ACL 2025 Papers

Language models (LMs) are indispensable tools shaping modern workflows, but their global effectiveness depends on understanding local socio-cultural contexts. To address this, we introduce SANSKRITI, a benchmark designed to evaluate language models’ comprehension of India’s rich cultural diversity. Comprising of 21,853 meticulously curated question-answer pairs spanning 28 states and 8 union territories, SANSKRITI is the largest dataset for testing Indian cultural knowledge. It covers sixteen key attributes of Indian culture namely rituals and ceremonies, history, tourism, cuisine, dance and music, costume, language, art, festivals, religion, medicine, transport, sports, nightlife and personalities, providing a comprehensive representation of India’s cultural tapestry. We evaluate SANSKRITI on leading Large Language Models (LLMs), Indic Language Models (ILMs), and Small Language Models(SLMs), revealing significant disparities in their ability to handle culturally nuanced queries, with many models struggling in region-specific contexts. By offering an extensive, culturally rich, and diverse dataset, SANSKRITI sets a new standard for assessing and improving the cultural understanding of LMs. We will share the dataset and findings publicly to support research on inclusive and culturally aware AI systems.

Speculative decoding (SD) has been demonstrated as an effective technique for lossless LLM inference acceleration.Retrieval-based SD methods, one kind of model-free method, have yielded promising speedup, but they often rely on single retrieval resources, inefficient retrieval methods, and are constrained to certain tasks. This paper presents a novel retrieval-based speculative decoding method that adapts the suffix automaton (SAM) for efficient and accurate draft generation by utilizing the generating text sequence and static text corpus. Unlike existing n-gram matching methods, SAM-Decoding finds the exact longest suffix match, achieving an average time complexity of O(1) per generation step of SAM update and suffix retrieval.It can also integrate with existing methods, adaptively selecting a draft generation strategy based on match length to generalize to broader domains. Extensive experiments on Spec-Bench show that our method is 18% faster than other retrieval-based SD methods. Additionally, when combined with advanced EAGLE-2, it provides an additional speedup of 3.28% – 11.13% across various-sized LLM backbones.

As Large Langue Models have been shown to memorize real-world facts, the need to update this knowledge in a controlled and efficient manner arises. Designed with these constraints in mind, Knowledge Editing (KE) approaches propose to alter specific facts in pretrained models. However, they have been shown to suffer from several limitations, including their lack of contextual robustness and their failure to generalize to logical implications related to the fact. To overcome these issues, we propose SAKE, a steering activation method that models a fact to be edited as a distribution rather than a single prompt. Leveraging Optimal Transport, SAKE alters the LLM behavior over a whole fact-related distribution, defined as paraphrases and logical implications. Several numerical experiments demonstrate the effectiveness of this method: SAKE is thus able to perform more robust edits than its existing counterparts.

Topic models are useful tools for discovering latent semantic structures in large textual corpora. Recent efforts have been oriented at incorporating contextual representations in topic modeling and have been shown to outperform classical topic models. These approaches are typically slow, volatile, and require heavy preprocessing for optimal results. We present Semantic Signal Separation (S3), a theory-driven topic modeling approach in neural embedding spaces. S3 conceptualizes topics as independent axes of semantic space and uncovers these by decomposing contextualized document embeddings using Independent Component Analysis. Our approach provides diverse and highly coherent topics, requires no preprocessing, and is demonstrated to be the fastest contextual topic model, being, on average, 4.5x faster than the runner-up BERTopic. We offer an implementation of S3, and all contextual baselines, in the Turftopic Python package.

Modeling latent representations in a hyperspherical space has proven effective for capturing directional similarities in high-dimensional text data, benefiting topic modeling. Variational autoencoder-based neural topic models (VAE-NTMs) commonly adopt the von Mises-Fisher prior to encode hyperspherical structure. However, VAE-NTMs often suffer from posterior collapse, where the KL divergence term in the objective function highly diminishes, leading to ineffective latent representations. To mitigate this issue while modeling hyperspherical structure in the latent space, we propose the Spherical Sliced Wasserstein Autoencoder for Topic Modeling (S2WTM). S2WTM employs a prior distribution supported on the unit hypersphere and leverages the Spherical Sliced-Wasserstein distance to align the aggregated posterior distribution with the prior. Experimental results demonstrate that S2WTM outperforms state-of-the-art topic models, generating more coherent and diverse topics while improving performance on downstream tasks.

Recent studies have demonstrated the effectiveness of LLM test-time scaling. However, existing approaches to incentivize LLMs’ deep thinking abilities generally require large-scale data or significant training efforts. Meanwhile, it remains unclear how to improve the thinking abilities of less powerful base models. In this work, we introduce S2R, an efficient framework that enhances LLM reasoning by teaching models to self-verify and self-correct during inference. Specifically, we first initialize LLMs with iterative self-verification and self-correction behaviors through supervised fine-tuning on carefully curated data. The self-verification and self-correction skills are then further strengthened by outcome-level and process-level reinforcement learning with minimized resource requirements. Our results demonstrate that, with only 3.1k behavior initialization samples, Qwen2.5-math-7B achieves an accuracy improvement from 51.0% to 81.6%, outperforming models trained on an equivalent amount of long-CoT distilled data. We also discuss the effect of different RL strategies on enhancing LLMs’ deep reasoning. Extensive experiments and analysis based on three base models across both in-domain and out-of-domain benchmarks validate the effectiveness of S2R.

Retrieval-Augmented Generation (RAG) systems combine external data retrieval with text generation and have become essential in applications requiring accurate and context-specific responses. However, their reliance on external data raises critical concerns about unauthorized collection and usage of personal information. To ensure compliance with data protection regulations like GDPR and detect improper use of data, we propose the Shadow RAG Auditing Data Provenance (S-RAG) framework. S-RAG enables users to determine whether their textual data has been utilized in RAG systems, even in black-box settings with no prior system knowledge. It is effective across open-source and closed-source RAG systems and resilient to defense strategies. Experiments demonstrate that S-RAG achieves an improvement in Accuracy by 19.9% (compared to the best baseline), while maintaining strong performance under adversarial defenses. Furthermore, we analyze how the auditor’s knowledge of the target system affects performance, offering practical insights for privacy-preserving AI systems. Our code is open-sourced online.

Knowledge editing emerges as a promising approach for updating target knowledge in Large Language Models (LLMs) in a timely manner, thereby preventing undesirable behaviors stemming from outdated, inaccurate, or incomplete knowledge. However, existing methods mainly focus on instance-level editing, which is prone to over-editing risk featuring knowledge degradation and general ability deterioration, due to redundant instance-specific modifications for knowledge. To mitigate the over-editing risk, we explore the rule-level editing problem that avoids case-by-case modification by generalizing rule-level knowledge to update rule-derived instances. We further construct a benchmark called RuleEdit for systematic evaluation on rule-level editing. Moreover, we propose a Rule-Transfer Editing (RTE) method to facilitate effective updates and generalizations of rule-level knowledge in LLMs. Experimental results highlight our significant improvements, with the enhancements of 28.1% in portability and 8.1% in average performance over the best-performing baselines for LLaMA-2-7B on RULEmix.

This paper introduces RuleArena, a novel and challenging benchmark designed to evaluate the ability of large language models (LLMs) to follow complex, real-world rules in reasoning. Covering three practical domains – airline baggage fees, NBA transactions, and tax regulations – RuleArena assesses LLMs’ proficiency in handling intricate natural language instructions that demand long-context understanding, logical reasoning, and accurate mathematical computation. Two key attributes distinguish RuleArena from traditional rule-based reasoning benchmarks: (1) it extends beyond standard first-order logic representations, and (2) it is grounded in authentic, practical scenarios, providing insights into the suitability and reliability of LLMs for real-world applications. Our findings reveal several notable limitations in LLMs: (1) they struggle to identify and apply the appropriate rules, frequently becoming confused by similar but distinct regulations, (2) they cannot consistently perform accurate mathematical computations, even when they correctly identify the relevant rules, and (3) in general, they perform poorly in the benchmark. We also observe a significant performance boost when LLMs are provided with external tools for oracle math and logic operations. These results highlight significant challenges and promising research directions in advancing LLMs’ rule-guided reasoning capabilities in real-life applications. Our codes and data are publicly available on https://github.com/skyriver-2000/rulearena.

The performance and usability of Large-Language Models (LLMs) are driving their use in explanation generation tasks. However, despite their widespread adoption, LLM explanations have been found to be unreliable, making it difficult for users to distinguish good from bad explanations. To address this issue, we present Rubrik’s CUBE–an education-inspired rubric and a dataset of 26k explanations, written and later quality-annotated using the rubric by both humans and six open- and closed-source LLMs. The CUBE dataset focuses on two reasoning and two language tasks, providing the necessary diversity for us to effectively test our proposed rubric. Using Rubrik, we find that explanations are influenced by both task and perceived difficulty. Low quality stems primarily from a lack of conciseness in LLM-generated explanations, rather than cohesion and word choice. The full dataset, rubric, and code are available at https://github.com/RubriksCube/rubriks_cube.

Large language models (LLMs) have achieved impressive performance but face high computational costs and latency, limiting their deployment in resource-constrained settings. In contrast, small-scale LLMs (SLMs) are more efficient yet struggle to capture evolving real-world knowledge. Retrieval-augmented generation (RAG) helps by integrating external knowledge, but imperfect retrieval can introduce distracting noise that misleads SLMs. We propose RoseRAG, a robust RAG framework for SLMs via Margin-aware Preference Optimization. RoseRAG employs multi-turn prompting for detailed reasoning, rejection sampling for high-quality explanations, and contrastive preference selection to refine responses by maximizing the likelihood gap between preferred and non-preferred outputs. By integrating these components into a margin-aware optimization process, RoseRAG robustly enhances the accuracy and reliability of SLMs for RAG applications. Extensive experiments on three open-domain question answering benchmarks indicate that our innovative RoseRAG surpasses state-of-the-art baselines significantly.

Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful outputs from the prefill-level lacks utilization of the model’s decoding outputs, leading to relatively lower effectiveness and robustness. 2) Rejecting potentially harmful outputs based on a single evaluation can significantly impair the model’s helpfulness. To address the above issues, we examine LLMs’ capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level. Our novel decoder-oriented, step-by-step defense architecture corrects the outputs of harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost safe decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model’s ability to discern hazardous information, maintaining its helpfulness compared to existing methods.

Large Language Models (LLMs) exhibit strong multilingual performance despite being predominantly trained on English-centric corpora. This raises a fundamental question: How do LLMs achieve such multilingual capabilities? Focusing on languages written in non-Roman scripts, we investigate the role of Romanization—the representation of non-Roman scripts using Roman characters—as a potential bridge in multilingual processing. Using mechanistic interpretability techniques, we analyze next-token generation and find that intermediate layers frequently represent target words in Romanized form before transitioning to native script, a phenomenon we term Latent Romanization. Further, through activation patching experiments, we demonstrate that LLMs encode semantic concepts similarly across native and Romanized scripts, suggesting a shared underlying representation. Additionally, for translation into non-Roman script languages, our findings reveal that when the target language is in Romanized form, its representations emerge earlier in the model’s layers compared to native script. These insights contribute to a deeper understanding of multilingual representation in LLMs and highlight the implicit role of Romanization in facilitating language transfer.

Human processing of idioms heavily depends on interpreting the surrounding context in which they appear. While large language models (LLMs) have achieved impressive performance on idiomaticity detection benchmarks, this success may be driven by reasoning shortcuts present in existing datasets. To address this, we introduce a novel, controlled contrastive dataset (DICE) specifically designed to assess whether LLMs can effectively leverage context to disambiguate idiomatic meanings. Furthermore, we investigate the influence of collocational frequency and sentence probability—proxies for human processing known to affect idiom resolution—on model performance. Our results show that LLMs frequently fail to resolve idiomaticity when it depends on contextual understanding, performing better on sentences deemed more likely by the model. Additionally, idiom frequency influences performance but does not guarantee accurate interpretation. Our findings emphasize the limitations of current models in grasping contextual meaning and highlight the need for more context-sensitive evaluation.

Role-playing agents (RPAs) are garnering increasing interests as a novel form of conversational AI. While previous research has predominantly concentrated on their ability to portray specified characters, we argue from a user-centered perspective that RPAs’ capability to advance the plot requires substantial improvements to deliver more engaging interaction. To bridge this gap, we propose RolePlot, a role-playing framework specifically designed to evaluate and enhance the plot-progression capabilities of RPAs. RolePlot begins by constructing a plot-progression dataset extended from human-written literary scripts and specially designed synthetic data, followed by narrative theory-driven manual annotation and automated labeling validated through human verification. We then exploit the over-parameterized embedding space of LLMs to detect a “trigger subspace” that identifies dialogue segments catalyzing plot transitions. When user’s inputs align with this subspace, we explicitly prompt RPAs to advance the plot. For evaluation, we simulate User-RPA interactions and track both the conversation longevity (measured in dialogue turns before disengagement) and users’ arousal levels across different stages. Empirically, our method improves RPAs’ capability to time plot developments, and more importantly, yielding a significant increase in conversation turns and sustained higher arousal levels, thereby confirming that users experience more immersive engagements.

Role-playing is important for Large Language Models (LLMs) to follow diverse instructions while maintaining role identity and the role’s pre-defined ability limits. Existing role-playing datasets mostly contribute to controlling role style and knowledge boundaries, but overlook role-playing in instruction-following scenarios. We introduce a fine-grained role-playing and instruction-following composite benchmark, named RoleMRC, including: (1) Multi-turn dialogues between ideal roles and humans, including free chats or discussions upon given passages; (2) Role-playing machine reading comprehension, involving response, refusal, and attempts according to passage answerability and role ability; (3) More complex scenarios with nested, multi-turn and prioritized instructions. The final RoleMRC features a 10.2k role profile meta-pool, 37.9k well-synthesized role-playing instructions, and 1.4k testing samples. We develop a pipeline to quantitatively evaluate the fine-grained role-playing and instruction-following capabilities of several mainstream LLMs, as well as models that are fine-tuned on our data. Moreover, cross-evaluation on external role-playing datasets confirms that models fine-tuned on RoleMRC enhances instruction-following without compromising general role-playing and reasoning capabilities. We also probe the neural-level activation maps of different capabilities over post-tuned LLMs. Access to our RoleMRC, RoleMRC-mix and Codes: https://github.com/LuJunru/RoleMRC.

Despite a growing literature finding that large language models (LLMs) exhibit demographic biases, reports with whom they align best are hard to generalize or even contradictory. In this work, we examine the alignment of LLMs with human annotations in five offensive language datasets, comprising approximately 220K annotations. While demographic traits, particularly race, influence alignment, these effects vary across datasets and are often entangled with other factors. Confounders introduced in the annotation process—such as document difficulty, annotator sensitivity, and within-group agreement—account for more variation in alignment patterns than demographic traits. Alignment increases with annotator sensitivity and group agreement, and decreases with document difficulty. Our results underscore the importance of multi-dataset analyses and confounder-aware methodologies in developing robust measures of demographic bias.

Embeddings as a Service (EaaS) is emerging as a crucial role in AI applications. Unfortunately, EaaS is vulnerable to model extraction attacks, highlighting the urgent need for copyright protection. Although some preliminary works propose applying embedding watermarks to protect EaaS, recent research reveals that these watermarks can be easily removed. Hence, it is crucial to inject robust watermarks resistant to watermark removal attacks. Existing watermarking methods typically inject a target embedding into embeddings through linear interpolation when the text contains triggers. However, this mechanism results in each watermarked embedding having the same component, which makes the watermark easy to identify and eliminate. Motivated by this, in this paper, we propose a novel embedding-specific watermarking (ESpeW) mechanism to offer robust copyright protection for EaaS. Our approach involves injecting unique, yet readily identifiable watermarks into each embedding. Watermarks inserted by ESpeW are designed to maintain a significant distance from one another and to avoid sharing common components, thus making it significantly more challenging to remove the watermarks. Moreover, ESpeW is minimally invasive, as it reduces the impact on embeddings to less than 1%, setting a new milestone in watermarking for EaaS. Extensive experiments on four popular datasets demonstrate that ESpeW can even watermark successfully against a highly aggressive removal strategy without sacrificing the quality of embeddings.

Anonymizing text that contains sensitive information is crucial for a wide range of applications. Existing techniques face the emerging challenges of the re-identification ability of large language models (LLMs), which have shown advanced capability in memorizing detailed information and reasoning over dispersed pieces of patterns to draw conclusions. When defending against LLM-based re-identification, anonymization could jeopardize the utility of the resulting anonymized data in downstream tasks. In general, the interaction between anonymization and data utility requires a deeper understanding within the context of LLMs. In this paper, we propose a framework composed of three key LLM-based components: a privacy evaluator, a utility evaluator and an optimization component, which work collaboratively to perform anonymization. Extensive experiments demonstrate that the proposed model outperforms existing baselines, showing robustness in reducing the risk of re-identification while preserving greater data utility in downstream tasks. We provide detailed studies on these core modules. To consider large-scale and real-time applications, we investigate the distillation of the anonymization capabilities into lightweight models. All of our code and datasets will be made publicly available at [Github URL].