NAACL 2025 论文列表

Instruction following is a core capability of modern Large language models (LLMs), making evaluating this capability essential to understanding these models. The Instruction Following Evaluation (IFEval) benchmark from the literature does this using objective criteria, offering a measure of LLM performance without subjective AI or human judgement. However, it only includes English instructions, limiting its ability to assess LLMs in other languages.We propose the Multilingual Instruction Following Evaluation (M-IFEval) benchmark, expanding the evaluation to French, Japanese, and Spanish, with both general and language-specific instructions. Applying this benchmark to 8 state-of-the-art LLMs, we find that benchmark performance across languages and instruction types can vary widely, underscoring the importance of a multilingual benchmark for evaluating LLMs in a diverse cultural context.

Large language models (LLMs) have demonstrated impressive capabilities in handling long contexts, but challenges remain in capturing relational knowledge spread far apart within text. Connecting long-distance knowledge is important for solving tasks as the context length increases: imagine reading a lengthy detective novel where seemingly trivial information introduced early on often becomes essential during the climactic reveal of the culprit. In this study, we expose the ”Lost in the Distance” phenomenon, where LLM performance of capturing the relational knowledge degrades significantly when the relational knowledge is separated by noise, i.e., unrelated sentences to solve a task. Specifically, we design an experiment in which we insert artificial noise between two related elements and observe model performance as the distance between them increases. Our findings show that while LLMs can handle edge noise with little impact, their ability to reason about distant relationships declines sharply as the intervening noise grows. These findings are consistent in both forward-looking prediction and backward-looking prediction settings. We validate this across various models (GPT-4, Gemini-1.5-pro, GPT-4o-mini, Gemini-1.5-flash, Claude-3.5-Sonnet) and tasks (causal reasoning and knowledge extraction). These results reveal a significant limitation in how LLMs process relational knowledge over long contexts. We release our code and data to support further research.

The proliferation of large language models (LLMs) in generating content raises concerns about text copyright. Watermarking methods, particularly logit-based approaches, embed imperceptible identifiers into text to address these challenges. However, the widespread usage of watermarking across diverse LLMs has led to an inevitable issue known as watermark collision during common tasks, such as paraphrasing or translation.In this paper, we introduce watermark collision as a novel and general philosophy for watermark attacks, aimed at enhancing attack performance on top of any other attacking methods. We also provide a comprehensive demonstration that watermark collision poses a threat to all logit-based watermark algorithms, impacting not only specific attack scenarios but also downstream applications.

In the recent past, a popular way of evaluating natural language understanding (NLU), was to consider a model’s ability to perform natural language inference (NLI) tasks. In this paper, we investigate if NLI tasks, that are rarely used for LLM evaluation, can still be informative for evaluating LLMs. Focusing on five different NLI benchmarks across six models of different scales, we investigate if they are able to discriminate models of different size and quality and how their accuracies develop during training. Furthermore, we investigate the extent to which the softmax distributions of models align with human distributions in cases where statements are ambiguous or vague. Overall, our results paint a positive picture for the NLI tasks: we find that they are able to discriminate well between models at various stages of training, yet are not (all) saturated. Furthermore, we find that while the similarity of model distributions with human label distributions increases with scale, it is still much higher than the similarity between two populations of humans, making it a potentially interesting statistic to consider.

Accelerating inference in Large Language Models (LLMs) is critical for real-time interactions, as they have been widely incorporated into real-world services. Speculative decoding, a fully algorithmic solution, has gained attention for improving inference speed by drafting and verifying tokens, thereby generating multiple tokens in a single forward pass. However, current drafting strategies usually require significant fine-tuning or have inconsistent performance across tasks. To address these challenges, we propose Hierarchy Drafting (HD), a novel lossless drafting approach that organizes various token sources into multiple databases in a hierarchical framework based on temporal locality. In the drafting step, HD sequentially accesses multiple databases to obtain draft tokens from the highest to the lowest locality, ensuring consistent acceleration across diverse tasks and minimizing drafting latency. Our experiments on Spec-Bench using LLMs with 7B and 13B parameters demonstrate that HD outperforms existing database drafting methods, achieving robust inference speedups across model sizes, tasks, and temperatures.

Large Language Models (LLMs), exemplified by Claude and LLama, have exhibited impressive proficiency in tackling a myriad of Natural Language Processing (NLP) tasks. Yet, in pursuit of the ambitious goal of attaining Artificial General Intelligence (AGI), there remains ample room for enhancing LLM capabilities. Chief among these is the pressing need to bolster long-context comprehension. Numerous real-world scenarios demand LLMs to adeptly reason across extended contexts, such as multi-turn dialogues or agent workflow. Hence, recent advancements have been dedicated to stretching the upper bounds of long-context comprehension, with models like Claude 3 accommodating up to 200k tokens, employing various techniques to achieve this feat. Aligned with this progression, we propose a leaderboard LongLeader that seeks to comprehensively assess different long-context comprehension abilities of diverse LLMs and context length extension strategies across meticulously selected benchmarks. Specifically, we aim to address the following questions: 1) Do LLMs genuinely deliver the long-context proficiency they purport? 2) Which benchmarks offer reliable metrics for evaluating long-context comprehension? 3) What technical strategies prove effective in extending the understanding of longer contexts? We streamline the evaluation process for LLMs on the benchmarks, offering open-source access to the benchmarks and maintaining a dedicated website for leaderboards. We will continuously curate new datasets and update models to the leaderboards.

The k-nearest-neighbor language model (kNN-LM), one of the retrieval-augmented language models, improves the perplexity for given text by directly accessing a large datastore built from any text data during inference.A widely held hypothesis for the success of kNN-LM is that its explicit memory, i.e., the datastore, enhances predictions for long-tail phenomena.However, prior works have primarily shown its ability to retrieve long-tail contexts, leaving the model’s performance remain underexplored in estimating the probabilities of long-tail target tokens during inference.In this paper, we investigate the behavior of kNN-LM on low-frequency tokens, examining prediction probability, retrieval accuracy, and token distribution in the datastore.Our experimental results reveal that kNN-LM does not improve prediction performance for low-frequency tokens but mainly benefits high-frequency tokens regardless of long-tail contexts in the datastore.

Effective organization of in-context learning (ICL) demonstrations is key to improving the quality of large language model (LLM) responses. To create better sample-label pairs that instruct LLM understanding, we introduce logit separability, a criterion to assess the clarity of both samples and class-related words at the logit level. This facilitates the optimization of sample and label selection, enhancing the precision of information provided in ICL demonstrations. Additionally, we find that incorporating multiple class-related words for each sample, rather than relying on a single class name, improves performance by offering a broader range of label information. Building on these insights, we propose LICL, a logit separability-based method that jointly organizes samples and integrates multiple class-related words into each sample-label pair. Evaluations across seven classification datasets show that this approach significantly improves ICL performance by providing clearer instructions and richer label information.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but their performance in complex logical reasoning tasks remains unsatisfactory. Although some prompting methods, such as Chain-of-Thought, can improve the reasoning ability of LLMs to some extent, they suffer from an unfaithful issue where derived conclusions may not align with the generated reasoning chain. To address this issue, some studies employ the approach of propositional logic to further enhance logical reasoning abilities of LLMs. However, the potential omissions in the extraction of logical expressions in these methods can cause information loss in the logical reasoning process, thereby generating incorrect results. To this end, we propose Logic-of-Thought (LoT) prompting which employs propositional logic to generate expanded logical information descriptions and utilizes them as an additional augmentation to original contexts, thereby ensuring information completeness and enhancing logical reasoning ability. LoT is orthogonal to existing prompting methods and can be seamlessly integrated with them. Extensive experiments demonstrate that LoT boosts the performance of various prompting methods with a striking margin across five logical reasoning tasks. In particular, LoT enhances Chain-of-Thought’s performance on the ReClor dataset by +4.35%, improves Chain-of-Thought with Self-Consistency’s performance on the RuleTaker dataset by +3.52%, and boosts performance of Tree-of-Thoughts on the ProofWriter dataset by +8%.

Currently, large language models (LLMs) have achieved impressive performance in natural language processing tasks. However, LLMs still exhibit many hallucinations when analyzing system logs, which is due to the implicit knowledge and rules in logs that LLMs cannot capture. Based on this, we propose LogRules, a lightweight log analysis framework that generates and utilizes rules through LLMs. LogRules consists of three stages: an induction stage, an alignment stage, and a reasoning stage. Firstly, in the induction stage, an strong LLM (e.g., GPT-4o-mini) is tasked with generating a series of rules related to logs, which are then validated on the training set. When the rules are confirmed to produce correct reasoning results, they are added to a rule repository. Secondly, considering that the LLMs with small size (≈8B parameters) still face challenges in utilizing rules, we design an alignment method based on rule-case contrastive preference optimization (CPO) to effectively enhance the rule reasoning capabilities of these LLMs. Finally, in the reasoning stage, the LLM constructs prompt using the rule repository and performs log analysis on the test set. Experiments show that LogRules outperforms LLM-based methods in log parsing and anomaly detection tasks, and achieves better performance compared to case-based methods.

In recent years, the use of prompts to guide the output of Large Language Models have increased dramatically. However, even the best of experts struggle to choose the correct words to stitch up a prompt for the desired task. To solve this, LLM driven prompt optimization emerged as an important problem. Existing prompt optimization methods optimize a prompt globally, where in all the prompt tokens have to be optimized over a large vocabulary while solving a complex task. The large optimization space (tokens) leads to insufficient guidance for a better prompt. In this work, we introduce Local Prompt Optimization (LPO) that integrates with any general automatic prompt engineering method. We identify the optimization tokens in a prompt and nudge the LLM to focus only on those tokens in its optimization step. We observe remarkable performance improvements on Math Reasoning (GSM8k and MultiArith) and BIG-bench Hard benchmarks across various automatic prompt engineering methods. Further, we show that LPO converges to the optimal prompt faster than global methods.

Large Language Models (LLMs) have demonstrated remarkable success as general-purpose task solvers across various fields. However, their capabilities remain limited when addressing domain-specific problems, particularly in downstream NLP tasks. Research has shown that models fine-tuned on instruction-based downstream NLP datasets outperform those that are not fine-tuned. While most efforts in this area have primarily focused on resource-rich languages like English and broad domains, little attention has been given to multilingual settings and specific domains. To address this gap, this study focuses on developing a specialized LLM, LlamaLens, for analyzing news and social media content in a multilingual context. To the best of our knowledge, this is the first attempt to tackle both domain specificity and multilinguality, with a particular focus on news and social media. Our experimental setup includes 18 tasks, represented by 52 datasets covering Arabic, English, and Hindi. We demonstrate that LlamaLens outperforms the current state-of-the-art (SOTA) on 23 testing sets, and achieves comparable performance on 8 sets. We make the models and resources publicly available for the research community (https://huggingface.co/QCRI).

Adverse Drug Reactions (ADRs) from psychiatric medications are the leading cause of hospitalizations among mental health patients. With healthcare systems and online communities facing limitations in resolving ADR-related issues, Large Language Models (LLMs) have the potential to fill this gap. Despite the increasing capabilities of LLMs, past research has not explored their capabilities in detecting ADRs related to psychiatric medications or in providing effective harm reduction strategies. To address this, we introduce the **Psych-ADR** benchmark and the **A**dverse **D**rug Reaction **R**esponse **A**ssessment (**ADRA**) framework to systematically evaluate LLM performance in detecting ADR expressions and delivering expert-aligned mitigation strategies. Our analyses show that LLMs struggle with understanding the nuances of ADRs and differentiating between types of ADRs. While LLMs align with experts in terms of expressed emotions and tone of the text, their responses are more complex, harder to read, and only 70.86% aligned with expert strategies. Furthermore, they provide less actionable advice by a margin of 12.32% on average. Our work provides a comprehensive benchmark and evaluation framework for assessing LLMs in strategy-driven tasks within high-risk domains.

Synthetic data generation has become an increasingly popular way of training models without the need for large, manually labeled datasets. For tasks like text embedding, synthetic data offers diverse and scalable training examples, significantly reducing the cost of human annotation. However, most current approaches rely heavily on proprietary models like GPT-4, which are expensive and inefficient for generating large-scale embedding data. In this paper, we introduce SPEED, a framework that aligns open-source small models (8B) to efficiently generate large-scale synthetic embedding data. Through supervised fine-tuning, preference optimization, and self-improvement, SPEED enables small open-source models to produce high-quality data. Remarkably, SPEED uses only less than 1/10 of the GPT API calls, outperforming the state-of-the-art embedding model E5_mistral when both are trained solely on their synthetic data. Using this efficient generator, we conduct a comprehensive study on how various factors within the alignment pipeline impact data quality and reveal the scaling law for synthetic embedding data. Our codes and models are released in https://github.com/haon-chen/SPEED.

We introduce LiteWebAgent, an open-source suite for VLM-based web agent applications. Our framework addresses a critical gap in the web agent ecosystem with a production-ready solution that combines minimal serverless backend configuration, intuitive user and browser interfaces, and extensible research capabilities in agent planning, memory, and tree search. For the core LiteWebAgent agent framework, we implemented a simple yet effective baseline using recursive function calling, providing with decoupled action generation and action grounding. In addition, we integrate advanced research components such as agent planning, agent workflow memory, and tree search in a modular and extensible manner. We then integrate the LiteWebAgent agent framework with frontend and backend as deployed systems in two formats: (1) a production Vercel-based web application, which provides users with an agent-controlled remote browser, (2) a Chrome extension leveraging LiteWebAgent’s API to control an existing Chrome browser via CDP (Chrome DevTools Protocol). The LiteWebAgent framework is available at https://github.com/PathOnAI/LiteWebAgent, with deployed frontend at https://lite-web-agent.vercel.app/.

Understanding how linguistic knowledge is encoded in language models is crucial for improving their generalisation capabilities. In this paper, we investigate the processing of morphosyntactic phenomena, by leveraging a recently proposed method for probing language models via Shapley Head Values (SHVs). Using the English language BLiMP dataset, we test our approach on two widely used models, BERT and RoBERTa, and compare how linguistic constructions such as anaphor agreement and filler-gap dependencies are handled. Through quantitative pruning and qualitative clustering analysis, we demonstrate that attention heads responsible for processing related linguistic phenomena cluster together. Our results show that SHV-based attributions reveal distinct patterns across both models, providing insights into how language models organize and process linguistic information. These findings support the hypothesis that language models learn subnetworks corresponding to linguistic theory, with potential implications for cross-linguistic model analysis and interpretability in Natural Language Processing (NLP).

The semi-supervised learning (SSL) strategy in lightweight models requires reducing annotated samples and facilitating cost-effective inference. However, the constraint on model parameters, imposed by the scarcity of training labels, limits the SSL performance. In this paper, we introduce PS-NET, a novel framework tailored for semi-supervised text mining with lightweight models. PS-NET incorporates online distillation to train lightweight student models by imitating the Teacher model. It also integrates an ensemble of student peers that collaboratively instruct each other. Additionally, PS-NET implements a constant adversarial perturbation schema to further self-augmentation by progressive generalizing. Our PS-NET, equipped with a 2-layer distilled BERT, exhibits notable performance enhancements over SOTA lightweight SSL frameworks of FLiText and Disco in SSL text classification with extremely rare labelled data.

As large language models (LLMs) continue to evolve, leaderboards play a significant role in steering their development. Existing leaderboards often prioritize model capabilities while overlooking safety concerns, leaving a significant gap in responsible AI development. To address this gap, we introduce Libra-Leaderboard, a comprehensive framework designed to rank LLMs through a balanced evaluation of performance and safety. Combining a dynamic leaderboard with an interactive LLM arena, Libra-Leaderboard encourages the joint optimization of capability and safety. Unlike traditional approaches that average performance and safety metrics, Libra-Leaderboard uses a distance-to-optimal-score method to calculate the overall rankings. This approach incentivizes models to achieve a balance rather than excelling in one dimension at the expense of some other ones. In the first release, Libra-Leaderboard evaluates 26 mainstream LLMs from 14 leading organizations, identifying critical safety challenges even in state-of-the-art models.

Recent advancements in code completion models have primarily focused on local file contexts. However, these studies do not fully capture the complexity of real-world software development, which often requires the use of rapidly-evolving public libraries. To address this gap, we introduce LibEvolutionEval, a comprehensive study that emphasizes the need to understand library evolution to perform accurate in-line code completions. LibEvolutionEvaloffers a version-specific code-completion task across eight libraries as they evolve over the years, along with an in-depth analysis of the evolution of two widely used and well-maintained public libraries: PyTorch and Matplotlib. We evaluate several popular models and find that public library evolution significantly affects their performance. To mitigate this, we explored how retrieving version-specific library documentation and prompt-based techniques can enhance model capability in dealing with these fast-evolving packages. This suggests a promising path forward for better handling fast-evolving libraries. Our tasks will be made publicly available upon acceptance.