NAACL 2025 Papers

Instruction tuning is an important step in making language models useful for direct user interaction. However, the legal domain is underrepresented in typical instruction datasets (e.g., only 10 out of 1600+ tasks in Super-NaturalInstructions). To study whether instruction tuning on legal datasets is necessary for strong legal reasoning, we aggregate 58 annotated legal datasets and write instructions for each, creating LawInstruct. LawInstruct covers 17 global jurisdictions, 24 languages and a total of 12M examples across diverse tasks such as legal QA, summarization of court cases, and legal argument mining. We evaluate our models on LegalBench, measuring legal reasoning across five categories in 162 challenging and realistic legal tasks, and MMLU, to measure potential drops in general reasoning capabilities. We find that legal-specific instruction tuning on Flan-T5 – yielding FLawN-T5 – improves performance on LegalBench across all model sizes, with an aggregate increase of 15 points or 50% over Flan-T5 for the base size. No model size shows performance drops in MMLU. We publish LawInstruct as a resource for further study of instruction tuning in the legal domain.

Instruction-following LLMs have recently allowed systems to discover hidden concepts from a collection of unstructured documents based on a natural language description of the purpose of the discovery (i.e., goal). Still, the quality of the discovered concepts remains mixed, as it depends heavily on LLM’s reasoning ability and drops when the data is noisy or beyond LLM’s knowledge. We present Instruct-LF, a goal-oriented latent factor discovery system that integrates LLM’s instruction-following ability with statistical models to handle large, noisy datasets where LLM reasoning alone falls short. Instruct-LF uses LLMs to propose fine-grained, goal-related properties from documents, estimates their presence across the dataset, and applies gradient-based optimization to uncover hidden factors, where each factor is represented by a cluster of co-occurring properties. We evaluate latent factors produced by Instruct-LF on movie recommendation, text-world navigation, and legal document categorization tasks. These interpretable representations improve downstream task performance by 5-52% than the best baselines and were preferred 1.8 times as often as the best alternative, on average, in human evaluation.

Ancient Chinese Poetry (ACP), a critical aspect of Chinese cultural heritage, presents unique challenges for Large Language Models (LLMs). One of the most pressing challenges is the significant hallucination issues faced by LLMs due to data scarcity and limited ability of general LLMs when dealing with ACP. To address these challenges, this paper constructs the ACP-Corpus, which encompasses 1.1 million ancient poems and 990K related texts, designed to enhance the training and performance of LLMs. Alongside this, we develop the ACP-QA dataset, comprising over 12 million question-answer pairs across 24 task categories, and the ACP-Eval dataset for rigorous evaluation purposes, containing 7,050 entries. Building on this resources, we propose the ACP-RAG framework, a specialized Retrieval-Augmented Generation (RAG) approach that significantly improves the performance of LLMs in the domain of ancient poetry from 49.2% to 89.0%. The ACP-RAG contains five modules of semantic coarse-grained retrieval, semantic fine-grained retrieval, keyword retrieval, keyword matching, and context filtering. Experiments show that ACP-RAG achieves a promising response accuracy of 89.0%, surpassing existing LLMs by a remarkable margin. We believe this work not only advances the capabilities of LLMs in processing ancient Chinese poetry but also contributes to the preservation and innovative development within this rich literary tradition. The datasets and code are available at https://github.com/SCUT-DLVCLab/ACP-RAG.

Large language models (LLMs) have shown superior capabilities in translating figurative language compared to neural machine translation (NMT) systems. However, the impact of different prompting methods and LLM-NMT combinations on idiom translation has yet to be thoroughly investigated. This paper introduces two parallel datasets of sentences containing idiomatic expressions for Persian→English and English→Persian translations, with Persian idioms sampled from our PersianIdioms resource, a collection of 2,200 idioms and their meanings, with 700 including usage examples.Using these datasets, we evaluate various open- and closed-source LLMs, NMT models, and their combinations. Translation quality is assessed through idiom translation accuracy and fluency. We also find that automatic evaluation methods like LLM-as-a-judge, BLEU, and BERTScore are effective for comparing different aspects of model performance. Our experiments reveal that Claude-3.5-Sonnet delivers outstanding results in both translation directions. For English→Persian, combining weaker LLMs with Google Translate improves results, while Persian→English translations benefit from single prompts for simpler models and complex prompts for advanced ones.

Detecting anomalies or out-of-distribution (OOD) samples is critical for maintaining the reliability and trustworthiness of machine learning systems. Recently, Large Language Models (LLMs) have demonstrated their effectiveness not only in natural language processing but also in broader applications due to their advanced comprehension and generative capabilities. The integration of LLMs into anomaly and OOD detection marks a significant shift from the traditional paradigm in the field. This survey focuses on the problem of anomaly and OOD detection under the context of LLMs. We propose a new taxonomy to categorize existing approaches into two classes based on the role played by LLMs. Following our proposed taxonomy, we further discuss the related work under each of the categories and finally discuss potential challenges and directions for future research in this field. We also provide an up-to-date reading list of relevant papers: https://github.com/rux001/Awesome-LLM-Anomaly-OOD-Detection.

Language Confusion is a phenomenon where Large Language Models (LLMs) generate text that is neither in the desired language, nor in a contextually appropriate language. This phenomenon presents a critical challenge in text generation by LLMs, often appearing as erratic and unpredictable behavior. We hypothesize that there are linguistic regularities to this inherent vulnerability in LLMs and shed light on patterns of language confusion across LLMs. We introduce a novel metric, Language Confusion Entropy, designed to directly measure and quantify this confusion, based on language distributions informed by linguistic typology and lexical variation. Comprehensive comparisons with the Language Confusion Benchmark (Marchisio et al., 2024) confirm the effectiveness of our metric, revealing patterns of language confusion across LLMs. We further link language confusion to LLM security, and find patterns in the case of multilingual embedding inversion attacks. Our analysis demonstrates that linguistic typology offers theoretically grounded interpretation, and valuable insights into leveraging language similarities as a prior for LLM alignment and security.

Causal inference has demonstrated significant potential to enhance Natural Language Processing (NLP) models in areas such as predictive accuracy, fairness, robustness, and explainability by capturing causal relationships among variables. The rise of generative Large Language Models (LLMs) has greatly impacted various language processing tasks. This survey focuses on research that evaluates or improves LLMs from a causal view in the following areas: reasoning capacity, fairness and safety issues, explainability, and handling multimodality. Meanwhile, LLMs can assist in causal inference tasks, such as causal relationship discovery and causal effect estimation, by leveraging their generation ability and knowledge learned during pre-training. This review explores the interplay between causal inference frameworks and LLMs from both perspectives, emphasizing their collective potential to further the development of more advanced and robust artificial intelligence systems.

Human bilinguals often use similar brain regions to process multiple languages, depending on when they learned their second language and their proficiency. In large language models (LLMs), how are multiple languages learned and encoded? In this work, we explore the extent to which LLMs share representations of morphsyntactic concepts such as grammatical number, gender, and tense across languages. We train sparse autoencoders on Llama-3-8B and Aya-23-8B, and demonstrate that abstract grammatical concepts are often encoded in feature directions shared across many languages. We use causal interventions to verify the multilingual nature of these representations; specifically, we show that ablating only multilingual features decreases classifier performance to near-chance across languages. We then use these features to precisely modify model behavior in a machine translation task; this demonstrates both the generality and selectivity of these feature’s roles in the network. Our findings suggest that even models trained predominantly on English data can develop robust, cross-lingual abstractions of morphosyntactic concepts.

Citation practices are crucial in shaping the structure of scientific knowledge, yet they are often influenced by contemporary norms and biases. The emergence of Large Language Models (LLMs) introduces a new dynamic to these practices. Interestingly, the characteristics and potential biases of references recommended by LLMs that entirely rely on their parametric knowledge, and not on search or retrieval-augmented generation, remain unexplored. Here, we analyze these characteristics in an experiment using a dataset from AAAI, NeurIPS, ICML, and ICLR, published after GPT-4’s knowledge cut-off date. In our experiment, LLMs are tasked with suggesting scholarly references for the anonymized in-text citations within these papers. Our findings reveal a remarkable similarity between human and LLM citation patterns, but with a more pronounced high citation bias, which persists even after controlling for publication year, title length, number of authors, and venue. The results hold for both GPT-4, and the more capable models GPT-4o and Claude 3.5 where the papers are part of the training data. Additionally, we observe a large consistency between the characteristics of LLM’s existing and non-existent generated references, indicating the model’s internalization of citation patterns. By analyzing citation graphs, we show that the references recommended are embedded in the relevant citation context, suggesting an even deeper conceptual internalization of the citation networks. While LLMs can aid in citation generation, they may also amplify existing biases, such as the Matthew effect, and introduce new ones, potentially skewing scientific knowledge dissemination.

Large Language Models (LLMs) struggle to directly generate correct plans for complex multi-constraint planning problems, even with self-verification and self-critique. For example, a U.S. domestic travel planning benchmark TravelPlanner was proposed in Xie et al. (2024), where the best LLM OpenAI o1-preview can only find viable travel plans with a 10% success rate given all needed information. In this work, we tackle this by proposing an LLM-based planning framework that formalizes and solves complex multi-constraint planning problems as constrained satisfiability problems, which are further consumed by sound and complete satisfiability solvers. We start with TravelPlanner as the primary use case and show that our framework achieves a success rate of 93.9% and is effective with diverse paraphrased prompts. More importantly, our framework has strong zero-shot generalizability, successfully handling unseen constraints in our newly created unseen international travel dataset and generalizing well to new fundamentally different domains. Moreover, when user input queries are infeasible, our framework can identify the unsatisfiable core, provide failure reasons, and offers personalized modification suggestions. We show that our framework can modify and solve for an average of 81.6% and 91.7% unsatisfiable queries from two datasets and prove with ablations that all key components of our framework are effective and necessary.

Large Language Models have demonstrated impressive reasoning capabilities across multiple languages. However, the relationship between capabilities in different languages is less explored. In this work, we decompose the process of reasoning tasks into two separated components: knowledge retrieval and knowledge-free reasoning, and analyze the relationship between cross-lingual transferability and these two components. With adapted commonsense reasoning datasets and constructed knowledge-free reasoning datasets, we show that the knowledge-free reasoning capability can be nearly perfectly transferred across various source-target language directions despite the secondary impact of resource in some specific target languages, while cross-lingual knowledge retrieval significantly hinders the transfer. Moreover, by analyzing the hidden states and feed-forward network neuron activation during the reasoning, we show that higher similarity of hidden representations and larger overlap of activated neurons could explain the better cross-lingual transferability of knowledge-free reasoning than knowledge retrieval. Thus, we hypothesize that knowledge-free reasoning shares similar neurons in different languages for reasoning, while knowledge is stored separately in different languages.

The advancement of Large Language Models (LLMs) has greatly improved our ability to process complex language. However, accurately detecting logical fallacies remains a significant challenge. This study presents a novel and effective prompt formulation approach for logical fallacy detection, applicable in both supervised (fine-tuned) and unsupervised (zero-shot) settings. Our method enriches input text by incorporating implicit contextual information—counterarguments, explanations, and goals—which we query for validity within the argument’s context. We then rank these queries based on confidence scores to inform classification. We evaluate our approach across multiple datasets from 5 domains, covering 29 distinct fallacy types, using models from GPT and LLaMA series. The results show substantial improvements over state-of-the-art models: up to a 0.57 increase in F1-score in zero-shot settings and up to 0.45 in fine-tuned models. Extensive analyses further illustrate why and how our method excels.

While affective expressions on social media have been extensively studied, most research has focused on the Western context. This paper explores cultural differences in affective expressions by comparing valence and arousal on Twitter/X (geolocated to the US) and Sina Weibo (in Mainland China). Using the NRC-VAD lexicon to measure valence and arousal, we identify distinct patterns of emotional expression across both platforms. Our analysis reveals a functional representation between valence and arousal, showing a negative offset in contrast to traditional lab-based findings which suggest a positive offset. Furthermore, we uncover significant cross-cultural differences in arousal, with US users displaying higher emotional intensity than Chinese users, regardless of the valence of the content. Finally, we conduct a comprehensive language analysis correlating n-grams and LDA topics with affective dimensions to deepen our understanding of how language and culture shape emotional expression. These findings contribute to a more nuanced understanding of affective communication across cultural and linguistic contexts on social media.

We examine the pre-training dynamics of language models, focusing on their ability to copy text from preceding context—a fundamental skill for various LLM applications, including in-context learning (ICL) and retrieval-augmented generation (RAG). We propose a novel perspective that Transformer-based language models develop copying abilities similarly to grokking, which refers to sudden generalization on test set long after the model fit to the training set. Our experiments yield three arguments: (1) The pre-training loss decreases rapidly, while the context copying ability of models initially lags and then abruptly saturates. (2) The speed of developing copying ability is independent of the number of tokens trained, similarly to how grokking speed is unaffected by dataset size as long as the data distribution is preserved. (3) Induction heads, the attention heads responsible for copying, form from shallow to deep layers during training, mirroring the development of circuits in deeper layers during grokking. We contend that the connection between grokking and context copying can provide valuable insights for more effective language model training, ultimately improving in-context performance. For example, we demonstrated that techniques that enhance grokking, such as regularization, either accelerate or enhance the development of context copying.

Language model systems are used to process diverse categories of input requests, ranging from improving creative writing to solving programming challenges. It would be useful to know which categories they are good at. However, existing evaluations compare model performance on pre-defined categories, failing to reflect a system’s performance on finer-grained or novel ones. We propose to automatically search for finer-grained categories based on inputs where a system performs well or poorly, and describe them in natural language. To search for these categories, we propose a large number of candidate category descriptions, e.g. “Communication Improvement”, find the subset of inputs that match the category descriptions, and calculate the performance on these categories; then we sort these categories based on their performance, thereby highlighting those that score high or low. As one application, we apply our method to compare LLaMA 3-70B and Claude 3 Opus, which have similar Elo-ratings on Chatbot Arena; our method finds the former is weaker at making text more professional and humorous while better at providing psychological insights, depicting a more nuanced picture of model performance.

Crosswords are a form of word puzzle that require a solver to demonstrate a high degree of proficiency in natural language understanding, wordplay, reasoning, and world knowledge, along with adherence to character and length constraints. In this paper we tackle the challenge of solving crosswords with large language models (LLMs). We demonstrate that the current generation of language models shows significant competence at deciphering cryptic crossword clues and outperforms previously reported state-of-the-art (SoTA) results by a factor of 2-3 in relevant benchmarks. We also develop a search algorithm that builds off this performance to tackle the problem of solving full crossword grids with out-of-the-box LLMs for the very first time, achieving an accuracy of 93% on New York Times crossword puzzles. Additionally, we demonstrate that LLMs generalize well and are capable of supporting answers with sound rationale.

Studies of human psychology have demonstrated that people are more motivated to extend empathy to in-group members than out-group members (Cikara et al., 2011). In this study, we investigate how this aspect of intergroup relations in humans is replicated by LLMs in an emotion intensity prediction task. In this task, the LLM is given a short description of an experience a person had that caused them to feel a particular emotion; the LLM is then prompted to predict the intensity of the emotion the person experienced on a numerical scale. By manipulating the group identities assigned to the LLM’s persona (the “perceiver”) and the person in the narrative (the “experiencer”), we measure how predicted emotion intensities differ between in-group and out-group settings. We observe that LLMs assign higher emotion intensity scores to in-group members than out-group members. This pattern holds across all three types of social groupings we tested: race/ethnicity, nationality, and religion. We perform an in-depth analysis on Llama-3.1-8B, the model which exhibited strongest intergroup bias among those tested.

Though English sentences are typically inflexible vis-à-vis word order, constituents often show far more variability in ordering. One prominent theory presents the notion that constituent ordering is directly correlated with constituent weight: a measure of the constituent’s length or complexity. Such theories are interesting in the context of natural language processing (NLP), because while recent advances in NLP have led to significant gains in the performance of large language models (LLMs), much remains unclear about how these models process language, and how this compares to human language processing. In particular, the question remains whether LLMs display the same patterns with constituent movement, and may provide insights into existing theories on when and how the shift occurs in human language. We compare a variety of LLMs with diverse properties to evaluate broad LLM performance on four types of constituent movement: heavy NP shift, particle movement, dative alternation, and multiple PPs. Despite performing unexpectedly around particle movement, LLMs generally align with human preferences around constituent ordering.

Large language models (LLMs) frequently make errors when handling even simple numerical problems, such as comparing two small numbers. A natural hypothesis is that these errors stem from how LLMs represent numbers, and specifically, whether their representations of numbers capture their numeric values. We tackle this question from the observation that LLM errors on numerical tasks are often distributed across the digits of the answer rather than normally around its numeric value. Through a series of probing experiments and causal interventions, we show that LLMs internally represent numbers with individual circular representations per-digit in base 10.This digit-wise representation, as opposed to a value representation, sheds light on the error patterns of models on tasks involving numerical reasoning and could serve as a basis for future studies on analyzing numerical mechanisms in LLMs.