AACL 2025 Papers

We assess the capabilities of large language models on tasks involving Buddhist texts written in Sanskrit and Classical Tibetan—two typologically distinct, low-resource historical languages. To this end, we introduce DharmaBench, a benchmark suite comprising 13 classification and detection tasks grounded in Buddhist textual traditions: six in Sanskrit and seven in Tibetan, with four shared across both. The tasks are curated from scratch, tailored to the linguistic and cultural characteristics of each language. We evaluate a range of models, from proprietary systems like GPT-4o to smaller, domain-specific open-weight models, analyzing their performance across tasks and languages. All datasets and code are publicly released, under the CC-BY-4 License and the Apache-2.0 License respectively, to support research on historical language processing and the development of culturally inclusive NLP systems.

Document Visual Question Answering (DocVQA) models often produce overconfident or ethically misaligned responses, especially under uncertainty. Existing models like LayoutLMv3, UDOP, and DONUT focus on accuracy but lack ethical calibration. We propose HonestVQA, a model-agnostic, self-supervised framework that aligns model confidence with correctness using weighted loss and contrastive learning. We introduce two new metrics—Honesty Score (H-Score) and Ethical Confidence Index (ECI)—to evaluate ethical alignment. HonestVQA improves accuracy and F1 by up to 4.3% across SpDocVQA, InfographicsVQA, and SROIE datasets, while reducing overconfidence. It also generalizes well across domains, achieving 78.9% accuracy and 76.1% F1-score.

Large Language Models (LLMs) demonstrate strong multilingual capabilities but are costly to deploy due to their size and computational demands. To mitigate this, compression techniques such as pruning and quantization are widely used. However, these methods face two key limitations: (1) they assume access to high-quality instruction or calibration data, which is often unavailable for low-resource languages; and (2) they aim to preserve multilingual generality, making them inefficient for language-specific applications. We introduce LangCompress, a language-aware compression framework that enhances existing compression methods for targeted deployment. LangCompress is method-agnostic and improves state-of-the-art pruning and quantization approaches. It features two core components: an iterative self-supervised pipeline for generating instruction data in the target language, and a vocabulary simplification strategy that reduces the LM head to focus on key tokens. Experiments on perplexity, translation, and summarization tasks show that LangCompress improves performance in the target language. The code and data are publicly available.

We introduce InterChart, a diagnostic benchmark that evaluates how well vision-language models (VLMs) reason across multiple related charts, a task central to real-world applications such as scientific reporting, financial analysis, and public policy dashboards. Unlike prior benchmarks focusing on isolated, visually uniform charts, InterChart challenges models with diverse question types ranging from entity inference and trend correlation to numerical estimation and abstract multi-step reasoning grounded in 2–3 thematically or structurally related charts. We organize the benchmark into three tiers of increasing difficulty: (1) factual reasoning over individual charts, (2) integrative analysis across synthetically aligned chart sets, and (3) semantic inference over visually complex, real-world chart pairs. Our evaluation of state-of-the-art open- and closed-source VLMs reveals consistent and steep accuracy declines as chart complexity increases. We find that models perform better when we decompose multi-entity charts into simpler visual units, underscoring their struggles with cross-chart integration. By exposing these systematic limitations, InterChart provides a rigorous framework for advancing multimodal reasoning in complex, multi-visual environments.

By augmenting Large Language Models (LLMs) with external tools, their capacity to solve complex problems has been significantly enhanced. However, despite ongoing advancements in the parsing capabilities of LLMs, incorporating all available tools simultaneously in the prompt remains impractical due to the vast number of external tools. Consequently, it is essential to provide LLMs with a precise set of tools tailored to the specific task, considering both quantity and quality. Current tool retrieval methods primarily focus on refining the ranking list of tools and directly packaging a fixed number of top-ranked tools as the tool set. However, these approaches often fail to equip LLMs with the optimal set of tools prior to execution, since the optimal number of tools for different tasks could be different, resulting in inefficiencies such as redundant or unsuitable tools, which impede immediate access to the most relevant tools. This paper addresses the challenge of recommending precise toolsets for LLMs. We introduce the problem of tool recommendation, define its scope, and propose a novel Precision-driven Tool Recommendation (PTR) approach. PTR captures an initial, concise set of tools by leveraging historical tool bundle usage and dynamically adjusts the tool set by performing tool matching, culminating in a multi-view-based tool addition. Additionally, we present a new dataset, RecTools, and a metric, TRACC, designed to evaluate the effectiveness of tool recommendation for LLMs. We further validate our design choices through comprehensive experiments, demonstrating promising accuracy across two open benchmarks and our RecTools dataset.

State space models (SSMs) are a promising alternative to transformers for language modeling because they use fixed memory during inference. However, this fixed memory usage requires some information loss in the hidden state when processing long sequences. While prior work has studied the sequence length at which this information loss occurs, it does not characterize the types of information SSM language models (LMs) tend to forget. In this paper, we address this knowledge gap by identifying the types of tokens (e.g., parts of speech, named entities) and sequences (e.g., code, math problems) that are more frequently forgotten by SSM LMs. We achieve this by training an auto-encoder to reconstruct sequences from the SSM’s hidden state, and measure information loss by comparing inputs with their reconstructions. We perform experiments using the Mamba family of SSM LMs (130M–1.4B) on sequences ranging from 4–256 tokens. Our results show significantly higher rates of information loss on math-related tokens (e.g., numbers, variables), mentions of organization entities, and alternative dialects to Standard American English. We then examine the frequency that these tokens appear in Mamba’s pretraining data and find that less prevalent tokens tend to be the ones Mamba is most likely to forget. By identifying these patterns, our work provides clear direction for future research to develop methods that better control Mamba’s ability to retain important information.

When captioning an image, people describe objects in diverse ways, such as by using different terms and/or including details that are perceptually noteworthy to them. Descriptions can be especially unique across languages and cultures. Modern vision-language models (VLMs) gain understanding of images with text in different languages often through training on machine translations of English captions. However, this process relies on input content written from the perception of English speakers, leading to a perceptual bias. In this work, we outline a framework to address this bias. We specifically use a small amount of native speaker data, nearest-neighbor example guidance, and multimodal LLM reasoning to augment captions to better reflect descriptions in a target language. When adding the resulting rewrites to multilingual CLIP finetuning, we improve on German and Japanese text-image retrieval case studies (up to +3.5 mean recall, +4.4 on native vs. translation errors). We also propose a mechanism to build understanding of object description variation across languages, and offer insights into cross-dataset and cross-language generalization.

Factor analysis studies have shown that the primary dimensions of word meaning are Valence (V), Arousal (A), and Dominance (D). Existing lexicons such as the NRC VAD Lexicon, published in 2018, include VAD association ratings for words. Here, we present a complement to it, which has human ratings of valence, arousal, and dominance for ∼10k English Multiword Expressions (MWEs) and their constituent words. We also increase the coverage of unigrams, especially words that have become more common since 2018. In all, the new NRC VAD Lexicon v2 now has entries for ∼10k MWEs and ∼25k words, in addition to the entries in v1. We show that the associations are highly reliable. We use the lexicon to examine emotional characteristics of MWEs, including: 1. The degree to which MWEs (idioms, noun compounds, and verb particle constructions) exhibit strong emotionality; 2. The degree of emotional compositionality in MWEs. The lexicon enables a wide variety of research in NLP, Psychology, Public Health, Digital Humanities, and Social Sciences. The NRC VAD Lexicon v2 is freely available through the project web- page: http://saifmohammad.com/ WebPages/nrc-vad.html

Large Language Models (LLMs) often produce factoid hallucinations - plausible yet incorrect answers. A common mitigation strategy is model alignment, which improves factual accuracy by training on curated (factual, non-factual) pairs. However, this approach often relies on a stronger model (e.g., GPT-4) or an external knowledge base to assess factual correctness that may not always be accessible. Addressing this, we propose Atomic Consistency Preference Optimization (ACPO), a self-supervised preference-tuning method that enhances factual accuracy without external supervision. ACPO leverages atomic consistency signals (i.e., the agreement of individual facts across multiple stochastic responses) to identify high- and low-quality data pairs for model alignment. Despite being fully self-supervised, ACPO outperforms the strong supervised alignment baseline by 1.95 points averaged across Phi-3 and Llama3 on the LongFact and BioGen datasets, demonstrating its effectiveness in improving factual reliability without relying on external models or knowledge bases.

As large language models (LLMs) become increasingly embedded in our daily lives, evaluating their quality and reliability across diverse contexts has become essential. While comprehensive benchmarks exist for assessing LLM performance in English, there remains a significant gap in evaluation resources for other languages. Moreover, because most LLMs are trained primarily on data rooted in European and American cultures, they often lack familiarity with non-Western cultural contexts. To address this limitation, our study focuses on the Persian language and Iranian culture. We introduce 19 new evaluation datasets specifically designed to assess LLMs on topics such as Iranian law, Persian grammar, Persian idioms, and university entrance exams. Using these datasets, we benchmarked 41 prominent LLMs, aiming to bridge the existing cultural and linguistic evaluation gap in the field. The evaluation results are publicly available on our live leaderboard: https://huggingface.co/spaces/opll-org/Open-Persian-LLM-Leaderboard

In this work, we address the challenge of generalizable audio deepfake detection (ADD) across diverse speech synthesis paradigms—including conventional text-to-speech (TTS) systems and modern diffusion or flow-matching (FM) based generators. Prior work has mostly targeted individual synthesis families and often fails to generalize across paradigms due to overfitting to generation-specific artifacts. We hypothesize that synthetic speech, irrespective of its generative origin, leaves behind shared structural distortions in the embedding space that can be aligned through geometry-aware modeling. To this end, we propose RHYME, a unified detection framework that fuses utterance-level embeddings from diverse pretrained speech encoders using non-Euclidean projections. RHYME maps representations into hyperbolic and spherical manifolds—where hyperbolic geometry excels at modeling hierarchical generator families, and spherical projections capture angular, energy-invariant cues such as periodic vocoder artifacts. The fused representation is obtained via Riemannian barycentric averaging, enabling synthesis invariant alignment. RHYME outperforms individual PTMs and homogeneous fusion baselines, achieving top performance and setting new state-of-the-art in cross-paradigm ADD.

Predictive modeling of hospital patient data is challenging due to its structured format, irregular timing of measurements, and variation in data representation across institutions. While traditional models often struggle with such inconsistencies, Large Language Models (LLMs) offer a flexible alternative. In this work, we propose a method for verbalizing structured Electronic Health Records (EHRs) into a format suitable for LLMs and systematically examine how to include time-stamped clinical observations—such as lab tests and vital signs—from previous time points in the prompt. We study how different ways of structuring this temporal information affect predictive performance, and whether fine-tuning alone enables LLMs to effectively reason over such data. Evaluated on two real-world hospital datasets and MIMIC-IV, our approach achieves strong in-hospital and cross-hospital performance, laying the groundwork for more generalizable clinical modeling.

Realignment is a promising strategy to improve cross-lingual transfer in multilingual language models. However, empirical results are mixed and often unreliable, particularly for typologically distant or low-resource languages (LRLs) compared to English. Moreover, word realignment tools often rely on high-quality parallel data, which can be scarce or noisy for many LRLs. In this work, we conduct an extensive empirical study to investigate whether realignment truly benefits from using all available languages, or if strategically selected subsets can offer comparable or even improved cross-lingual transfer, and study the impact on LRLs. Our controlled experiments show that realignment can be particularly effective for LRLs and that using carefully selected, linguistically diverse subsets can match full multilingual alignment, and even outperform it for unseen LRLs. This indicates that effective realignment does not require exhaustive language coverage and can reduce data collection overhead, while remaining both efficient and robust when guided by informed language selection.

Document retrieval is a critical challenge in information retrieval systems, where the goal is to efficiently retrieve relevant documents in response to a given query. Dense retrieval methods, which utilize vector embeddings to represent semantic information, require effective indexing to ensure fast and accurate retrieval. Existing methods, such as MEVI, have attempted to address this by using hierarchical K-Means for clustering, but they often face limitations in computational efficiency and retrieval accuracy. In this paper, we introduce the Hildoc Index, a novel document indexing approach that leverages the Hilbert Curve to map document embeddings onto a one-dimensional space. This innovative representation facilitates efficient clustering using a 1D quantile-based algorithm, ensuring uniform partition sizes and preserving the inherent structure of the data. As a result, Hildoc Index not only reduces training complexity but also enhances retrieval accuracy and speed during inference. Our method can be seamlessly integrated into both dense retrieval systems and hybrid ensemble systems. Through comprehensive experiments on standard benchmarks like MSMARCO Passage and Natural Questions, we demonstrate that the Hildoc Index significantly outperforms the current state-of-the-art MEVI in terms of both retrieval speed and recall. These results underscore the Hildoc Index as a solution for fast and accurate dense document retrieval.

The rapid progress and widespread availability of text-to-image (T2I) generation models have heightened concerns about the misuse of AI-generated visuals, particularly in the context of misinformation campaigns. Existing AI-generated image detection (AGID) methods often overfit to known generators and falter on outputs from newer or unseen models. To systematically address this generalization gap, we introduce the Visual Counter Turing Test (VCT^2), a comprehensive benchmark of 166,000 images, comprising both real and synthetic prompt-image pairs produced by six state-of-the-art (SoTA) T2I systems: Stable Diffusion 2.1, SDXL, SD3 Medium, SD3.5 Large, DALL·E 3, and Midjourney 6. We curate two distinct subsets: COCO_AI, featuring structured captions from MS COCO, and Twitter_AI, containing narrative-style tweets from The New York Times. Under a unified zero-shot evaluation, we benchmark 17 leading AGID models and observe alarmingly low detection accuracy, 58% on COCO_AI and 58.34% on Twitter_AI. To transcend binary classification, we propose the Visual AI Index (V_AI), an interpretable, prompt-agnostic realism metric based on twelve low-level visual features, enabling us to quantify and rank the perceptual quality of generated outputs with greater nuance. Correlation analysis reveals a moderate inverse relationship between V_AI and detection accuracy: Pearson rho of -0.532 on COCO_AI and rho of -0.503 on Twitter_AI; suggesting that more visually realistic images tend to be harder to detect, a trend observed consistently across generators. We release COCO_AI and Twitter_AI to catalyze future advances in robust AGID and perceptual realism assessment.

Question answering in low-resource languages faces critical challenges when models encounter questions beyond their knowledge boundaries, often producing confident but incorrect answers. We propose Knowledge-Enhanced Reinforcement Learning for Question Answering (KERLQA), a novel approach that combines knowledge graph integration with reinforcement learning to enable principled abstention decisions. Unlike existing refusal-tuned methods that make binary decisions based solely on internal confidence, KERLQA implements a three-way decision process: answer with internal knowledge, answer with external knowledge assistance, or abstain. Using a composite reward function that jointly optimizes for correctness, appropriate abstention, and efficient knowledge utilization, we train policies via PPO and DPO with dynamic calibration for low-resource settings. Experiments on CommonsenseQA and OpenBookQA across English and four South African languages show KERLQA achieves improved F1 scores, with up to 6.2 point improvements in low-resource languages. Our analysis reveals that KERLQA reduces false positive abstention rates by 30% while expanding the boundary of answerable questions through external knowledge integration.

Clinical interventions often hinge on age: medications and procedures safe for adults may be harmful to children or ineffective for older adults. However, as language models are increasingly integrated into biomedical evidence synthesis workflows, it remains uncertain whether these systems preserve such crucial demographic distinctions. To address this gap, we evaluate how well state-of-the-art language models retain age-related information when generating abstractive summaries of biomedical studies.We construct DemogSummary, a novel age-stratified dataset of systematic review primary studies, covering child, adult, and older adult populations. We evaluate three prominent summarisation-capable LLMs, Qwen (open-source), Longformer (open-source) and GPT-4.1 Nano (proprietary), using both standard metrics and a newly proposed Demographic Salience Score (DSS), which quantifies age-related entity retention and hallucination.Our results reveal systematic disparities across models and age groups: demographic fidelity is lowest for adult-focused summaries, and underrepresented populations are more prone to hallucinations. These findings highlight the limitations of current LLMs in faithful and bias-free summarisation and point to the need for fairness-aware evaluation frameworks and summarisation pipelines in biomedical NLP.

The translation of idiomatic expressions often results in misunderstandings and inaccuracies, affecting everyday communication as well as machine translation systems. This paper introduces Idiom-aware Vietnamese Translation (IDiAT), a new framework for the evaluation of idiomatic translation for Vietnamese, along with state-of-the-art results for this task. We collect and curate a high-quality Vietnamese-English idiom set that serves as a resource for in-context learning (ICL). IDiAT’s evaluation benchmark includes both idiomatic and non-idiomatic text pairs to assess general translation quality and idiomatic translation performance. We leverage ICL in large language models to augment few-shot demonstrations with idiom and topic descriptions and consequently improve the translation accuracy. Empirical results demonstrate that our IDiAT-based ICL outperforms traditional supervised methods using only a few data samples. Multiple evaluations confirm the effectiveness of our proposed approach. Though focusing on the Vietnamese language, our approach advances idiomatic translation and contributes to the development of culturally aware translation systems, paving the way for future research in low-resource languages. The experimental materials are publicly available.

The recent development of LLMs is remarkable, but they still struggle to handle cross-lingual homographs effectively. This research focuses on the cross-lingual homographs between Japanese and Chinese—the spellings of the words are the same, but their meanings differ entirely between the two languages. We introduce a new benchmark dataset named Doppelganger-JC to evaluate the ability of LLMs to handle them correctly. We provide three kinds of tasks for evaluation: word meaning tasks, word meaning in context tasks, and translation tasks. Through the evaluation, we found that LLMs’ performance in understanding and using homographs is significantly inferior to that of humans. We pointed out the significant issue of homograph shortcut, which means that the model tends to preferentially interpret the cross-lingual homographs in its easy-to-understand language. We investigate the potential cause of this homograph shortcut from a linguistic perspective and pose that it is difficult for LLMs to recognize a word as a cross-lingual homograph, especially when it shares the same part-of-speech (POS) in both languages. The data and code is publicly available here: https://github.com/0017-alt/Doppelganger-JC.git.