ACL 2025 Papers

Recent advances in table understanding have focused on instruction-tuning large language models (LLMs) for table-related tasks. However, existing research has overlooked the impact of hyperparameter choices, and also lacks a comprehensive evaluation of the out-of-domain table understanding ability and the general capabilities of these table LLMs. In this paper, we evaluate these abilities in existing table LLMs, and find significant declines in both out-of-domain table understanding and general capabilities as compared to their base models. Through systematic analysis, we show that hyperparameters, such as learning rate, can significantly influence both table-specific and general capabilities. Contrary to the previous table instruction-tuning work, we demonstrate that smaller learning rates and fewer training instances can enhance table understanding while preserving general capabilities. Based on our findings, we introduce TAMA, a TAble LLM instruction-tuned from LLaMA 3.1 8B Instruct, which achieves performance on par with, or surpassing GPT-3.5 and GPT-4 on table tasks, while maintaining strong out-of-domain generalization and general capabilities. Our findings highlight the potential for reduced data annotation costs and more efficient model development through careful hyperparameter selection. We open-source the project and our models.

Existing benchmarks that assess Language Models (LMs) as Language Agents (LAs) for tool use primarily focus on stateless, single-turn interactions or partial evaluations, such as tool selection in a single turn, overlooking the inherent stateful nature of interactions in multi-turn applications. To fulfill this gap, we propose DialogTool, a multi-turn dialogue dataset with stateful tool interactions considering the whole life cycle of tool use, across six key tasks in three stages: 1) tool creation; 2) tool utilization: tool awareness, tool selection, tool execution; and 3) role-consistent response: response generation and role play. Furthermore, we build VirtualMobile – an embodied virtual mobile evaluation environment to simulate API calls and assess the robustness of the created APIs. Taking advantage of these artifacts, we conduct comprehensive evaluation on 13 distinct open- and closed-source LLMs and provide detailed analysis at each stage, revealing that the existing state-of-the-art LLMs still cannot perform well to use tools over long horizons .

Semantic Parsing aims to capture the meaning of a sentence and convert it into a logical, structured form. Previous studies show that semantic parsing enhances the performance of smaller models (e.g., BERT) on downstream tasks. However, it remains unclear whether the improvements extend similarly to LLMs. In this paper, our empirical findings reveal that, unlike smaller models, directly adding semantic parsing results into LLMs reduces their performance. To overcome this, we propose SENSE, a novel prompting approach that embeds semantic hints within the prompt. Experiments show that SENSE consistently improves LLMs’ performance across various tasks, highlighting the potential of integrating semantic information to improve LLM capabilities.

Reward models (RMs) play a crucial role in reinforcement learning from human feedback (RLHF), aligning model behavior with human preferences. However, existing benchmarks for reward models show a weak correlation with the performance of optimized policies, suggesting that they fail to accurately assess the true capabilities of RMs. To bridge this gap, we explore several evaluation designs through the lens of reward overoptimization, i.e., a phenomenon that captures both how well the reward model aligns with human preferences and the dynamics of the learning signal it provides to the policy. The results highlight three key findings on how to construct a reliable benchmark: (i) it is important to minimize differences between chosen and rejected responses beyond correctness, (ii) evaluating reward models requires multiple comparisons across a wide range of chosen and rejected responses, and (iii) given that reward models encounter responses with diverse representations, responses should be sourced from a variety of models. However, we also observe that a extremely high correlation with degree of overoptimization leads to comparatively lower correlation with certain downstream performance. Thus, when designing a benchmark, it is desirable to use the degree of overoptimization as a useful tool, rather than the end goal.

With the advent of neural language models, the performance of code generation has been significantly boosted. However, the problem of repetitions during the generation process continues to linger. Previous work has primarily focused on content repetition, which is merely a fraction of the broader repetition problem in code generation. A more prevalent and challenging problem is structural repetition. In structural repetition, the repeated code appears in various patterns but possesses a fixed structure, which can be inherently reflected in grammar. In this paper, we formally define structural repetition and propose an efficient decoding approach called RPG, which stands for Repetition Penalization based on Grammar, to alleviate the repetition problems in code generation for LLMs. Specifically, RPG first leverages grammar rules to identify repetition problems during code generation, and then strategically decays the likelihood of critical tokens that contribute to repetitions, thereby mitigating them in code generation. To facilitate this study, we construct a new dataset CodeRepetEval to comprehensively evaluate approaches for mitigating the repetition problems in code generation. Extensive experimental results demonstrate that RPG substantially outperforms the best-performing baselines on CodeRepetEval dataset as well as HumanEval and MBPP benchmarks, effectively reducing repetitions and enhancing the quality of generated code.

Investigating bias in large language models (LLMs) is crucial for developing trustworthy AI. While prompt-based through prompt engineering is common, its effectiveness relies on the assumption that models inherently understand biases. Our study systematically analyzed this assumption using the BBQ and StereoSet benchmarks on both open-source models as well as commercial GPT model. Experimental results indicate that prompt-based is often superficial; for instance, the Llama2-7B-Chat model misclassified over 90% of unbiased content as biased, despite achieving high accuracy in identifying bias issues on the BBQ dataset. Additionally, specific evaluation and question settings in bias benchmarks often lead LLMs to choose “evasive answers”, disregarding the core of the question and the relevance of the response to the context. Moreover, the apparent success of previous methods may stem from flawed evaluation metrics. Our research highlights a potential “false prosperity” in prompt-base efforts and emphasizes the need to rethink bias evaluation metrics to ensure truly trustworthy AI. We will release our data and code upon acceptance.

With the increasing demand for substantial amounts of high-quality data to train large language models (LLMs), efficiently filtering large web corpora has become a critical challenge. For this purpose, KenLM, a lightweight n-gram-based language model that operates on CPUs, is widely used. However, the traditional method of training KenLM utilizes only high-quality data and, consequently, does not explicitly learn the linguistic patterns of low-quality data. To address this issue, we propose an ensemble approach that leverages two contrasting KenLMs: (i) Good KenLM, trained on high-quality data; and (ii) Bad KenLM, trained on low-quality data. Experimental results demonstrate that our approach significantly reduces noisy content while preserving high-quality content compared to the traditional KenLM training method. This indicates that our method can be a practical solution with minimal computational overhead for resource-constrained environments.

One of the goals of automatic evaluation metrics in grammatical error correction (GEC) is to rank GEC systems such that it matches human preferences. However, current automatic evaluations are based on procedures that diverge from human evaluation. Specifically, human evaluation derives rankings by aggregating sentence-level relative evaluation results, e.g., pairwise comparisons, using a rating algorithm, whereas automatic evaluation averages sentence-level absolute scores to obtain corpus-level scores, which are then sorted to determine rankings. In this study, we propose an aggregation method for existing automatic evaluation metrics which aligns with human evaluation methods to bridge this gap. We conducted experiments using various metrics, including edit-based metrics, -gram based metrics, and sentence-level metrics, and show that resolving the gap improves results for the most of metrics on the SEEDA benchmark.We also found that even BERT-based metrics sometimes outperform the metrics of GPT-4.

Understanding human preferences is crucial for improving foundation models and building personalized AI systems. However, preferences are inherently diverse and complex, making it difficult for traditional reward models to capture their full range. While fine-grained preference data can help, collecting it is expensive and hard to scale. In this paper, we introduce Decomposed Reward Models (DRMs), a novel approach that extracts diverse human preferences from binary comparisons without requiring fine-grained annotations. Our key insight is to represent human preferences as vectors and analyze them using Principal Component Analysis (PCA). By constructing a dataset of embedding differences between preferred and rejected responses, DRMs identify orthogonal basis vectors that capture distinct aspects of preference. These decomposed rewards can be flexibly combined to align with different user needs, offering an interpretable and scalable alternative to traditional reward models. We demonstrate that DRMs effectively extract meaningful preference dimensions (e.g., helpfulness, safety, humor) and adapt to new users without additional training. Our results highlight DRMs as a powerful framework for personalized and interpretable LLM alignment.

Vision-language Models (VLMs) have emerged as general-purpose tools for addressing a variety of complex computer vision problems. Such models have been shown to be highly capable, but, at the same time, lacking some basic visual understanding skills. In this paper, we set out to understand the limitations of SoTA VLMs on fundamental visual tasks (object classification, spatial understanding, and ability to delineate individual object instances through counting), by constructing a series of tests that probe which components of design, specifically, may be lacking. Importantly, we go significantly beyond the current benchmarks, which simply measure the final performance of VLM response, by also comparing and contrasting it to the performance of probes trained directly on features obtained from the visual encoder, intermediate vision-language projection and LLM-decoder output. In doing so, we uncover shortcomings in VLMs and make a number of important observations about their capabilities, robustness and how they process visual information. We hope our insights will guide progress in further improving VLMs.

Knowledge-intensive multi-hop question answering (QA) tasks, which require integrating evidence from multiple sources to address complex queries, often necessitate multiple rounds of retrieval and iterative generation by large language models (LLMs). However, incorporating many documents and extended contexts poses challenges—such as hallucinations and semantic drift—for lightweight LLMs with fewer parameters. This work proposes a novel framework called DEC (Dynamic Enhancement Chain). DEC first decomposes complex questions into logically coherent subquestions to form a hallucination-free reasoning chain. It then iteratively refines these subquestions through context-aware rewriting to generate effective query formulations. For retrieval, we introduce a lightweight discriminative keyword extraction module that leverages extracted keywords to achieve targeted, precise document recall with relatively low computational overhead. Extensive experiments on three multi-hop QA datasets demonstrate that DEC performs on par with or surpasses state-of-the-art benchmarks while significantly reducing token consumption. Notably, our approach attains state-of-the-art results on models with 8B parameters, showcasing its effectiveness in various scenarios, particularly in resource-constrained environments.

Despite the widespread use of ‘artificial intelligence’ (AI) framing in Natural Language Processing (NLP) research, it is not clear what researchers mean by ”intelligence”. To that end, we present the results of a survey on the notion of ”intelligence” among researchers and its role in the research agenda. The survey elicited complete responses from 303 researchers from a variety of fields including NLP, Machine Learning (ML), Cognitive Science, Linguistics, and Neuroscience.We identify 3 criteria of intelligence that the community agrees on the most: generalization, adaptability, & reasoning.Our results suggests that the perception of the current NLP systems as ”intelligent” is a minority position (29%).Furthermore, only 16.2% of the respondents see developing intelligent systems as a research goal, and these respondents are more likely to consider the current systems intelligent.

Improving cultural competence of language technologies is important. However most recent works rarely engage with the communities they study, and instead rely on synthetic setups and imperfect proxies of culture. In this work, we take a human-centered approach to discover and measure language-based cultural norms, and cultural competence of LLMs. We focus on a single kind of culture, *research cultures*, and a single task, *adapting writing across research cultures*. Through a set of interviews with interdisciplinary researchers, who are experts at moving between cultures, we create a framework of structural, stylistic, rhetorical, and citational norms that vary across research cultures. We operationalise these features with a suite of computational metrics and use them for (a) surfacing latent cultural norms in human-written research papers at scale; and (b) highlighting the lack of cultural competence of LLMs, and their tendency to homogenize writing. Overall, our work illustrates the efficacy of a human-centered approach to measuring cultural norms in human-written and LLM-generated texts.

Generating unbiased summaries in real-world settings such as political perspective summarization remains a crucial application of Large Language Models (LLMs). Yet, existing evaluation frameworks rely on traditional metrics for measuring key attributes such as coverage and faithfulness without verifying their applicability, and efforts to develop improved summarizers are still nascent. We address these gaps by (1) identifying reliable metrics for measuring perspective summary quality, and (2) investigating the efficacy of LLM-based methods beyond zero-shot inference. Namely, we build a test set for benchmarking metric reliability using human annotations and show that traditional metrics underperform compared to language model–based metrics, which prove to be strong evaluators. Using these metrics, we show that reranking-based methods yield strong results, and preference tuning with synthetically generated and reranking-labeled data further boosts performance. Our findings aim to contribute to the reliable evaluation and development of perspective summarization methods.

Rational speakers are supposed to know what they know and what they do not know, and to generate expressions matching the strength of evidence. In contrast, it is still a challenge for current large language models to generate corresponding utterances based on the assessment of facts and confidence in an uncertain real-world environment. While it has recently become popular to estimate and calibrate confidence of LLMs with verbalized uncertainty, what is lacking is a careful examination of the linguistic knowledge of uncertainty encoded in the latent space of LLMs. In this paper, we draw on typological frameworks of epistemic expressions to evaluate LLMs’ knowledge of epistemic modality, using controlled stories. Our experiments show that the performance of LLMs in generating epistemic expressions is limited and not robust, and hence the expressions of uncertainty generated by LLMs are not always reliable. To build uncertainty-aware LLMs, it is necessary to enrich semantic knowledge of epistemic modality in LLMs.

Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks – ranging from harmful content generation to broader societal harms – pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering – simple vector arithmetic for steering model’s behavior during inference – to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.

Reinforcement Learning from Human Feedback (RLHF) has become an essential technique for enhancing pre-trained large language models (LLMs) to generate responses that align with human preferences and societal values. Although RLHF has shown promise, the training of reward models (RMs) still faces the challenge of reward hacking, motivating recent works to prevent RMs from finding shortcuts that bypass the intended optimization objectives by identifying simplistic patterns such as response length. Besides the issue of length bias, our work firstly reveals that prompt-template bias learned by RMs can also cause reward hacking when dealing with some marginal samples, resulting in LLMs preferring to generate responses in a specific format after RLHF fine-tuning, regardless of the format requested in the prompt. To this end, we propose a low-cost but effective method, namely Prompt Bias Calibration (PBC), to estimate the prompt-template bias term during reward modeling, which can be utilized to calibrate reward scores in the following RL fine-tuning process. Then, we show that our PBC method can be flexibly combined with existing algorithms of removing length bias, leading to a further improvement in the aspect of enhancing the quality of generated responses.

Retrieval-Augmented Generation (RAG) enhances factual accuracy by integrating external knowledge, yet it introduces a critical issue: erroneous or biased retrieval can mislead generation, compounding hallucinations, a phenomenon we term Hallucination on Hallucination. To address this, we propose Debate-Augmented RAG (DRAG), a training-free framework that integrates Multi-Agent Debate (MAD) mechanisms into both retrieval and generation stages. In retrieval, DRAG employs structured debates among proponents, opponents, and judges to refine retrieval quality and ensure factual reliability. In generation, DRAG introduces asymmetric information roles and adversarial debates, enhancing reasoning robustness and mitigating factual inconsistencies. Evaluations across multiple tasks demonstrate that DRAG improves retrieval reliability, reduces RAG-induced hallucinations, and significantly enhances overall factual accuracy. Our code is available at https://github.com/Huenao/Debate-Augmented-RAG.

Retrieval-augmented generation (RAG) improves the service quality of large language models by retrieving relevant documents from credible literature and integrating them into the context of the user query.Recently, the rise of the cloud RAG service has made it possible for users to query relevant documents conveniently.However, directly sending queries to the cloud brings potential privacy leakage.In this paper, we are the first to formally define the privacy-preserving cloud RAG service to protect the user query and propose RemoteRAG as a solution regarding privacy, efficiency, and accuracy.For privacy, we introduce (n,𝜖)-DistanceDP to characterize privacy leakage of the user query and the leakage inferred from relevant documents.For efficiency, we limit the search range from the total documents to a small number of selected documents related to a perturbed embedding generated from (n,𝜖)-DistanceDP, so that computation and communication costs required for privacy protection significantly decrease.For accuracy, we ensure that the small range includes target documents related to the user query with detailed theoretical analysis.Experimental results also demonstrate that RemoteRAG can resist existing embedding inversion attack methods while achieving no loss in retrieval under various settings.Moreover, RemoteRAG is efficient, incurring only 0.67 seconds and 46.66KB of data transmission (2.72 hours and 1.43 GB with the non-optimized privacy-preserving scheme) when retrieving from a total of 105 documents.