ACL 2025 Papers

Encoder-only transformer models such as BERT offer a great performance-size tradeoff for retrieval and classification tasks with respect to larger decoder-only models. Despite being the workhorse of numerous production pipelines, there have been limited Pareto improvements to BERT since its release. In this paper, we introduce ModernBERT, bringing modern model optimizations to encoder-only models and representing a major Pareto improvement over older encoders. Trained on 2 trillion tokens with a native 8192 sequence length, ModernBERT models exhibit state-of-the-art results on a large pool of evaluations encompassing diverse classification tasks and both single and multi-vector retrieval on different domains (including code). In addition to strong downstream performance, ModernBERT is also the most speed and memory efficient encoder and is designed for inference on common GPUs.

Large language models (LLMs) excel in complex reasoning tasks, and distilling their reasoning capabilities into smaller models has shown promise. However, we uncover an interesting phenomenon, which we term the Small Model Learnability Gap: small models (3B parameters) do not consistently benefit from long chain-of-thought (CoT) reasoning or distillation from larger models. Instead, they perform better when fine-tuned on shorter, simpler reasoning chains that better align with their intrinsic learning capacity. To address this, we propose Mix Distillation, a simple yet effective strategy that balances reasoning complexity by combining long and short CoT examples or reasoning from both larger and smaller models. Our experiments demonstrate that Mix Distillation significantly improves small model reasoning performance compared to training on either data alone. These findings highlight the limitations of direct strong model distillation and underscore the importance of adapting reasoning complexity for effective reasoning capability transfer.

Augmenting large language models (LLMs) with external context significantly improves their performance in natural language processing (NLP) tasks. However, LLMs struggle to answer queries reliably when the provided context lacks information, often resorting to ungrounded speculation or internal knowledge. Groundedness – generating responses strictly supported by the context – is essential for ensuring factual consistency and trustworthiness. This study focuses on detecting whether a given query is grounded in a document provided in context before the costly answer generation by LLMs. Such a detection mechanism can significantly reduce both inference time and resource consumption. We show that lightweight, task-specific encoder models such as RoBERTa and NomicBERT, fine-tuned on curated datasets, can achieve accuracy comparable to state-of-the-art LLMs, such as Llama3 8B and GPT4o, in groundedness detection while reducing inference latency by orders of magnitude.

Recent remarkable advances in Large Language Models (LLMs) have led to innovations in various domains such as education, healthcare, and finance, while also raising serious concerns that they can be easily misused for malicious purposes. Most previous research has focused primarily on observing how jailbreak attack techniques bypass safety mechanisms like Reinforcement Learning through Human Feedback (RLHF). However, whether there are neurons within LLMs that directly govern aggression has not been sufficiently investigated. To fill this gap, this study identifies specific neurons (“aggression neurons”) closely related to the expression of aggression and systematically analyzes how manipulating them affects the model’s overall aggression. Specifically, using a large-scale synthetic text corpus (aggressive and non-aggressive), we measure the activation frequency of each neuron, then apply masking and activation techniques to quantitatively evaluate changes in aggression by layer and by manipulation ratio. Experimental results show that, in all models, manipulating only a small number of neurons can increase aggression by up to 33%, and the effect is even more extreme when aggression neurons are concentrated in certain layers. Moreover, even models of the same scale exhibit nonlinear changes in aggression patterns, suggesting that simple external safety measures alone may not be sufficient for complete defense.

Large Language Models (LLMs) have shown exciting performance in listwise passage ranking. Due to the limited input length, existing methods often adopt the sliding window strategy. Such a strategy, though effective, is inefficient as it involves repetitive and serialized processing, which usually re-evaluates relevant passages multiple times. As a result, it incurs redundant API costs, which are proportional to the number of inference tokens. The development of long-context LLMs enables the full ranking of all passages within a single inference, avoiding redundant API costs. In this paper, we conduct a comprehensive study of long-context LLMs for ranking tasks in terms of efficiency and effectiveness. Surprisingly, our experiments reveal that full ranking with long-context LLMs can deliver superior performance in the supervised fine-tuning setting with a huge efficiency improvement. Furthermore, we identify two limitations of fine-tuning the full ranking model based on existing methods: (1) sliding window strategy fails to produce a full ranking list as a training label, and (2) the language modeling loss cannot emphasize top-ranked passage IDs in the label. To alleviate these issues, we propose a new complete listwise label construction approach and a novel importance-aware learning objective for full ranking. Experiments show the superior performance of our method over baselines.

We present an end-to-end framework for generating synthetic users for evaluating interactive agents designed to encourage positive behavior changes, such as in health and lifestyle coaching. The synthetic users are grounded in health and lifestyle conditions, specifically sleep and diabetes management in this study, to ensure realistic interactions with the health coaching agent. Synthetic users are created in two stages: first, structured data are generated grounded in real-world health and lifestyle factors in addition to basic demographics and behavioral attributes; second, full profiles of the synthetic users are developed conditioned on the structured data. Interactions between synthetic users and the coaching agent are simulated using generative agent-based models such as Concordia, or directly by prompting a language model. Using two independently-developed agents for sleep and diabetes coaching as case studies, the validity of this framework is demonstrated by analyzing the coaching agent’s understanding of the synthetic users’ needs and challenges. Finally, through multiple blinded evaluations of user-coach interactions by human experts, we demonstrate that our synthetic users with health and behavioral attributes more accurately portray real human users with the same attributes, compared to generic synthetic users not grounded in such attributes. The proposed framework lays the foundation for efficient development of conversational agents through extensive, realistic, and grounded simulated interactions.

We introduce *Slam*, a recipe for training high-quality Speech Language Models (SLMs) on a single academic GPU in 24 hours. We do so through empirical analysis of model initialisation and architecture, synthetic training data, preference optimisation with synthetic data and tweaking all other components. We empirically demonstrate that this training recipe also scales well with more compute getting results on par with leading SLMs in a fraction of the compute cost. We hope these insights will make SLM training and research more accessible. In the context of SLM scaling laws, our results far outperform predicted compute optimal performance, giving an optimistic view to SLM feasibility. See code, data, models, samples - https://pages.cs.huji.ac.il/adiyoss-lab/slamming .

Large language models (LLMs) have demonstrated exceptional capabilities across a wide range of tasks, from text generation to complex problem-solving. LLM APIs provide easy access to these models by streamlining deployment and usage. Combining LLMs with complementary strengths has been shown to yield substantial performance gains over a monolithic LLM. However, invoking a fixed set of LLM APIs for each query incurs higher API costs and increased inference latency. To address these limitations, we propose SkyLLM, a system composed of a set of estimators and an API selector, which federates multiple LLM APIs and dynamically assigns a non-empty subset of these APIs to each query prior to inference under cost and latency budgets. The selected subset consists of either a single LLM or multiple LLMs. A single LLM efficiently handles simple queries at low cost, whereas multiple LLMs are employed for more complex queries to overcome performance limitations. We evaluate SkyLLM against individual LLMs and representative ensemble LLM methods from the literature. SkyLLM achieves the highest accuracy under a high budget. It can also be cost-effective, matching the most accurate individual LLM while cutting costs by 67.8%.

As language models evolve to tackle complex, multifaceted tasks, their evaluation must adapt to capture this intricacy. A granular, skill-specific understanding of model capabilities can empower researchers to make informed model development plans. In this paper, we introduce SkillVerse, an unsupervised tree-structured diagnosis framework for understanding model proficiency in specific abilities. With LLM as a judge, SkillVerse first critiques the model responses, and then organizes them into a hierarchical structure termed dendrogram. Given proficiency at arbitrary levels of granularity, SkillVerse is flexible to produce insights of behaviors of modern large models. We also demonstrate its efficacy in two downstream tasks: 1) improving model in-context learning by 25% using a tree-search algorithm to select more informative few-shot demonstrations, and 2) accurately predicting new model weaknesses with a 55% success rate, 22% higher than without SkillVerse.

Large Language Models (LLMs) are increasingly used to assess NLP tasks due to their ability to generate human-like judgments. Single LLMs were used initially, however, recent work suggests using multiple LLMs as judges yields improved performance. An important step in exploiting multiple judgements is the combination stage, aggregation. Existing methods in NLP either assign equal weight to all LLM judgments or are designed for specific tasks such as hallucination detection. This work focuses on aggregating predictions from multiple systems where no reference labels are available. A new method called SkillAggregation is proposed, which learns to combine estimates from LLM judges without needing additional data or ground truth. It extends the Crowdlayer aggregation method, developed for image classification, to exploit the judge estimates during inference. The approach is compared to a range of standard aggregation methods on HaluEval-Dialogue, TruthfulQA and Chatbot Arena tasks. SkillAggregation outperforms Crowdlayer on all tasks, and yields the best performance over all approaches on the majority of tasks.

Recently, language models (LMs) have produced excellent results in many natural language processing (NLP) tasks. However, their effectiveness is highly dependent on available pre-training resources, which is particularly challenging for low-resource languages such as Sinhala. Furthermore, the scarcity of benchmarks to evaluate LMs is also a major concern for low-resource languages. In this paper, we address these two challenges for Sinhala by (i) collecting the largest monolingual corpus for Sinhala, (ii) training multiple LMs on this corpus and (iii) compiling the first Sinhala NLP benchmark (Sinhala-GLUE) and evaluating LMs on it. We show the Sinhala LMs trained in this paper outperform the popular multilingual LMs, such as XLM-R and existing Sinhala LMs in downstream NLP tasks. All the trained LMs are publicly available. We also make Sinhala-GLUE publicly available as a public leaderboard, and we hope that it will enable further advancements in developing and evaluating LMs for Sinhala.

Document Image Machine Translation (DIMT) aims to translate text within document images, facing generalization challenges due to limited training data and the complex interplay between visual and textual information. To address these challenges, we introduce M4Doc, a novel single-to-mix Modality alignment framework leveraging Multimodal Large Language Models (MLLMs). M4Doc aligns an imageonly encoder with the multimodal representations of an MLLM, pre-trained on large-scale document image datasets. This alignment enables a lightweight DIMT model to learn crucial visual-textual correlations during training. During inference, M4Doc bypasses the MLLM, maintaining computational efficiency while benefiting from its multimodal knowledge. Comprehensive experiments demonstrate substantial improvements in translation quality, especially in cross-domain generalization and challenging document image scenarios. The code will be released upon acceptance.

Simultaneous speech translation (SST) outputs translations in parallel with streaming speech input, balancing translation quality and latency. While large language models (LLMs) have been extended to handle the speech modality, streaming remains challenging as speech is pre-pended as a prompt for the entire generation process. To unlock LLM streaming capability, this paper proposes SimulS2S-LLM, which trains speech LLMs offline and employs a test-time policy to guide simultaneous inference. SimulS2S-LLM alleviates the mismatch between training and inference by extracting boundary-aware speech prompts that allows it to be better matched with text input data. SimulS2S-LLM achieves simultaneous speech-to-speech translation (Simul-S2ST) by predicting discrete output speech tokens and then synthesising output speech using a pre-trained vocoder. An incremental beam search is designed to expand the search space of speech token prediction without increasing latency. Experiments on the CVSS speech data show that SimulS2S-LLM offers a better translation quality-latency trade-off than existing methods that use the same training data, such as improving ASR-BLEU scores by 3 points at similar latency.

Recent advancements in large language models (LLMs) have shown impressive versatility across various tasks. To eliminate their hallucinations, retrieval-augmented generation (RAG) has emerged as a powerful approach, leveraging external knowledge sources like knowledge graphs (KGs). In this paper, we study the task of KG-driven RAG and propose a novel Similar Graph Enhanced Retrieval-Augmented Generation (SimGRAG) method. It effectively addresses the challenge of aligning query texts and KG structures through a two-stage process: (1) query-to-pattern, which uses an LLM to transform queries into a desired graph pattern, and (2) pattern-to-subgraph, which quantifies the alignment between the pattern and candidate subgraphs using a graph semantic distance (GSD) metric. We also develop an optimized retrieval algorithm that efficiently identifies the top-k subgraphs within 1-second on a 10-million-scale KG. Extensive experiments show that SimGRAG outperforms state-of-the-art KG-driven RAG methods in both question answering and fact verification. Our code is available at https://github.com/YZ-Cai/SimGRAG.

To meet the demands of content moderation, online platforms have resorted to automated systems. Newer forms of real-time engagement (e.g., users commenting on live streams) on platforms like Twitch exert additional pressures on the latency expected of such moderation systems. Despite their prevalence, relatively little is known about the effectiveness of these systems. In this paper, we conduct an audit of Twitch’s automated moderation tool (AutoMod) to investigate its effectiveness in flagging hateful content. For our audit, we create streaming accounts to act as siloed test beds, and interface with the live chat using Twitch’s APIs to send over 107,000 comments collated from 4 datasets. We measure AutoMod‘s accuracy in flagging blatantly hateful content containing misogyny, racism, ableism and homophobia. Our experiments reveal that a large fraction of hateful messages, up to 94% on some datasets, bypass moderation. Contextual addition of slurs to these messages results in 100% removal, revealing AutoMod‘s reliance on slurs as a hate signal. We also find that contrary to Twitch’s community guidelines, AutoMod blocks up to 89.5% of benign examples that use sensitive words in pedagogical or empowering contexts. Overall, our audit points to large gaps in AutoMod‘s capabilities and underscores the importance for such systems to understand context effectively.

A persistent challenge in sign language video processing, including the task of sign language to written language translation, is how we train efficient model given the nature of videos. Informed by the nature and linguistics of signed languages, our proposed method focuses on just the most relevant parts in a signing video: the face, hands and body posture of the signer. However, instead of using pose estimation coordinates from off-the-shelf pose tracking models, which have inconsistent performance for hands and faces, we propose to learn the complex handshapes and rich facial expressions of sign languages in a self-supervised fashion. Our approach is based on learning from individual frames (rather than video sequences) and is therefore much more efficient than prior work on sign language pre-training. Compared to a recent model trained on publicly avaiable data that established a new state of the art in sign language translation on the How2Sign dataset, our approach yields similar translation performance, using less than 3% of the compute.

Deaf and Hard-of-Hearing (DHH) users increasingly utilize Large Language Models (LLMs), yet face significant challenges due to these models’ limited understanding of sign language grammar, multimodal sign inputs, and Deaf cultural contexts. Further, current approaches that try to address these limitations, frequently reduce sign language processing (SLP) to traditional translation tasks, neglecting the multimodal and linguistic complexity inherent in signed languages. In this paper, we present an empirical investigation informed by learning theory into natively integrating sign language support within LLMs, directly addressing the documented needs of DHH users. We introduce the first text-based and multimodal LLMs capable of sign language processing called SignAlignLM, and propose new prompting and fine-tuning strategies incorporating sign linguistic rules and conventions. We show that LLMs can be generalized interfaces for both spoken and signed languages if trained with a multitasking paradigm. Our code and model checkpoints are open-source.

Data visualizations, such as bar charts and histograms, are essential for analyzing and exploring data, enabling the effective communication of insights. While existing methods have been proposed to translate natural language descriptions into visualization queries, they focus solely on spoken languages, overlooking sign languages, which comprise about 200 variants used by 70 million Deaf and Hard-of-Hearing (DHH) individuals. To fill this gap, this paper proposes Sign2Vis, a sign language interface that enables the DHH community to engage more fully with data analysis. We first construct a paired dataset that includes sign language pose videos and their corresponding visualization queries. Using this dataset, we evaluate a variety of models, including both pipeline-based and end-to-end approaches. Extensive experiments, along with a user study involving 15 participants, demonstrate the effectiveness of Sign2Vis. Finally, we share key insights from our evaluation and highlight the need for more accessible and user-centered tools to support the DHH community in interactive data analytics.

Often, the needs and visual abilities differ between the annotator group and the end user group. Generating detailed diagram descriptions for blind and low-vision (BLV) users is one such challenging domain. Sighted annotators could describe visuals with ease, but existing studies have shown that direct generations by them are costly, bias-prone, and somewhat lacking by BLV standards. In this study, we ask sighted individuals to assess—rather than produce—diagram descriptions generated by vision-language models (VLM) that have been guided with latent supervision via a multi-pass inference. The sighted assessments prove effective and useful to professional educators who are themselves BLV and teach visually impaired learners. We release Sightation, a collection of diagram description datasets spanning 5k diagrams and 137k samples for completion, preference, retrieval, question answering, and reasoning training purposes and demonstrate their fine-tuning potential in various downstream tasks.