NAACL 2025 论文列表

Piaget’s Theory of Cognitive Development (PTC) posits that the development of cognitive levels forms the foundation for human learning across various abilities. As Large Language Models (LLMs) have recently shown remarkable abilities across a wide variety of tasks, we are curious about the cognitive levels of current LLMs: to what extent they have developed and how this development has been achieved. To this end, we construct a benchmark CogLM (Cognitive Ability Evaluation for Language Model) based on PTC to assess the cognitive levels of LLMs. CogLM comprises 1,220 questions spanning 10 cognitive abilities crafted by more than 20 human experts, providing a comprehensive testbed for the cognitive levels of LLMs. Through extensive experiments across multiple mainstream LLMs with CogLM, we find that: (1) In our testing framework, advanced LLMs (such as GPT-4) have demonstrated human-like cognitive abilities, comparable to those of a 20-year-old human. (2) The parameter size and optimization objective are two key factors affecting the cognitive levels of LLMs. (3) The performance on downstream tasks is positively correlated with the level of cognitive abilities. These findings fill the gap in research on the cognitive abilities of LLMs, tracing the development of LLMs from a cognitive perspective and guiding the future direction of their evolution.

Large Language Models (LLMs) excel in stand-alone code tasks like HumanEval and MBPP, but struggle with handling entire code repositories. This challenge has prompted research on enhancing LLM-codebase interaction at a repository scale. Current solutions rely on similarity-based retrieval or manual tools and APIs, each with notable drawbacks. Similarity-based retrieval often has low recall in complex tasks, while manual tools and APIs are typically task-specific and require expert knowledge, reducing their generalizability across diverse code tasks and real-world applications. To mitigate these limitations, we introduce CodexGraph, a system that integrates LLM agents with graph database interfaces extracted from code repositories. By leveraging the structural properties of graph databases and the flexibility of the graph query language, CodexGraph enables the LLM agent to construct and execute queries, allowing for precise, code structure-aware context retrieval and code navigation. We assess CodexGraph using three benchmarks: CrossCodeEval, SWE-bench, and EvoCodeBench. Additionally, we develop five real-world coding applications. With a unified graph database schema, CodexGraph demonstrates competitive performance and potential in both academic and real-world environments, showcasing its versatility and efficacy in software engineering. Our code and demo will be released soon.

Pretrained on massive amounts of code and text data, large language models (LLMs) have demonstrated remarkable achievements in performing code generation tasks. With additional execution-based feedback, these models can act as agents with capabilities to self-refine and improve generated code autonomously. However, on challenging coding tasks with extremely large search space, current agentic approaches still struggle with multi-stage planning, generating, and debugging. To address this problem, we propose CodeTree, a framework for LLM agents to efficiently explore the search space in different stages of the code generation process. Specifically, we adopted a unified tree structure to explicitly explore different coding strategies, generate corresponding coding solutions, and subsequently refine the solutions. In each stage, critical decision-making (ranking, termination, expanding) of the exploration process is guided by both the environmental execution-based feedback and LLM-agent-generated feedback. We comprehensively evaluated CodeTree on 7 code generation benchmarks and demonstrated the significant performance gains of CodeTree against strong baselines. Using GPT-4o as the base model, we consistently achieved top results of 95.1% on HumanEval, 98.7% on MBPP, and 43.0% on CodeContests. On the challenging SWEBench benchmark, our approach led to significant performance gains, achieving a 31.9% solving rate.

In this paper, we propose CodeSCM, a Structural Causal Model (SCM) for analyzing multi-modal code generation using large language models (LLMs). By applying interventions to CodeSCM, we measure the causal effects of different prompt modalities, such as natural language, code, and input-output examples, on the model. CodeSCM introduces latent mediator variables to separate the code and natural language semantics of a multi-modal code generation prompt. Using the principles of Causal Mediation Analysis on these mediators we quantify direct effects representing the model’s spurious leanings. We find that, in addition to natural language instructions, input-output examples significantly influence code generation.

While language models (LMs) excel at generating code, many programs are difficult to generate using only parametric knowledge. Despite the success of retrieval-augmented generation (RAG) in text-centric tasks, its potential for code generation remains under-explored. This work introduces CodeRAG-bench, a holistic retrieval-augmented code generation benchmark covering tasks like basic programming, open-domain, and repository-level problems and provides reproducible evaluations on both retrieval and end-to-end code generation performance. We further create a diverse, open datastore for code retrieval, aggregating sources such as competition solutions, tutorials, library documentation, StackOverflow posts, and GitHub repositories. Based on CodeRAG-bench, we conduct large-scale evaluations of 10 retrievers and 10 LMs and systematically analyze when retrieval can benefit code generation models and identify remaining challenges. We find that while retrieving high-quality contexts improves code generation, retrievers often struggle to fetch useful contexts, and generators face limitations in using those contexts effectively. We hope CodeRAG-bench encourages further development in code-oriented RAG methods.

Assuring the data quality of tabular datasets is essential for the efficiency of the diverse tabular downstream tasks (like summarization and fact-checking). Data-wrangling tasks effectively address the challenges associated with structured data processing to improve the quality of tabular data. Traditional statistical methods handle numeric data efficiently but often fail to understand the semantic context of the textual data in tables. Deep learning approaches are resource-intensive, requiring task and dataset-specific training. Addressing these shortcomings, we present an automated system that leverages LLMs to generate executable code for data-wrangling tasks like missing value imputation, error detection, and error correction. Our system aims to identify inherent patterns in the data while leveraging external knowledge, effectively addressing both memory-independent and memory-dependent tasks.

Code Language Models have been trained togenerate accurate solutions, typically with noregard for runtime. On the other hand, previousworks that explored execution optimisationhave observed corresponding drops infunctional correctness. To that end, we introduceCode-Optimise, a framework that incorporatesboth correctness (passed, failed) andruntime (quick, slow) as learning signals viaself-generated preference data. Our frameworkis both lightweight and robust as it dynamicallyselects solutions to reduce overfitting whileavoiding a reliance on larger models for learningsignals. Code-Optimise achieves significantimprovements in pass@k while decreasingthe competitive baseline runtimes by anadditional 6% for in-domain data and up to3% for out-of-domain data. As a by-product,the average length of the generated solutionsis reduced by up to 48% on MBPP and 23%on HumanEval, resulting in faster and cheaperinference. The generated data and codebaseis open-sourced at https://github.com/huawei-noah/HEBO/tree/Code_Optimise.

Retrieval-Augmented Generation (RAG) models excel in knowledge-intensive tasks, especially under few-shot learning constraints. We introduce CoRAG, a framework extending RAG to collaborative settings, where clients jointly train a shared model using a collaborative passage store. To evaluate CoRAG, we introduce CRAB, a benchmark for collaborative homogeneous open-domain question answering. Our experiments demonstrate that CoRAG consistently outperforms both parametric collaborative learning methods and locally trained RAG models in low-resource scenarios. Further analysis reveals the critical importance of relevant passages within the shared store, the surprising benefits of incorporating irrelevant passages, and the potential for hard negatives to negatively impact performance. This introduces a novel consideration in collaborative RAG: the trade-off between leveraging a collectively enriched knowledge base and the potential risk of incorporating detrimental passages from other clients. Our findings underscore the viability of CoRAG, while also highlighting key design challenges and promising avenues for future research.

Automatic code question answering aims to generate precise answers to questions about code by analyzing code snippets. To provide an appropriate answer, it is necessary to accurately understand the relevant part of the code and correctly interpret the intent of the question. However, in real-world scenarios, the questioner often provides only a portion of the code along with the question, making it challenging to find an answer. The responder should be capable of providing a suitable answer using such limited information. We propose a knowledge-based framework, CoRAC, an automatic code question responder that enhances understanding through selective API document retrieval and question semantic intent clustering. We evaluate our method on three real-world benchmark datasets and demonstrate its effectiveness through various experiments. We also show that our method can generate high-quality answers compared to large language models, such as ChatGPT.

Legal professionals often struggle with lengthy judgments and require efficient summarization for quick comprehension. To address this challenge, we investigate the need for structured planning in legal case summarization, particularly through event-centric representations that reflect the narrative nature of legal case documents. We propose our framework, CoPERLex, which operates in three stages: first, it performs content selection to identify crucial information from the judgment; second, the selected content is utilized to generate intermediate plans through event-centric representations modeled as Subject-Verb-Object tuples; and finally, it generates coherent summaries based on both the content and the structured plan. Our experiments on four legal summarization datasets demonstrate the effectiveness of integrating content selection and planning components, highlighting the advantages of event-centric plans over traditional entity-centric approaches in the context of legal judgements.

Large language models (LLMs) often retain outdated or incorrect information from pre-training, which undermines their reliability. While model editing methods have been developed to address such errors without full re-training, they frequently suffer from knowledge conflicts, where outdated information interferes with new knowledge. In this work, we propose Conflict-free Model Editing (CoME), a novel framework that enhances the accuracy of knowledge updates in LLMs by selectively removing outdated knowledge. CoME leverages unlearning to mitigate knowledge interference, allowing new information to be integrated without compromising relevant linguistic features. Through experiments on GPT-J and LLaMA-3 using Counterfact and ZsRE datasets, we demonstrate that CoME improves both editing accuracy and model reliability when applied to existing editing methods. Our results highlight that the targeted removal of outdated knowledge is crucial for enhancing model editing effectiveness and maintaining the model’s generative performance.

We introduce CluSanT, a novel text sanitization framework based on Metric Local Differential Privacy (MLDP). Our framework consists of three components: token clustering, cluster embedding, and token sanitization. For the first, CluSanT employs Large Language Models (LLMs) to create—a set of potential substitute tokens which we meaningfully cluster. Then, we develop a parameterized cluster embedding that balances the trade-off between privacy and utility. Lastly, we propose a MLDP algorithm which sanitizes/substitutes sensitive tokens in a text with the help of our embedding. Notably, our MLDP-based framework can be tuned with parameters such that (1) existing state-of-the-art (SOTA) token sanitization algorithms can be described—and improved—via our framework with extremal values of our parameters, and (2) by varying our parameters, we allow for a whole spectrum of privacy-utility tradeoffs between the two SOTA. Our experiments demonstrate CluSanT’s balance between privacy and semantic coherence, highlighting its capability as a valuable framework for privacy-preserving text sanitization.

Chinese literary classics hold significant cultural and educational value, offering deep insights into morality, history, and human nature. These works often include classical Chinese and complex narratives, making them difficult for children to read. To bridge this gap, we introduce a child-friendly literary adaptation (CLA) task to adapt the Chinese literary classic into engaging and accessible text for children. However, recent large language models (LLMs) overlook children’s reading preferences (i.e., vivid character portrayals, concise narrative structures, and appropriate readability with simpler words and sentences), which poses challenges in CLA. In this paper, we propose a method called InstructChild, which augments the LLM with these preferences for adaptation. Specifically, we first obtain the characters’ personalities and narrative structure as additional information for fine-grained instruction tuning. Then, we devise a readability metric as the reward to align the LLM with the children’s reading level. Finally, a lookahead decoding strategy is applied to improve the readability of the generated text during inference. To support the evaluation of CLA task, we construct the Classic4Children dataset, which comprises both the original and child-friendly versions of the Four Great Classical Novels of Chinese literature. Experimental results show that our InstructChild significantly improves performance in automatic and human evaluation.

In this work, we explore the challenges of developing interactive assistants that resolve ambiguity by asking their users clarifying questions. Specifically, we develop a task-agnostic framework for evaluating a system’s ability to determine when to ask for clarification. Determining when to ask for clarification is a challenging task that requires systems to consider the demands of the individual user (i.e., how much they prioritize speed and usability versus carefulness) and the distribution of interpretations for a given request (i.e., whether an ambiguous request has one dominant, inferable interpretation). Using this framework, we evaluate systems for determining when to clarify across three NLP applications: QA, MT, and NLI. Finally, we introduce present a novel uncertainty estimation approach, IntentSim, that determines the utility of asking a clarifying question by estimating the entropy over user intents. Our method consistently outperforms existing uncertainty estimation approaches at identifying predictions that will benefit from clarification. Furthermore, we find that IntentSim is robust, demonstrating improvements across a wide range of NLP tasks and LMs. Together, our work lays foundation for further studies on clarifying interactions with LM assistants.

Recent advances in natural language processing and the increased use of large language models have exposed new security vulnerabilities, such as backdoor attacks. Previous backdoor attacks require input manipulation after model distribution to activate the backdoor, posing limitations in real-world applicability. Addressing this gap, we introduce a novel Claim-Guided Backdoor Attack (CGBA), which eliminates the need for such manipulations by utilizing inherent textual claims as triggers. CGBA leverages claim extraction, clustering, and targeted training to trick models to misbehave on targeted claims without affecting their performance on clean data. CGBA demonstrates its effectiveness and stealthiness across various datasets and models, significantly enhancing the feasibility of practical backdoor attacks. Our code and data will be available at https://github.com/minkyoo9/CGBA.

E-commerce live streaming in China, particularly on platforms like Douyin, has become a major sales channel, but hosts often use morphs to evade scrutiny and engage in false advertising. This study introduces the Live Auditory Morph Resolution (LiveAMR) task to detect such violations. Unlike previous morph research focused on text-based evasion in social media and underground industries, LiveAMR targets pronunciation-based evasion in health and medical live streams. We constructed the first LiveAMR dataset with 86,790 samples and developed a method to transform the task into a text-to-text generation problem. By leveraging large language models (LLMs) to generate additional training data, we improved performance and demonstrated that morph resolution significantly enhances live streaming regulation.

In industrial LLM development, evaluating large language models (LLMs) is critical for tasks like benchmarking internal models and detecting regressions during fine-tuning, but existing benchmark aggregation methods, such as Elo-based systems, can be resource-intensive, public facing, and time-consuming. Here, we describe Chatbot Arena Estimate (CAE), a practical framework for aggregating performance across diverse benchmarks. The framework, developed and widely adopted within our organization, addresses the need for quick, accurate, and cost-efficient evaluations of LLMs. CAE generates two primary metrics: a “Goodness” score (answer accuracy) and a “Fastness” score (cost or queries per second, QPS). These metrics allow for model ranking both overall and within specific subdomains, enabling informed decisions during model iteration and deployment. We demonstrate CAE’s effectiveness by comparing it with existing benchmarks, including the full Chatbot Arena and the MMLU leaderboard. Notably, our approach achieves higher Pearson correlation with Chatbot Arena Elo scores than MMLU’s correlation with Chatbot Arena Elo scores, validating its reliability for real-world LLM evaluation.

This paper aims to efficiently enable large language models (LLMs) to use external knowledge and goal guidance in conversational recommender system (CRS) tasks. Advanced LLMs (e.g., ChatGPT) are limited in domain-specific CRS tasks for 1) generating grounded responses with recommendation-oriented knowledge, or 2) proactively leading the conversations through different dialogue goals. In this work, we first analyze those limitations through a comprehensive evaluation, showing the necessity of external knowledge and goal guidance which contribute significantly to the recommendation accuracy and language quality. In light of this finding, we propose a novel ChatCRS framework to decompose the complex CRS task into several sub-tasks through the implementation of 1) a knowledge retrieval agent using a tool-augmented approach to reason over external Knowledge Bases and 2) a goal-planning agent for dialogue goal prediction. Experimental results on two multi-goal CRS datasets reveal that ChatCRS sets new state-of-the-art benchmarks, improving language quality of informativeness by 17% and proactivity by 27%, and achieving a tenfold enhancement in recommendation accuracy.

Property inheritance—a phenomenon where novel properties are projected from higher level categories (e.g., birds) to lower level ones (e.g., sparrows)—provides a unique window into how humans organize and deploy conceptual knowledge. It is debated whether this ability arises due to explicitly stored taxonomic knowledge vs. simple computations of similarity between mental representations. How are these mechanistic hypotheses manifested in contemporary language models? In this work, we investigate how LMs perform property inheritance with behavioral and causal representational analysis experiments. We find that taxonomy and categorical similarities are not mutually exclusive in LMs’ property inheritance behavior. That is, LMs are more likely to project novel properties from one category to the other when they are taxonomically related and at the same time, highly similar. Our findings provide insight into the conceptual structure of language models and may suggest new psycholinguistic experiments for human subjects.