NAACL 2025 Papers

General-purpose language models (LMs) are aligned to diverse user intents, but fall short when it comes to specific applications. While finetuning is the default method for customized alignment, human annotations are often unavailable in various customization scenarios. Based on the observation that one of the main issues of LM customization is constraint adherence, we investigate the feasibility of using constraints as a bridge from general LMs to customized ones. We investigate common constraints in NLP tasks, categorize them into three classes based on the types of their arguments, and propose a unified framework, ACT (Aligning to ConsTraints), to automatically produce supervision signals for user alignment with constraints. Specifically, ACT uses constraint verifiers, which are typically easy to implement in practice, to compute constraint satisfaction rate (CSR) of each response. It samples multiple responses for each prompt and collect preference labels based on their CSR automatically. Subsequently, ACT adapts the LM to the target task through a ranking-based learning process. Experiments on fine-grained entity typing, abstractive summarization, and temporal question answering show that ACT is able to enhance LMs’ capability to adhere to different classes of constraints, thereby improving task performance comparable to or approaching that of finetuning with labeled data.

Text simplification is crucial for improving accessibility and comprehension for English as a Second Language (ESL) learners. This study goes a step further and aims to facilitate ESL learners’ language acquisition by simplification. Specifically, we propose simplifying complex sentences to appropriate levels for learners while also increasing vocabulary coverage of the target level in the simplifications. We achieve this without a parallel corpus by conducting reinforcement learning on a large language model. Our method employs token-level and sentence-level rewards, and iteratively trains the model on its self-generated outputs to guide the model to search for simplification hypotheses that satisfy the target attributes. Experiment results on CEFR-SP and TurkCorpus datasets show that the proposed method can effectively increase the frequency and diversity of vocabulary of the target level by more than 20% compared to baseline models, while maintaining high simplification quality.

Large language models (LLMs) are increasingly used as automated judges to evaluate recommendation systems, search engines, and other subjective tasks, where relying on human evaluators can be costly, time-consuming, and unscalable. LLMs offer an efficient solution for continuous, automated evaluation. However, since the systems that are built and improved with these judgments are ultimately designed for human use, it is crucial that LLM judgments align closely with human evaluators to ensure such systems remain human-centered. On the other hand, aligning LLM judgments with human evaluators is challenging due to individual variability and biases in human judgments. We propose a simple yet effective framework to align LLM judgments with individual human evaluators or their aggregated judgments, without retraining or fine-tuning the LLM. Our approach learns a linear mapping between the LLM’s outputs and human judgments, achieving over 142% average improvement in agreement across 29 tasks with only a small number of calibration examples used for training. Notably, our method works in zero-shot and few-shot settings, exceeds inter-human agreement on four out of six tasks, and enables smaller LLMs to achieve performance comparable to that of larger models.

Realignment techniques are often employed to enhance cross-lingual transfer in multilingual language models, still, they can sometimes degrade performance in languages that differ significantly from the fine-tuned source language. This paper introduces AlignFreeze, a method that freezes either the layers’ lower half or upper half during realignment. Through controlled experiments on 4 tasks, 3 models, and in 35 languages, we find that realignment affects all the layers but can be the most detrimental to the lower ones. Freezing the lower layers can prevent performance degradation. Particularly, AlignFreeze improves Part-of-Speech (PoS) tagging performances in languages where full realignment fails: with XLM-R, it provides improvements of more than one standard deviation in accuracy in seven more languages than full realignment.

This paper introduces the novel task of multimodal puzzle solving, framed within the context of visual question-answering. We present a new dataset, AlgoPuzzleVQA designed to challenge and evaluate the capabilities of multimodal language models in solving algorithmic puzzles that necessitate both visual understanding, language understanding, and complex algorithmic reasoning. We create the puzzles to encompass a diverse array of mathematical and algorithmic topics such as boolean logic, combinatorics, graph theory, optimization, search, etc., aiming to evaluate the gap between visual data interpretation and algorithmic problem-solving skills. The dataset is generated automatically from code authored by humans. All our puzzles have exact solutions that can be found from the algorithm without tedious human calculations. It ensures that our dataset can be scaled up arbitrarily in terms of reasoning complexity and dataset size. Our investigation reveals that multimodal language models such as GPT4V and Gemini exhibit limited performance in puzzle-solving tasks. We find that their performance is near random in a multi-choice question-answering setup for a significant number of puzzles. The findings emphasize the challenges of integrating visual, language, and algorithmic knowledge for solving complex reasoning problems.

In-context Learning (ICL) has become the primary method for performing natural language tasks with Large Language Models (LLMs). The knowledge acquired during pre-training is crucial for this few-shot capability, providing the model with task priors. However, recent studies have shown that ICL predominantly relies on retrieving task priors rather than “learning” to perform tasks. This limitation is particularly evident in complex subjective domains such as emotion and morality, where priors significantly influence posterior predictions. In this work, we examine whether this is the result of the aggregation used in corresponding datasets, where trying to combine low-agreement, disparate annotations might lead to annotation artifacts that create detrimental noise in the prompt. Moreover, we evaluate the posterior bias towards certain annotators by grounding our study in appropriate, quantitative measures of LLM priors. Our results indicate that aggregation is a confounding factor in the modeling of subjective tasks, and advocate focusing on modeling individuals instead. However, aggregation does not explain the entire gap between ICL and the state of the art, meaning other factors in such tasks also account for the observed phenomena. Finally, by rigorously studying annotator-level labels, we find that it is possible for minority annotators to both better align with LLMs and have their perspectives further amplified.

Large language models (LLMs) are increasingly leveraged to empower autonomous agents to simulate human beings in various fields of behavioral research. However, evaluating their capacity to navigate complex social interactions remains a challenge. Previous studies face limitations due to insufficient scenario diversity, complexity, and a single-perspective focus. To this end, we introduce AgentSense: Benchmarking Social Intelligence of Language Agents through Interactive Scenarios. Drawing on Dramaturgical Theory, AgentSense employs a bottom-up approach to create 1,225 diverse social scenarios constructed from extensive scripts. We evaluate LLM-driven agents through multi-turn interactions, emphasizing both goal completion and implicit reasoning. We analyze goals using ERG theory and conduct comprehensive experiments. Our findings highlight that LLMs struggle with goals in complex social scenarios, especially high-level growth needs, and even GPT-4o requires improvement in private information reasoning.

Next location prediction plays a crucial role in various real-world applications. Recently, due to the limitation of existing deep learning methods, attempts have been made to apply large language models (LLMs) to zero-shot next location prediction task. However, they directly generate the final output using LLMs without systematic design, which limits the potential of LLMs to uncover complex mobility patterns and underestimates their extensive reserve of global geospatial knowledge. In this paper, we introduce AgentMove, a systematic agentic prediction framework to achieve generalized next location prediction. In AgentMove, we first decompose the mobility prediction task and design specific modules to complete them, including spatial-temporal memory for individual mobility pattern mining, world knowledge generator for modeling the effects of urban structure and collective knowledge extractor for capturing the shared patterns among population. Finally, we combine the results of three modules and conduct a reasoning step to generate the final predictions. Extensive experiments utilizing mobility data from two distinct sources reveal that AgentMove surpasses the leading baseline by 3.33% to 8.57% across 8 out of 12 metrics and it shows robust predictions with various LLMs as base and also less geographical bias across cities. Our codes are available via https://github.com/tsinghua-fib-lab/AgentMove.

Speech technologies are transforming interactions across various sectors, from healthcare to call centers and robots, yet their performance on African-accented conversations remains underexplored. We introduce Afrispeech-Dialog, a benchmark dataset of 50 simulated medical and non-medical African-accented English conversations, designed to evaluate automatic speech recognition (ASR) and related technologies. We assess state-of-the-art (SOTA) speaker diarization and ASR systems on long-form, accented speech, comparing their performance with native accents and discover a 10%+ performance degradation. Additionally, we explore medical conversation summarization capabilities of large language models (LLMs) to demonstrate the impact of ASR errors on downstream medical summaries, providing insights into the challenges and opportunities for speech technologies in the Global South. Our work highlights the need for more inclusive datasets to advance conversational AI in low-resource settings.

Hate speech and abusive language are global phenomena that need socio-cultural background knowledge to be understood, identified, and moderated. However, in many regions of the Global South, there have been several documented occurrences of (1) absence of moderation and (2) censorship due to the reliance on keyword spotting out of context. Further, high-profile individuals have frequently been at the center of the moderation process, while large and targeted hate speech campaigns against minorities have been overlooked.These limitations are mainly due to the lack of high-quality data in the local languages and the failure to include local communities in the collection, annotation, and moderation processes. To address this issue, we present AfriHate: a multilingual collection of hate speech and abusive language datasets in 15 African languages. Each instance in AfriHate is a tweet annotated by native speakers familiar with the regional culture. We report the challenges related to the construction of the datasets and present various classification baseline results with and without using LLMs. We find that model performance highly depends on the language and that multilingual models can help boost performance in low-resource settings.

Automatic speech recognition (ASR) systems have traditionally been evaluated using English datasets, with the word error rate (WER) serving as the predominant metric. WER’s simplicity and ease of interpretation have contributed to its widespread adoption, particularly for English. However, as ASR systems expand to multilingual contexts, WER fails in various ways, particularly with morphologically complex languages or those without clear word boundaries. Our work documents the limitations of WER as an evaluation metric and advocates for the character error rate (CER) as the primary metric in multilingual ASR evaluation. We show that CER avoids many of the challenges WER faces and exhibits greater consistency across writing systems. We support our proposition by conducting human evaluations of ASR transcriptions in three languages—Malayalam, English, and Arabic—which exhibit distinct morphological characteristics. We show that CER correlates more closely with human judgments than WER, even for English. To facilitate further research, we release our human evaluation dataset for future benchmarking of ASR metrics. Our findings suggest that CER should be prioritized, or at least supplemented, in multilingual ASR evaluations to account for the varying linguistic characteristics of different languages.

As the integration of large language models into daily life is on the rise, there is still a lack of dataset for *advising on subjective and personal dilemmas*. To address this gap, we introduce AdvisorQA, which aims to improve LLMs’ capability to offer advice for deeply subjective concerns, utilizing the LifeProTips Reddit forum. This forum features a dynamic interaction where users post advice-seeking questions, receiving an average of 8.9 advice per query, with 164.2 upvotes from hundreds of users, embodying a *collective intelligence*. Therefore, we’ve completed a dataset encompassing daily life questions, diverse corresponding responses, and majority vote ranking, which we use to train a helpfulness metric. In baseline experiments, models aligned with AdvisorQA dataset demonstrated improved helpfulness through our automatic metric, as well as GPT-4 and human evaluations. Additionally, we expanded the independent evaluation axis to include harmlessness. AdvisorQA marks a significant leap in enhancing QA systems to provide subjective, helpful, and harmless advice, showcasing LLMs’ improved understanding of human subjectivity.

Evaluation of large language models (LLMs) in low-resource languages like Persian has received less attention than in high-resource languages like English. Existing evaluation approaches for Persian LLMs generally lack comprehensive frameworks, limiting their ability to assess models’ performance over a wide range of tasks requiring considerable cultural and contextual knowledge, as well as a deeper understanding of Persian literature and style. This paper first aims to fill this gap by providing two new benchmarks, PeKA and PK-BETS, on topics such as history, literature, and cultural knowledge, as well as challenging the present state-of-the-art models’ abilities in a variety of Persian language comprehension tasks. These datasets are meant to reduce data contamination while providing an accurate assessment of Persian LLMs. The second aim of this paper is the general evaluation of LLMs across the current Persian benchmarks to provide a comprehensive performance overview. By offering a structured evaluation methodology, we hope to promote the examination of LLMs in the Persian language.

With the success of integrating large language models into the development of conversational systems, many studies have shown the effectiveness of retrieving and augmenting external knowledge for informative responses. While many existing studies agree on the necessity of Retrieval Augmented Generation (RAG), further investigation into the necessity and value of applying RAG to every turn of the conversation is needed. In this study, we propose to investigate the need for each turn of system response to be augmented with external knowledge. In particular, by leveraging human judgements on the binary choice of adaptive augmentation, we develop RAGate, a gating model, which models conversation context and relevant inputs to predict if a conversational system requires RAG for improved responses. We conduct extensive experiments on devising and applying RAGate to conversational models, joined with well-rounded analyses of various conversational scenarios. Our experimental results and analysis indicate the effective application of RAGate in RAG-based conversational systems in identifying if system responses require RAG to generate high-quality responses with high confidence. This study also identifies and shows the correlation between the generation’s confidence level and the relevance of the augmented knowledge. We have also released the implementation code and resources in https://github.com/wangxieric/RAGate.

Recent advances on instruction fine-tuning have led to the development of various prompting techniques for large language models, such as explicit reasoning steps. However, the success of techniques depends on various parameters, such as the task, language model, and context provided. Finding an effective prompt is, therefore, often a trial-and-error process. Most existing approaches to automatic prompting aim to optimize individual techniques instead of compositions of techniques and their dependence on the input. To fill this gap, we propose an adaptive prompting approach that predicts the optimal prompt composition ad-hoc for a given input. We apply our approach to social bias detection, a highly context-dependent task that requires semantic understanding. We evaluate it with three large language models on three datasets, comparing compositions to individual techniques and other baselines. The results underline the importance of finding an effective prompt composition. Our approach robustly ensures high detection performance, and is best in several settings. Moreover, first experiments on other tasks support its generalizability.

Large Language Model (LLM) agents exhibit remarkable performance across diverse applications by using external tools to interact with environments. However, integrating external tools introduces security risks, such as indirect prompt injection (IPI) attacks. Despite defenses designed for IPI attacks, their robustness remains questionable due to insufficient testing against adaptive attacks.In this paper, we evaluate eight different defenses and bypass all of them using adaptive attacks, consistently achieving an attack success rate of over 50%.This reveals critical vulnerabilities in current defenses. Our research underscores the need for adaptive attack evaluation when designing defenses to ensure robustness and reliability.The code is available at https://github.com/uiuc-kang-lab/AdaptiveAttackAgent.

Text simplification aims to enhance the clarity and comprehensibility of a complex text while preserving its original meaning. Previous research on the automatic evaluation of text simplification has primarily focused on sentence simplification, with commonly used metrics such as SARI and advanced metrics such as LENS being trained and evaluated at the sentence level. However, these metrics often underperform on longer texts. In our study, we propose a novel approach to adapt existing sentence-level metrics for paragraph- or document-level simplification. We benchmark our approach against a wide variety of existing reference-based and reference-less metrics across multiple domains. Empirical results demonstrate that our approach outperforms traditional sentence-level metrics in terms of correlation with human judgment. Furthermore, we evaluate the sensitivity and robustness of various metrics to different types of errors produced by existing text simplification systems.

Recent advances in large language models (LLMs) have demonstrated potential for LLM agents. To facilitate the training for these agents with both linguistic feedback and non-linguistic reward signals, we introduce Learning through Communication (LTC). We design a universal buffer to store all the feedback, and an iterative pipeline to enable an LLM agent to explore and update its policy in an given environment. To optimize agent interactions for task-specific learning with our universal buffer and pipeline, we introduce diverse communication patterns tailored for both single-agent and multi-agent environments. We evaluate the efficacy of our LTC approach on four diverse datasets: ALFWorld (single-agent), HotpotQA (multi-agent collaboration), Chameleon (multi-agent competition), and GSM8k (multi-agent teacher-student). On these data sets, LTC outperforms the supervised instruction fine-tuning baselines by 3.6% to 12%. These results highlight the versatility and efficiency of LTC in facilitating online adaptation for LLM agents.

Knowledge conflict arises from discrepancies between information in the context of a large language model (LLM) and the knowledge stored in its parameters. This can hurt performance when using standard decoding techniques, which tend to ignore the context. Existing test-time contrastive methods seek to address this by comparing the LLM’s output distribution with and without the context and adjust the model according to the contrast between them. However, we find that these methods frequently misjudge the degree of conflict and struggle to handle instances that vary in their amount of conflict, with static methods over-adjusting when conflict is absent. We propose a fine-grained, instance-level approach called AdaCAD, which dynamically infers the weight of adjustment based on the degree of conflict, as measured by the Jensen-Shannon divergence between distributions representing contextual and parametric knowledge. Across four LLMs, six question-answering (QA) and three summarization datasets, we demonstrate that AdaCAD consistently outperforms other decoding baselines with average QA accuracy gains of 14.21% (absolute) over a static contrastive baseline, and improves the factuality of summaries by 6.19 (AlignScore). Lastly, we show that while contrastive baselines hurt performance when conflict is absent, AdaCAD mitigates these losses, making it more applicable to real-world datasets in which some examples have conflict and others do not.