NAACL 2025 Papers

Large Language Models (LLMs) are increasingly serving as evaluators in Natural Language Generation (NLG) tasks; this is often referred to as “LLM-as-a-judge” paradigm. However, the capabilities of LLMs in evaluating NLG quality remain underexplored. Current studies depend on human assessments and simple metrics that fail to capture the discernment of LLMs across diverse NLG tasks. To address this gap, we propose the Discernment of Hierarchical Perturbation (DHP) benchmarking framework, which provides quantitative discernment scores for LLMs. This framework leverages hierarchically perturbed text data and statistical tests to systematically measure the NLG evaluation capabilities of LLMs. We re-established six evaluation datasets for this benchmark, covering four NLG tasks: Summarization, Story Completion, Question Answering, and Translation. Our comprehensive benchmarking of five major LLM families provides critical insight into their strengths and limitations as NLG evaluators. Our dataset is available at https://huggingface.co/datasets/YCWANGVINCE/DHP_Benchmark.

Query-focused tabular summarization is an emerging task in table-to-text generation that synthesizes a summary response from tabular data based on user queries. Traditional transformer-based approaches face challenges due to token limitations and the complexity of reasoning over large tables. To address these challenges, we introduce DETQUS (Decomposition-Enhanced Transformers for QUery-focused Summarization), a system designed to improve summarization accuracy by leveraging tabular decomposition alongside a fine-tuned encoder-decoder model. DETQUS employs a large language model to selectively reduce table size, retaining only query-relevant columns while preserving essential information. This strategy enables more efficient processing of large tables and enhances summary quality. Our approach, equipped with table-based QA model Omnitab, achieves a ROUGE-L score of 0.4437, outperforming the previous state-ofthe- art REFACTOR model (ROUGE-L: 0.422). These results highlight DETQUS as a scalable and effective solution for query-focused tabular summarization, offering a structured alternative to more complex architectures.

Automatic radiology report generation has attracted considerable attention with the rise of computer-aided diagnostic systems. Due to the inherent biases in medical imaging data, generating reports with precise clinical details is challenging yet crucial for accurate diagnosis. To this end, we design a disease description graph that encapsulates comprehensive and pertinent disease information. By aligning visual features with the graph, our model enhances the quality of the generated reports. Furthermore, we introduce a novel informed prompting method which increases the accuracy of short-gram predictions, acting as an implicit bag-of-words planning for surface realization. Notably, this informed prompt succeeds with a three-layer decoder, reducing the reliance on conventional prompting methods that require extensive model parameters. Extensive experiments on two widely-used datasets, IU-Xray and MIMIC-CXR, demonstrate that our method outperforms previous state-of-the-art models.

Dead code introduces several challenges in software development, such as increased binary size and maintenance difficulties. It can also obscure logical errors and be exploited for obfuscation in malware. For LLM-based code-related tasks, dead code introduces vulnerabilities that can mislead these models, raising security concerns. Although modern compilers and IDEs offer dead code elimination, sophisticated patterns can bypass these tools. A universal approach that includes classification, location, explanation, and correction is needed, yet current tools often require significant manual effort. We present DCE-LLM, a framework for automated dead code elimination using a small CodeBERT model with an attribution-based line selector to efficiently locate suspect code. LLMs then generate judgments and explanations, fine-tuned on a large-scale, annotated dead code dataset to provide detailed explanations and patches. DCE-LLM outperforms existing tools, with advanced unreachability detection, automated correction, and support for multiple programming languages. Experimental results show DCE-LLM achieves over 94% F1 scores for unused and unreachable code, significantly surpassing GPT-4o by 30%.

Zero-shot in-context learning (ZS-ICL) aims to conduct in-context learning (ICL) without using human-annotated demonstrations.Existing ZS-ICL methods either use large language models (LLMs) to generate (input, label) pairs as pseudo-demonstrations or leverage historical pseudo-demonstrations to help solve the current problem.They assume that all problems are from the same task and traverse them in a random order.However, in real-world scenarios, problems usually come from diverse tasks, and only a few belong to the same task.The random traversing order may generate unreliable pseudo-demonstrations and lead to error accumulation.To address this problem, we reformulate ZS-**ICL** as a planning problem and propose a **D**emonstration-**AW**are Mo**N**te Carlo Tree Search (MCTS) approach (DAWN-ICL), which leverages MCTS to strategically plan the problem-solving trajectories for ZS-ICL.In addition, to achieve effective and efficient Q value estimation, we propose a demonstration-aware Q-value function and use it to enhance the selection phase and accelerate the expansion and simulation phases in MCTS.Extensive experiments demonstrate the effectiveness and efficiency of DAWN-ICL on in-domain and cross-domain scenarios, and it even outperforms ICL using human-annotated demonstrations.The code is available at https://github.com/txy77/MCTS4ZSICL.

As large language models (LLMs) generate more human-like texts, concerns about the side effects of AI-generated texts (AIGT) have grown. So, researchers have developed methods for detecting AIGT. However, two challenges remain. First, the performance of detecting black-box LLMs is low because existing models focus on probabilistic features. Second, most AIGT detectors have been tested on a single-candidate setting, which assumes that we know the origin of an AIGT and which may deviate from the real-world scenario. To resolve these challenges, we propose DART, which consists of four steps: rephrasing, semantic parsing, scoring, and multiclass classification. We conducted three experiments to test the performance of DART. The experimental result shows that DART can discriminate multiple black-box LLMs without probabilistic features and the origin of AIGT.

Large Language Models (LLMs) have achieved significant success in open-domain question answering. However, they continue to face challenges such as hallucinations and knowledge cutoffs. These issues can be mitigated through in-context learning by providing LLMs with relevant context before generating answers. Recent literature proposes Knowledge Graph Prompting (KGP) which integrates knowledge graphs with an LLM-based traversal agent to substantially enhance document retrieval quality. However, KGP requires costly fine-tuning with large datasets and remains prone to hallucination. In this paper, we propose CuriousLLM, an enhancement that integrates a curiosity-driven reasoning mechanism into an LLM agent. This mechanism enables the agent to generate relevant follow-up questions, thereby guiding the information retrieval process more efficiently.Central to our approach is the development of the new Follow-upQA dataset, which includes questions and supporting evidence as input, with follow-up questions serving as ground truths. These follow-up questions either inquire about what is still missing to fully answer the user’s query or use special tokens to signify that the retrieved evidence is sufficient. Our experiments show that CuriousLLM significantly boosts LLM performance in multi-document question answering (MD-QA), circumventing the substantial computational costs and latency from the original KGP framework.

Large language models, despite their remarkable success in recent years, still exhibit severe cultural bias. Therefore, in this paper, we introduce CultureInstruct, a large-scale instruction-tuning dataset designed to reduce cultural bias in LLMs. CultureInstruct is constructed with an automatic pipeline, utilizing public web sources and a specialized LLM to generate instruction. Our data comprises 430K instructions, ranging from classic NLP tasks to complex reasoning. CultureInstruct also covers 11 most relevant topics to cultural knowledge, making it highly diverse. Our experiments show that fine-tuning LLMs with CultureInstruct results in consistent improvements across three types of cultural benchmarks, including (i) general cultural knowledge, (ii) human opinions and values, and (iii) linguistic cultural bias. Our best model, Qwen2-Instruct 72B + CultureInstruct, outperforms GPT-4o Mini and GPT-4o with 18.47% and 13.07% average relative improvements on cultural benchmarks.

Reinforcement learning with human feedback (RLHF) is shown to largely benefit from precise reward models (RMs). However, recent studies in reward modeling schemes are skewed towards English, limiting the applicability of RLHF in multilingual alignments. In this work, we investigate the cross-lingual transfer of RMs trained in diverse languages, primarily from English. Our experimental results demonstrate the strong cross-lingual transfer of English RMs, exceeding target language RMs by 3~4% average increase in Multilingual RewardBench. Furthermore, we analyze the cross-lingual transfer of RMs through the representation shifts. Finally, we perform multilingual alignment to exemplify how cross-lingual transfer in RM propagates to enhanced multilingual instruction-following capability.

Do speakers of different languages talk differently about what they see? Behavioural and cognitive studies report cultural effects on perception; however, these are mostly limited in scope and hard to replicate. In this work, we conduct the first large-scale empirical study of cross-lingual variation in image descriptions. Using a multimodal dataset with 31 languages and images from diverse locations, we develop a method to accurately identify entities mentioned in captions and present in the images, then measure how they vary across languages. Our analysis reveals that pairs of languages that are geographically or genetically closer tend to mention the same entities more frequently. We also identify entity categories whose saliency is universally high (such as animate beings), low (clothing accessories) or displaying high variance across languages (landscape). In a case study, we measure the differences in a specific language pair (e.g., Japanese mentions clothing far more frequently than English). Furthermore, our method corroborates previous small-scale studies, including 1) Rosch et al. (1976)’s theory of basic-level categories, demonstrating a preference for entities that are neither too generic nor too specific, and 2) Miyamoto et al. (2006)’s hypothesis that environments afford patterns of perception, such as entity counts. Overall, our work reveals the presence of both universal and culture-specific patterns in entity mentions.

There has been increasing interest in building multilingual foundation models for NLP and speech research. This paper examines how to expand the speech translation capability of these models with restricted data. Whisper, a speech foundation model with strong performance on speech recognition and English translation, is used as the example model. Using speech-to-speech retrieval to analyse the audio representations generated by the encoder, we show that utterances from different languages are mapped to a shared semantic space. This shared embedding space can then be leveraged for zero-shot cross-lingual transfer in speech translation. By fine-tuning the Whisper decoder with only English-to-Chinese speech translation data, improved performance for translation to Chinese can be obtained for multiple languages, in addition to English. Furthermore, for languages related to those seen in training it is possible to perform speech translation, despite the model never seeing the language in training, or being able to perform transcription.

While knowledge evaluation in large language models has predominantly focused on academic subjects like math and physics, these assessments often fail to capture the practical demands of real-world professions. In this paper, we introduce IndoCareer, a dataset comprising 8,834 multiple-choice questions designed to evaluate performance in vocational and professional certification exams across various fields. With a focus on Indonesia, IndoCareer provides rich local contexts, spanning six key sectors: (1) healthcare, (2) insurance and finance, (3) creative and design, (4) tourism and hospitality, (5) education and training, and (6) law. Our comprehensive evaluation of 27 large language models shows that these models struggle particularly in fields with strong local contexts, such as insurance and finance. Additionally, while using the entire dataset, shuffling answer options generally maintains consistent evaluation results across models, but it introduces instability specifically in the insurance and finance sectors.

While much work on web agents emphasizes the promise of autonomously performing tasks on behalf of users, in reality, agents often fallshort on complex tasks in real-world contexts and modeling user preference. This presents an opportunity for humans to collaborate with the agent and leverage the agent’s capabilities effectively. We propose CowPilot, a frame- work supporting autonomous as well as human-agent co llaborative w eb navigation, and evaluation across task success and task efficiency. CowPilot reduces the number of steps humans need to perform by allowing agents to propose next steps, while users are able to pause, reject, or take alternative actions. During execution, users can interleave their actions with the agent’s by overriding suggestions or resuming agent control when needed. We conducted case studies on five common websites and found that the human-agent collaborative mode achieves the highest success rate of 95% while requiring humans to perform only 15.2% of the total steps. Even with human interventions during task execution, the agent successfully drives up to half of task success on its own. CowPilot can serve as a useful tool for data collection and agent evaluation across websites, which we believe will enable research in how users and agents can work together. Video demonstrations are available at https://oaishi.github.io/cowpilot.html

Fairness in multi-document summarization (MDS) measures whether a system can generate a summary fairly representing information from documents with different social attribute values. Fairness in MDS is crucial since a fair summary can offer readers a comprehensive view. Previous works focus on quantifying summary-level fairness using Proportional Representation, a fairness measure based on Statistical Parity. However, Proportional Representation does not consider redundancy in input documents and overlooks corpus-level unfairness. In this work, we propose a new summary-level fairness measure, Equal Coverage, which is based on coverage of documents with different social attribute values and considers the redundancy within documents. To detect the corpus-level unfairness, we propose a new corpus-level measure, Coverage Parity. Our human evaluations show that our measures align more with our definition of fairness. Using our measures, we evaluate the fairness of thirteen different LLMs. We find that Claude3-sonnet is the fairest among all evaluated LLMs. We also find that almost all LLMs overrepresent different social attribute values. The code is available at https://github.com/leehaoyuan/coverage_fairness

A binary decision task, like yes-no questions or answer verification, reflects a significant real-world scenario such as where users look for confirmation about the correctness of their decisions on specific issues. In this work, we observe that language models exhibit a negative bias in the binary decisions of complex reasoning tasks. Based on our observations and the rationale about attention-based model dynamics, we propose a negative attention score (NAS) to systematically and quantitatively formulate negative bias. Based on NAS, we identify attention heads that attend to negative tokens provided in the instructions as answer candidate of binary decisions, regardless of the question in the prompt, and validate their association with the negative bias. Additionally, we propose the negative attention score alignment (NASA) method, which is a parameter-efficient fine-tuning technique to address the extracted negatively biased attention heads. Experimental results from various domains of reasoning tasks and large model search space demonstrate that NASA significantly reduces the gap between precision and recall caused by negative bias while preserving their generalization abilities.

The fusion of speech and language in the era of large language models has garnered significant attention. Discrete speech token is often utilized in text-to-speech tasks for speech compression and portability, which is convenient for joint training with text and have good compression efficiency. However, we found that the discrete speech tokenizer still suffers from information loss. Therefore, we propose a simple yet effective continuous speech tokenizer named Cont-SPT, and a text-to-speech model based on continuous speech tokens. Our results show that the speech language model based on the continuous speech tokenizer has better continuity and higher estimated Mean Opinion Scores (MoS). This enhancement is attributed to better information preservation rate of the continuous speech tokenizer across both low and high frequencies in the frequency domain. The code and resources for Cont-SPT can be found in https://github.com/Yixing-Li/Continuous-Speech-Tokenizer.

The field of sign language translation (SLT) is still in its infancy, as evidenced by the low translation quality, even when using deep learn- ing approaches. Probably because of this, many common approaches in other machine learning fields have not been explored in sign language. Here, we focus on continual learning for mul- tilingual SLT. We experiment with three con- tinual learning methods and compare them to four more naive baseline and fine-tuning ap- proaches. We work with four sign languages (ASL, BSL, CSL and DGS) and three spo- ken languages (Chinese, English and German). Our results show that incremental fine-tuning is the best performing approach both in terms of translation quality and transfer capabilities, and that continual learning approaches are not yet fully competitive given the current SOTA in SLT.

Meta-evaluation of automatic evaluation metrics—assessing evaluation metrics themselves—is crucial for accurately benchmarking natural language processing systems and has implications for scientific inquiry, production model development, and policy enforcement. While existing approaches to metric meta-evaluation focus on general statements about the absolute and relative quality of metrics across arbitrary system outputs, in practice, metrics are applied in highly contextual settings, often measuring the performance for a highly constrained set of system outputs. For example, we may only be interested in evaluating a specific model or class of models. We introduce a method for contextual metric meta-evaluation by comparing the local metric accuracy of evaluation metrics. Across translation, speech recognition, and ranking tasks, we demonstrate that the local metric accuracies vary both in absolute value and relative effectiveness as we shift across evaluation contexts. This observed variation highlights the importance of adopting context-specific metric evaluations over global ones.

There has recently been considerable interest in incorporating information retrieval into large language models (LLMs). Retrieval from a dynamically expanding external corpus of text allows a model to incorporate current events and can be viewed as a form of episodic memory. Here we demonstrate that pre-processing the external corpus into semi-structured “atomic facts” makes retrieval more efficient. More specifically, we demonstrate that our particular form of atomic facts improves performance on various question answering tasks when the amount of retrieved text is limited. Limiting the amount of retrieval reduces the size of the context and improves inference efficiency.