. Critically, in a rigorous, double-blind “Turing Test,” probes evolved by PMPO from non-expert seeds were judged by sociology experts to have achieved a level of quality, sophistication, and nuance that is comparable to, and even indistinguishable from, those handcrafted by domain experts. This work provides a systematic pathway to upgrade LLM measurement tools from artisanal artifacts to automated scientific instruments, offering an unprecedented and trustworthy tool for AI safety auditing and computational social science.

Vision-Language Models (VLMs) often struggle to balance visual and textual information when summarizing complex multimodal inputs, such as entire TV show episodes. In this paper, we propose a zero-shot video-to-text summarization approach that builds its own screenplay-like representation of an episode, effectively integrating key video moments, dialogue, and character information into a unified document. Unlike previous approaches, we simultaneously generate screenplays and name the characters in zero-shot, using only the audio, video, and transcripts as input. Additionally, we highlight that existing summarization metrics can fail to assess the multimodal content in summaries. To address this, we introduce MFactSum, a multimodal metric that evaluates summaries with respect to both vision and text modalities. Using MFactSum, we evaluate our screenplay summaries on the SummScreen3D dataset, demonstrating superiority against state-of-the-art VLMs such as Gemini 1.5 by generating summaries containing 20% more relevant visual information while requiring 75% less of the video as input.

State-of-the-art Large Language Models (LLMs) achieve high pass@1 on general benchmarks like HumanEval (Chen et al., 2021) but underperform on specialized suites such as ParEval (Nichols et al., 2024). Is this due to LLMs missing domain knowledge or insufficient prompt detail is given? To answer this, we introduce PartialOrderEval, which augments any code generation benchmark with a partial order of prompts from minimal to maximally detailed. Applying it to HumanEval and both serial and OpenMP subsets of ParEval, we measure how pass@1 scales with prompt specificity. Our experiments with Llama-3.x and Qwen2.5-Coder demonstrate varying degrees of prompt sensitivity across different tasks, and a qualitative analysis highlights explicit I/O specifications, edge-case handling, and stepwise breakdowns as the key drivers of prompt detail improvement.

Large language models (LLMs) are increasingly being used for complex research tasks such as literature review, idea generation, and scientific paper analysis, yet their ability to truly understand and process the intricate relationships within complex research papers, such as the logical links between claims and supporting evidence remains largely unexplored. In this study, we present CLAIM-BENCH, a comprehensive benchmark for evaluating LLMs’ capabilities in scientific claim-evidence extraction and validation, a task that reflects deeper comprehension of scientific argumentation. We systematically compare three approaches which are inspired by divide and conquer approaches, across six diverse LLMs, highlighting model-specific strengths and weaknesses in scientific comprehension. Through evaluation involving over 300 claim-evidence pairs across multiple research domains, we reveal significant limitations in LLMs’ ability to process complex scientific content. Our results demonstrate that closed-source models like GPT-4 and Claude consistently outperform open-source counterparts in precision and recall across claim-evidence identification tasks. Furthermore, strategically designed three-pass and one-by-one prompting approaches significantly improve LLMs’ abilities to accurately link dispersed evidence with claims, although this comes at increased computational cost. CLAIM-BENCH sets a new standard for evaluating scientific comprehension in LLMs, offering both a diagnostic tool and a path forward for building systems capable of deeper, more reliable reasoning across full-length papers.

Large language models (LLMs) are increasingly widely used as critical components of knowledge retrieval systems and agentic systems. These systems can benefit from knowledge-seeking capabilities of LLMs, in other words, curiosity. However, this capability has not been evaluated quantitatively. Towards bridging this gap, we propose an evaluation framework, CDQG (Curiosity-Driven Question Generation). The CDQG task prompts LLMs to generate questions about a statement introducing scientific knowledge, simulating a curious person when facing the statement for the first time. The CDQG dataset contains 1,988 statements including physics, chemistry, and mathematics with distinct levels of difficulty, general knowledge statements, and intentionally erroneous statements. We score the qualities of the questions generated by LLMs along multiple dimensions. These scores are validated by rigorous controlled ablation studies and human evaluations. While large models like GPT-4 and Mistral 8x7b can generate highly coherent and relevant questions, the smaller Phi-2 model is equally or more effective. This indicates that size does not solely determine a model’s knowledge acquisition potential. CDQG quantifies a critical model capability, and opens up research opportunities for developing future knowledge retrieval systems driven by LLMs.

We introduce a set of resampling-based methods for quantifying uncertainty and statistical precision of evaluation metrics in multilingual and/or multitask NLP benchmarks.We show how experimental variation in performance scores arises from both model and data-related sources, and that accounting for both of them is necessary to avoid substantially underestimating the overall variability over hypothetical replications.Using multilingual question answering, machine translation, and named entity recognition as example tasks, we also demonstrate how resampling methods are useful for quantifying the replication uncertainty of various quantities used in leaderboards such as model rankings and pairwise differences between models.

Understanding the latent space geometry of large language models (LLMs) is key to interpreting their behavior and improving alignment. Yet it remains unclear to what extent LLMs linearly organize representations related to semantic understanding. To explore this, we conduct a large-scale empirical study of hidden representations in 11 autoregressive models across six scientific topics. We find that high-level semantic information consistently resides in low-dimensional subspaces that form linearly separable representations across domains. This separability becomes more pronounced in deeper layers and under prompts that elicit structured reasoning or alignment behavior—even when surface content remains unchanged. These findings motivate geometry-aware tools that operate directly in latent space to detect and mitigate harmful and adversarial content. As a proof of concept, we train an MLP probe on final-layer hidden states as a lightweight latent-space guardrail. This approach substantially improves refusal rates on malicious queries and prompt injections that bypass both the model’s built-in safety alignment and external token-level filters.

Quantization offers a practical solution to deploy LLMs in resource-constraint environments. However, its impact on internal representations remains understudied, raising questions about the reliability of quantized models. In this study, we employ a range of interpretability techniques to investigate how quantization affects model and neuron behavior. We analyze multiple LLMs under 4-bit and 8-bit quantization. Our findings reveal that the impact of quantization on model calibration is generally minor. Analysis of neuron activations indicates that the number of dead neurons, i.e., those with activation values close to 0 across the dataset, remains consistent regardless of quantization. In terms of neuron contribution to predictions, we observe that smaller full precision models exhibit fewer salient neurons, whereas larger models tend to have more, with the exception of Llama-2-7B. The effect of quantization on neuron redundancy varies across models. Overall, our findings suggest that effect of quantization may vary by model and tasks, however, we did not observe any drastic change which may discourage the use of quantization as a reliable model compression technique.

Multi-hop Question Generation (QG) effectively evaluates reasoning but remains confined to text; Video Question Generation (VideoQG) is limited to zero-hop questions over single segments. To address this, we introduce VideoChain, a novel Multi-hop Video Question Generation (MVQG) framework designed to generate questions that require reasoning across multiple, temporally separated video segments. VideoChain features a modular architecture built on a modified BART backbone enhanced with video embeddings, capturing textual and visual dependencies. Using the TVQA+ dataset, we automatically construct the large-scale MVQ-60 dataset by merging zero-hop QA pairs, ensuring scalability and diversity. Evaluations show VideoChain’s strong performance across standard generation metrics: ROUGE-L (0.6454), ROUGE-1 (0.6854), BLEU-1 (0.6711), BERTScore-F1 (0.7967), and semantic similarity (0.8110). These results highlight the model’s ability to generate coherent, contextually grounded, and reasoning-intensive questions. To facilitate future research, we publicly release our code and dataset.

Recent approaches have shown impressive proficiency in extracting and leveraging parametric knowledge from Large-Language Models (LLMs) and Vision-Language Models (VLMs). In this work, we consider how we can improve the retrieval of videos related to complex real-world events by automatically extracting latent parametric knowledge about those events. We present Q2E: a Query-to-Event decomposition method for zero-shot multilingual text-to-video retrieval, adaptable across datasets, domains, LLMs, or VLMs. Our approach demonstrates that we can enhance the understanding of otherwise overly simplified human queries by decomposing the query using the knowledge embedded in LLMs and VLMs. We additionally show how to apply our approach to both visual and speech-based inputs. To combine this varied multimodal knowledge, we adopt entropy-based fusion scoring for zero-shot fusion. Q2E outperforms the previous SOTA on the MultiVENT dataset by 8 NDCG points, while improving on MSR-VTT and MSVD by 4 and 3 points, respectively, outperforming several existing retrieval methods, including many fine-tuned and SOTA zero-shot approaches. We have released both code and data.

We introduced PerCoR—Persian Commonsense Reasoning—the first large-scale Persian benchmark for commonsense reasoning. PerCoR contains 106K multiple-choice sentence-completion problems drawn from more than forty news, cultural and other web sources. We adopt a linguistically grounded, conjunction-based segmentation strategy to generate coherent prefix–continuation pairs. To create challenging distractors, we propose DRESS-AF—Distractor Ranking via Embedding Similarity Scoring and Adversarial Filtering—a generation-free adversarial filtering method that selects distractors from the pool of gold continuations while maximising model confusion. Human annotators score 89% on PerCoR, while OpenAI-o3 achieves the highest performance at 92.18%, followed closely by Claude-Sonnet-3.7 (91.17%). The strongest open-source model, DeepSeek-R1, reaches 82.51%, underscoring both the dataset’s difficulty and the remaining performance gap in Persian commonsense reasoning. We further show that DRESS-AF transfers to the English HellaSwag benchmark, increasing its difficulty without hurting human solvability. The dataset is available at https://huggingface.co/datasets/MCINext/PerCoR .

Quotation recommendation is an inherently asymmetric retrieval task, where the intended meaning of a quote often diverges from surface expressions, creating significant semantic shifts. Combined with minimal lexical overlap, this poses a core challenge for classic dense retrievers, which struggle to capture non-literal and rhetorical alignments. To bridge this semantic gap, we propose introducing controllable signals to guide the model’s attention toward abstract, context-relevant concepts. We propose CtrlShift, a framework that leverages a Variational Autoencoder (VAE) to capture latent associations between context and quotation, which is used to derive context-aware control signals to modulate semantic focus and support bidirectional alignment and rhetorical intent modeling. Experiments show that our method consistently outperforms baselines on the quotation recommendation task and can be effectively transfered to the general purposed benchmark. Further, CtrlShift integrates seamlessly with general-purpose generative models without additional fine-tuning, and provides satisfactory interpretability by generating textual explaination to uncover the model’s focus on abstract, citation-aligned semantics.

Traditional dialogue summarization primarily focuses on dialogue content, assuming it comprises adequate information for a clear summary. However, this assumption often fails for discussions grounded in shared background, where participants frequently omit context and use implicit references. This results in summaries that are confusing to readers unfamiliar with the background. To address this, we introduce Knowledge-Grounded Discussion Summarization (KGDS), a novel task that produces a supplementary background summary for context and a clear opinion summary with clarified references. To facilitate research, we construct the first KGDS benchmark, featuring news-discussion pairs and expert-created multi-granularity gold annotations for evaluating sub-summaries. We also propose a novel hierarchical evaluation framework with fine-grained and interpretable metrics. Our extensive evaluation of 12 advanced large language models (LLMs) reveals that KGDS remains a significant challenge. The models frequently miss key facts and retain irrelevant ones in background summarization, and often fail to resolve implicit references in opinion summary integration.

We present Large Temporal Model, a Large Language Model (LLM) that excels in Temporal Relation Classification (TRC). We show how a carefully designed fine-tuning strategy, using a novel two-step fine-tuning approach, can adapt LLMs for TRC. Our approach is focused on global TRC, enabling simultaneous classification of all temporal relations within a document. Unlike traditional pairwise methods, our approach performs global inference in a single step, improving both efficiency and consistency. Evaluations on the MATRES and OmniTemp benchmarks demonstrate that, for the first time, an LLM achieves state-of-the-art performance, outperforming previous pairwise and global TRC methods. Results show that our global approach produces more consistent and accurate temporal graphs. Ablation studies further validate the effectiveness of our two-step fine-tuning strategy, while analyses reveal why our approach succeeds in increasing performance and reducing inconsistencies.

The output of large language models (LLMs) is unstable, due both to non-determinism of the decoding process as well as to prompt brittleness. While the intrinsic non-determinism of LLM generation may mimic existing uncertainty in human annotations through distributional shifts in outputs, it is largely assumed, yet unexplored, that the prompt brittleness effect is unique to LLMs. This raises the question: do human annotators show similar sensitivity to prompt changes? If so, should prompt brittleness in LLMs be considered problematic? One may alternatively hypothesize that prompt brittleness correctly reflects human annotation variances. To fill this research gap, we systematically compare the effects of prompt modifications on LLMs and identical instruction modifications for human annotators, focusing on the question of whether humans are similarly sensitive to prompt perturbations. To study this, we prompt both humans and LLMs for a set of text classification tasks conditioned on prompt variations. Our findings indicate that both humans and LLMs exhibit increased brittleness in response to specific types of prompt modifications, particularly those involving the substitution of alternative label sets or label formats. However, the distribution of human judgments is less affected by typographical errors and reversed label order than that of LLMs.