ACL 2025 Papers

Grammar plays a critical role in natural language processing and text/code generation by enabling the definition of syntax, the creation of parsers, and guiding structured outputs. Although large language models (LLMs) demonstrate impressive capabilities across domains, their ability to infer and generate grammars has not yet been thoroughly explored. In this paper, we aim to study and improve the ability of LLMs for few-shot grammar generation, where grammars are inferred from sets of a small number of positive and negative examples and generated in Backus-Naur Form. To explore this, we introduced a novel dataset comprising 540 structured grammar generation challenges, devised 6 metrics, and evaluated 8 various LLMs against it. Our findings reveal that existing LLMs perform sub-optimally in grammar generation. To address this, we propose an LLM-driven hybrid genetic algorithm, namely HyGenar, to optimize grammar generation. HyGenar achieves substantial improvements in both the syntactic and semantic correctness of generated grammars across LLMs.

In this paper, we present HumanEval Pro and MBPP Pro, a series of benchmarks to evaluate LLMs on self-invoking code generation task. This task involves providing LLMs with a base problem alongside a related, more complex problem. The models must solve the base problem and leverage its solution to address the more complex one, thereby showcasing their capacity for progressive reasoning and problem-solving. This work features three key contributions. First, we propose a general recipe for generating more challenging versions of existing benchmarks. Second, from the analysis of experimental results over twenty large language models (LLM) on our benchmarks, we have two important observations: (i) Most LLMs excel in traditional code generation benchmarks like HumanEval and MBPP, but their performance declines on self-invoking tasks. For example, o1-mini achieves 96.2% pass@1 on HumanEval but only 76.2% on HumanEval Pro. (ii) On self-invoking code generation task, the instruction-tuned models demonstrate only marginal improvements compared to the base models. Third, we disclose the types of failure modes that exist in our evaluation results. All these results underscore the need for further advancements in this area and provide a new prospective to future research.

With a statistical analysis of arXiv paper abstracts, we report a marked drop in the frequency of several words previously identified as overused by ChatGPT, such as “delve”, starting soon after they were pointed out in early 2024. The frequency of certain other words favored by ChatGPT, such as “significant”, has instead kept increasing. These phenomena suggest that some authors of academic papers have adapted their use of large language models (LLMs), for example, by selecting outputs or applying modifications to the LLM-generated content. Such coevolution and cooperation of humans and LLMs thus introduce additional challenges to the detection of machine-generated text in real-world scenarios. Estimating the impact of LLMs on academic writing by examining word frequency remains feasible, and more attention should be paid to words that were already frequently employed, including those that have decreased in frequency due to LLMs’ disfavor. The coevolution between humans and LLMs also merits further study.

As Al advances in text generation, human trust in Al generated content remains constrained by biases that go beyond concerns of accuracy. This study explores how bias shapes the perception of AI versus human generated content. Through three experiments involving text rephrasing, news article summarization, and persuasive writing, we investigated how human raters respond to labeled and unlabeled content. While the raters could not differentiate the two types of texts in the blind test, they overwhelmingly favored content labeled as “Human Generated,” over those labeled “AI Generated,” by a preference score of over 30%. We observed the same pattern even when the labels were deliberately swapped. This human bias against AI has broader societal and cognitive implications, as it undervalues AI performance. This study highlights the limitations of human judgment in interacting with AI and offers a foundation for improving human-AI collaboration, especially in creative fields.

Large Language Models (LLMs) are becoming a common part of our lives, yet few studies have examined how they influence our behavior. Using a cooperative language game in which players aim to converge on a shared word, we investigate how people adapt their communication strategies when paired with either an LLM or another human. Our study demonstrates that LLMs exert a measurable influence on human communication strategies and that humans notice and adapt to these differences irrespective of whether they are aware they are interacting with an LLM. These findings highlight the reciprocal influence of human–AI dialogue and raise important questions about the long-term implications of embedding LLMs in everyday communication.

Should LLMs generate language that makes them seem human? Human-like language might improve user experience, but might also lead to deception, overreliance, and stereotyping. Assessing these potential impacts requires a systematic way to measure human-like tone in LLM outputs. We introduce HumT and SocioT, metrics for human-like tone and other dimensions of social perceptions in text data based on relative probabilities from an LLM. By measuring HumT across preference and usage datasets, we find that users prefer less human-like outputs from LLMs in many contexts. HumT also offers insights into the perceptions and impacts of anthropomorphism: human-like LLM outputs are highly correlated with warmth, social closeness, femininity, and low status, which are closely linked to the aforementioned harms. We introduce DumT, a method using HumT to systematically control and reduce the degree of human-like tone while preserving model performance. DumT offers a practical approach for mitigating risks associated with anthropomorphic language generation.

We study continued training and supervised fine-tuning (SFT) of a language model (LM) to make effective use of long-context information. We first establish a reliable evaluation protocol to guide model development—instead of perplexity or simple needle-in-a-haystack (NIAH) tests, we use a broad set of long-context downstream tasks, and we evaluate models after SFT as this better reveals long-context abilities. Supported by our robust evaluations, we run thorough experiments to decide the data mix for continued pre-training, the instruction tuning dataset, and many other design choices such as position extrapolation. We find that (1) code repositories and books are excellent sources of long data, but it is crucial to combine them with high-quality short-context data; (2) training with a sequence length beyond the evaluation length boosts long-context performance; (3) for SFT, using only short instruction datasets yields strong performance on long-context tasks. Our final model, ProLong-8B, which is initialized from Llama-3 and trained on 40B tokens, demonstrates state-of-the-art long-context performance among similarly sized models at a length of 128K. ProLong outperforms Llama-3.1-8B-Instruct on the majority of long-context tasks despite using only 5% as many tokens during long-context training. Additionally, ProLong can effectively process up to 512K tokens, one of the longest context windows of publicly available LMs.

Aligning powerful AI models on tasks that surpass human evaluation capabilities is the central problem of **superalignment**. To address this problem, weak-to-strong generalization aims to elicit the capabilities of strong models through weak supervisors and ensure that the behavior of strong models aligns with the intentions of weak supervisors without unsafe behaviors such as deception. Although weak-to-strong generalization exhibiting certain generalization capabilities, strong models exhibit significant overfitting in weak-to-strong generalization: Due to the strong fit ability of strong models, erroneous labels from weak supervisors may lead to overfitting in strong models. In addition, simply filtering out incorrect labels may lead to a degeneration in question quality, resulting in a weak generalization ability of strong models on hard questions. To mitigate overfitting in weak-to-strong generalization, we propose a two-stage framework that simultaneously improves the quality of supervision signals and the quality of input questions. Experimental results in three series of large language models and two mathematical benchmarks demonstrate that our framework significantly improves PGR (Performance Gap Recovered) compared to naive weak-to-strong generalization, even achieving up to 100% PGR on some models.

With the advancement of large language models (LLMs), intelligent models have evolved from mere tools to autonomous agents with their own goals and strategies for cooperating with humans. This evolution has birthed a novel paradigm in NLP, i.e., human-model cooperation, that has yielded remarkable progress in numerous NLP tasks in recent years. In this paper, we take the first step to present a thorough review of human-model cooperation, exploring its principles, formalizations, and open challenges. In particular, we introduce a new taxonomy that provides a unified perspective to summarize existing approaches. Also, we discuss potential frontier areas and their corresponding challenges. We regard our work as an entry point, paving the way for more breakthrough research in this regard.

Comparative Question Answering (CQA) lies at the intersection of Question Answering, Argument Mining, and Summarization. It poses unique challenges due to the inherently subjective nature of many questions and the need to integrate diverse perspectives. Although the CQA task can be addressed using recently emerged instruction-following Large Language Models (LLMs), challenges such as hallucinations in their outputs and the lack of transparent argument provenance remain significant limitations.To address these challenges, we construct a manually curated dataset comprising arguments annotated with their relevance. These arguments are further used to answer comparative questions, enabling precise traceability and faithfulness. Furthermore, we define explicit criteria for an “ideal” comparison and introduce a benchmark for evaluating the outputs of various Retrieval-Augmented Generation (RAG) models with respect to argument relevance. All code and data are publicly released to support further research.

Large Language Models (LLMs) have shown remarkable capabilities in knowledge-intensive tasks, while they remain vulnerable when encountering misinformation. Existing studies have explored the role of LLMs in combating misinformation, but there is still a lack of fine-grained analysis on the specific aspects and extent to which LLMs are influenced by misinformation. To bridge this gap, we present MisBench, the current largest and most comprehensive benchmark for evaluating LLMs’ behavior and knowledge preference toward misinformation. MisBench consists of 10,346,712 pieces of misinformation, which uniquely considers both knowledge-based conflicts and stylistic variations in misinformation. Empirical results reveal that while LLMs demonstrate comparable abilities in discerning misinformation, they still remain susceptible to knowledge conflicts and stylistic variations. Based on these findings, we further propose a novel approach called Reconstruct to Discriminate (RtD) to strengthen LLMs’ ability to detect misinformation. Our study provides valuable insights into LLMs’ interactions with misinformation, and we believe MisBench can serve as an effective benchmark for evaluating LLM-based detectors and enhancing their reliability in real-world applications. Codes and data are available at: https://github.com/GKNL/MisBench.

Compositionality is a key aspect of human intelligence, essential for reasoning and generalization. While transformer-based models have become the de facto standard for many language modeling tasks, little is known about how they represent compound words, and whether these representations are compositional. In this study, we test compositionality in Mistral, OpenAI Large, and Google embedding models, and compare them with BERT. First, we evaluate compositionality in the representations by examining six diverse models of compositionality (addition, multiplication, dilation, regression, etc.). We find that ridge regression, albeit linear, best accounts for compositionality. Surprisingly, we find that the classic vector addition model performs almost as well as any other model. Next, we verify that most embedding models are highly compositional, while BERT shows much poorer compositionality. We verify and visualize our findings with a synthetic dataset consisting of fully transparent adjective-noun compositions. Overall, we present a thorough investigation of compositionality.

Large Language Models (LLMs) have significantly improved the performance of unsupervised Event Argument Extraction (EAE) tasks. However, LLMs’ inherent preferences severely hinder their effectiveness in EAE, leading to what we term preference traps, namely, the Prior Knowledge Trap, the Sycophancy Hallucination Trap, and the Output Contradiction Trap. Existing approaches often fall into these traps due to misalignments between their prior knowledge, instructions, or output constraints and LLMs’ preferences, which significantly limits further performance gains. To address this issue, we propose Choose-After-Think (CAT), an unsupervised EAE framework designed to handle these preference traps through targeted measures. CAT innovatively divides the EAE task into two phases: identification of event information (argument roles) (Think Phase) and selection of the final answers from a candidate set (Choose Phase). This two-phase approach reduces the impact of individual token probability anomalies and ensures the integrity of EAE results. Experimental results demonstrate that CAT (based on the local 7B model, zero-shot setting) matches the performance of the best DeepSeek-R1 API model, with a significantly lower time cost.

Several studies have explored the mechanisms of large language models (LLMs) in coding tasks, but most have focused on programming languages (PLs) in a monolingual setting. In this paper, we investigate the relationship between multiple PLs and English in the concept space of LLMs. We perform a few-shot translation task on 21 PL pairs using two Llama-based models. By decoding the embeddings of intermediate layers during this task, we observe that the concept space is closer to English (including PL keywords) and assigns high probabilities to English tokens in the second half of the intermediate layers. We analyze neuron activations for 11 PLs and English, finding that while language-specific neurons are primarily concentrated in the bottom layers, those exclusive to each PL tend to appear in the top layers. For PLs that are highly aligned with multiple other PLs, identifying language-specific neurons is not feasible. These PLs also tend to have a larger keyword set than other PLs and are closer to the model’s concept space regardless of the input/output PL in the translation task. Our findings provide insights into how LLMs internally represent PLs, revealing structural patterns in the model’s concept space. Code is available at https://github.com/cisnlp/code-specific-neurons.

Large Language Models (LLMs) are increasingly deployed as autonomous agents for simulation and decision-making, necessitating a deeper understanding of their decision-making behaviour under risk. We investigate the relationship between LLMs’ personality traits and risk-propensity, applying Cumulative Prospect Theory (CPT) and the Big Five personality framework. We compare the behaviour of several LLMs to human baselines. Our findings show that the majority of the models investigated are risk-neutral rational agents, whilst displaying higher Conscientiousness and Agreeableness traits, coupled with lower Neuroticism. Interventions on Big Five traits, particularly Openness, influence the risk-propensity of several LLMs. Advanced models mirror human personality-risk patterns, suggesting that cognitive biases can be surfaced by optimal prompting. However, their distilled variants show no cognitive bias, suggesting limitations to knowledge transfer processes. Notably, Openness emerges as the most influential factor to risk-propensity, aligning with human baselines. In contrast, less advanced models demonstrate inconsistent generalization of the personality-risk relationship. This research advances our understanding of LLM behaviour under risk and highlights the potential and limitations of personality-based interventions in shaping LLM decision-making.

Despite the remarkable success of transformer-based large language models (LLMs) across various domains, understanding and enhancing their mathematical capabilities remains a significant challenge. In this paper, we conduct a rigorous theoretical analysis of LLMs’ mathematical abilities, with a specific focus on their arithmetic performances. We identify numerical precision as a key factor that influences their effectiveness in arithmetical tasks. Our results show that Transformers operating with low numerical precision fail to address arithmetic tasks, such as iterated addition and integer multiplication, unless the model size grows super-polynomially with respect to the input length. In contrast, Transformers with standard numerical precision can efficiently handle these tasks with significantly smaller model sizes. We further support our theoretical findings through empirical experiments that explore the impact of varying numerical precision on arithmetic tasks, providing valuable insights for improving the mathematical reasoning capabilities of LLMs.

Word Sense Disambiguation (WSD) is a key task in Natural Language Processing (NLP), involving selecting the correct meaning of a word based on its context. With Pretrained Language Models (PLMs) like BERT and DeBERTa now well established, significant progress has been made in understanding contextual semantics. Nevertheless, how well these models inherently disambiguate word senses remains uncertain. In this work, we evaluate several encoder-only PLMs across two popular inventories (i.e. WordNet and the Oxford Dictionary of English) by analyzing their ability to separate word senses without any task-specific fine-tuning. We compute centroids of word senses and measure similarity to assess performance across different layers. Our results show that DeBERTa-v3 delivers the best performance on the task, with the middle layers (specifically the 7th and 8th layers) achieving the highest accuracy, outperforming the output layer by approximately 15 percentage points. Our experiments also explore the inherent structure of WordNet and ODE sense inventories, highlighting their influence on the overall model behavior and performance. Finally, based on our findings, we develop a small, efficient model for the WSD task that attains robust performance while significantly reducing the carbon footprint.

Large language models (LLMs) exihibit increasingly sophisticated linguistic capabilities, yet the extent to which these behaviors reflect human-like cognition versus advanced pattern recognition remains an open question.In this study, we investigate how LLMs process the temporal meaning of linguistic aspect in narratives that were previously used in human studies. Using an Expert-in-the-Loop probing pipeline, we conduct a series of targeted experiments to assess whether LLMs construct semantic representations and pragmatic inferences in a human-like manner.Our findings show that LLMs over-rely on prototypicality, produce inconsistent aspectual judgments, and struggle with causal reasoning derived from aspect, raising concerns about their ability to fully comprehend narratives.These results suggest that LLMs process aspect fundamentally differently from humans and lack robust narrative understanding.Beyond these empirical findings, we develop a standardized experimental framework for the reliable assessment of LLMs’ cognitive and linguistic capabilities.

People can categorize the same entity at multiple taxonomic levels, such as basic (bear), superordinate (animal), and subordinate (grizzly bear). While prior research has focused on basic-level categories, this study is the first attempt to examine the organization of categories by analyzing exemplars produced at the subordinate level. We present a new Italian psycholinguistic dataset of human-generated exemplars for 187 concrete words. We then leverage these data to evaluate whether textual and vision LLMs produce meaningful exemplars that align with human category organization across three key tasks: exemplar generation, category induction, and typicality judgment. Our findings show a low alignment between humans and LLMs, consistent with previous studies. However, their performance varies notably across different semantic domains. Ultimately, this study highlights both the promises and the constraints of using AI-generated exemplars to support psychological and linguistic research.