NAACL 2025 论文列表

The recent introduction of the Assistants API highlights its potential for large language models (LLMs) in role-playing agents (RPA). However, maintaining consistent character personas remains a significant challenge due to variability in information extraction, which frequently omits critical elements such as backstory or interpersonal relationships. To address this limitation, we introduce CharacterGPT, a framework designed to dynamically reconstruct character personas through Character Persona Training (CPT). This approach incrementally updates personas by extracting traits from chapter-wise novel summaries, reflecting the progression of the narrative. Our framework is evaluated through Big Five personality evaluations and creative tasks, in which characters generate original narratives, demonstrating the efficacy of CharacterGPT in preserving persona consistency. The code and results are available at https://github.com/Jeiyoon/charactergpt

Role-playing is a crucial capability of Large Language Models (LLMs), enabling a wide range of practical applications, including intelligent non-player characters, digital twins, and emotional companions. Evaluating this capability in LLMs is challenging due to the complex dynamics involved in role-playing, such as maintaining character fidelity throughout a storyline and navigating open-ended narratives without a definitive ground truth. Current evaluation methods, which primarily focus on question-answering or conversational snapshots, fall short of adequately capturing the nuanced character traits and behaviors essential for authentic role-playing. In this paper, we propose CharacterBox, which is a simulation sandbox designed to generate situational fine-grained character behavior trajectories. These behavior trajectories enable a more comprehensive and in-depth evaluation of role-playing capabilities. CharacterBox consists of two main components: the character agent and the narrator agent. The character agent, grounded in psychological and behavioral science, exhibits human-like behaviors, while the narrator agent coordinates interactions between character agents and environmental changes. Additionally, we introduce two trajectory-based methods that leverage CharacterBox to enhance LLM performance. To reduce costs and facilitate the adoption of CharacterBox by public communities, we fine-tune two smaller models, CharacterNR and CharacterRM, as substitutes for GPT API calls, and demonstrate their competitive performance compared to advanced GPT APIs. The code is available at https://github.com/Paitesanshi/CharacterBox.

Recently, open community-driven platforms like Chatbot Arena that collect user preference data from site visitors have gained reputation as trustworthy publicly available benchmarks for LLM performance. While gold standard, it is often tricky to implement the required guardrails to collect high-quality annotations from humans. In this paper, we demonstrate that different source of bad annotations, both malicious and otherwise, can corrupt the reliability of open leaderboard rankings. In particular, we show that only 10% of poor quality votes by apathetic (site visitors not appropriately incentivized to give correct votes) or adversarial (bad actors seeking to inflate the ranking of a target model) annotators can change the rankings of models by up to 5 places on the leaderboard. Finally, we discuss open challenges in ensuring high quality human annotations.

Context:Despite the impressive capabilities of Large Language Models (LLMs), they pose significant risks in many domains and therefore require guardrails throughout the lifecycle.Problem:Many such guardrails are trained as classifiers with domain-specific human text datasets obtained from sources such as social media and they achieve reasonable performance against closed-domain benchmarks. When deployed in the real world, however, the guardrails have to deal with machine text in an open domain, and their performance deteriorates drastically, rendering them almost unusable due to a high level of false refusal.Solution:In this paper, using a self-harm detector as an example, we demonstrate the specific challenges facing guardrail deployment due to the data drift between training and production environments. More specifically, we formed two hypotheses about the potential causes, i.e. closed vs. open domain, human vs. LLM-generated text, and conducted five experiments to explore various potential remedies, including their respective advantages and disadvantages.Evaluation:While focusing on one example, our experience and knowledge of LLM guardrails give us great confidence that our work contributes to a more thorough understanding of guardrail deployment and can be generalized as a methodology to build more robust domain-specific guardrails in real-world applications.

Retrieval-augmented generation (RAG) with large language models (LLMs) is especially valuable in specialized domains, where precision is critical. To more specialize the LLMs into a target domain, domain-specific RAG has recently been developed by allowing the LLM to access the target domain early via finetuning. The domain-specific RAG makes more sense in resource-constrained environments like edge devices, as they should perform a specific task (e.g. personalization) reliably using only small-scale LLMs. While the domain-specific RAG is well-aligned with edge devices in this respect, it often relies on widely-used reasoning techniques like chain-of-thought (CoT). The reasoning step is useful to understand the given external knowledge, and yet it is computationally expensive and difficult for small-scale LLMs to learn it. Tackling this, we propose the Chain of Rank (CoR) which shifts the focus from intricate lengthy reasoning to simple ranking of the reliability of input external documents. Then, CoR reduces computational complexity while maintaining high accuracy, making it particularly suited for resource-constrained environments. We attain the state-of-the-art (SOTA) results in benchmarks, and analyze its efficacy.

Current research found the issue of Early Answering in large language models (LLMs), where the models already have an answer before generating the Chain-of-Thought (CoT). This phenomenon suggests a potential lack of necessary dependency between the predicted answer and the reasoning process. Consequently, two important questions arise: (1) Is CoT still necessary if the model already has an answer? (2) Can the correctness of the answer serve as valid evidence for the correctness of CoT? To address these questions, we propose a method, namely Chain-of-Probe (CoP), to probe changes in confidence during the model’s reasoning. The probing results show that in a significant number of question-answer cases, CoT appears to be unnecessary, and this necessity correlates with the simplicity of the task, defined by the reasoning steps required. Furthermore, by analyzing patterns in confidence change, we examine the correctness of the model’s reasoning. Our validation reveals that many responses, although correct in their final answer, contain errors in their reasoning process. To this end, we propose a strategic approach based on CoP to prioritize answers with correct reasoning among multiple candidates, thereby bolstering the reliability of the model’s reasoning.

The rise of LLMs has deflected a growing portion of human-computer interactions towards LLM-based chatbots.The remarkable abilities of these models allow users to interact using long, diverse natural language text covering a wide range of topics and styles. Phrasing these messages is a time and effort consuming task, calling for an autocomplete solution to assist users. We present **ChaI-TeA**: **Cha**t **I**n**te**raction **A**utocomplete; An autocomplete evaluation framework for LLM-based chatbot interactions. The framework includes a formal definition of the task, curated datasets and suitable metrics. We use it to evaluate 11 models on this task, finding that while current off-the-shelf models perform fairly, there is still much room for improvement, mainly in ranking of the generated suggestions. We provide insights for practitioners working on this task and open new research directions for researchers in the field. We release our framework to serve as a foundation for future research.

Autonomous LLM-based agents have emerged as a powerful paradigm for complex task execution, yet the field lacks standardized tools for development, deployment, and distribution. We present Cerebrum, an open-source platform that addresses this gap through three key components: (1) a comprehensive SDK featuring a modular four-layer architecture for agent development, encompassing LLM, memory, storage, and tool management; (2) a community-driven Agent Hub for sharing and discovering agents, complete with version control and dependency management; and (3) an interactive web interface for testing and evaluating agents. The platform’s effectiveness is demonstrated through implementations of various agent architectures, including Chain of Thought (CoT), ReAct, and tool-augmented agents. Cerebrum advances the field by providing a unified framework that standardizes agent development while maintaining flexibility for researchers and developers to innovate and distribute their work. Live url for demo can be found at https://app.aios.foundation. Code can be found at https://github.com/agiresearch/Cerebrum. Video demo can be found at https://app.aios.foundation/video-demo.

Generated texts from large language models (LLMs) have been shown to exhibit a variety of harmful, human-like biases against various demographics. These findings motivate research efforts aiming to understand and measure such effects. This paper introduces a causal formulation for bias measurement in generative language models. Based on this theoretical foundation, we outline a list of desiderata for designing robust bias benchmarks. We then propose a benchmark called OccuGender, with a bias-measuring procedure to investigate occupational gender bias. We test several state-of-the-art open-source LLMs on OccuGender, including Llama, Mistral, and their instruction-tuned versions. The results show that these models exhibit substantial occupational gender bias. Lastly, we discuss prompting strategies for bias mitigation and an extension of our causal formulation to illustrate the generalizability of our framework.

Email is a vital conduit for human communication across businesses, organizations, and broader societal contexts. In this study, we aim to model the intents, expectations, and responsiveness in email exchanges. To this end, we release SIZZLER, a new dataset containing 1800 emails annotated with nuanced types of intents and expectations. We benchmark models ranging from feature-based logistic regression to zero-shot prompting of large language models. Leveraging the predictive model for intent, expectations, and 14 other features, we analyze 11.3M emails from GMANE to study how linguistic and social factors influence the conversational dynamics in email exchanges. Through our causal analysis, we find that the email response rates are influenced by social status, argumentation, and in certain limited contexts, the strength of social connection.

Causality is essential in scientific research, enabling researchers to interpret true relationships between variables. These causal relationships are often represented by causal graphs, which are directed acyclic graphs. With the recent advancements in Large Language Models (LLMs), there is an increasing interest in exploring their capabilities in causal reasoning and their potential use to hypothesize causal graphs. These tasks necessitate the LLMs to encode the causal graph effectively for subsequent downstream tasks. In this paper, we introduce CausalGraph2LLM, a comprehensive benchmark comprising over 700k queries across diverse causal graph settings to evaluate the causal reasoning capabilities of LLMs. We categorize the causal queries into two types: graph-level and node-level queries. We benchmark both open-sourced and closed models for our study. Our findings reveal that while LLMs show promise in this domain, they are highly sensitive to the encoding used. Even capable models like GPT-4 and Gemini-1.5 exhibit sensitivity to encoding, with deviations of about 60%. We further demonstrate this sensitivity for downstream causal intervention tasks. Moreover, we observe that LLMs can often display biases when presented with contextual information about a causal graph, potentially stemming from their parametric memory.

Causal reasoning (CR) is a crucial aspect of intelligence, essential for problem-solving, decision-making, and understanding the world. While language models (LMs) can generate rationales for their outputs, their ability to reliably perform causal reasoning remains uncertain, often falling short in tasks requiring a deep understanding of causality. In this paper, we introduce CausalEval, a comprehensive review of research aimed at enhancing LMs for causal reasoning, coupled with an empirical evaluation of current models and methods. We categorize existing methods based on the role of LMs: either as reasoning engines or as helpers providing knowledge or data to traditional CR methods, followed by a detailed discussion of methodologies in each category. We then assess the performance of current LMs and various enhancement methods on a range of causal reasoning tasks, providing key findings and in-depth analysis. Finally, we present insights from current studies and highlight promising directions for future research. We aim for this work to serve as a comprehensive resource, fostering further advancements in causal reasoning with LMs.

Causal inference has been a pivotal challenge across diverse domains such as medicine and economics, demanding a complicated integration of human knowledge, mathematical reasoning, and data mining capabilities. Recent advancements in natural language processing (NLP), particularly with the advent of large language models (LLMs), have introduced promising opportunities for traditional causal inference tasks. This paper reviews recent progress in applying LLMs to causal inference, encompassing various tasks spanning different levels of causation. We summarize the main causal problems and approaches, and present a comparison of their evaluation results in different causal scenarios. Furthermore, we discuss key findings and outline directions for future research, underscoring the potential implications of integrating LLMs in advancing causal inference methodologies.

This paper introduces CaseSumm, a novel dataset for long-context summarization in the legal domain that addresses the need for longer and more complex datasets for summarization evaluation. We collect 25.6K U.S. Supreme Court (SCOTUS) opinions and their official summaries, known as “syllabuses.” Our dataset is the largest open legal case summarization dataset, and is the first to include summaries of SCOTUS decisions dating back to 1815.We also present a comprehensive evaluation of LLM-generated summaries using both automatic metrics and expert human evaluation, revealing discrepancies between these assessment methods. Our evaluation shows Mistral 7b, a smaller open-source model, outperforms larger models on most automatic metrics and successfully generates syllabus-like summaries. In contrast, human expert annotators indicate that Mistral summaries contain hallucinations. The annotators consistently rank GPT-4 summaries as clearer and exhibiting greater sensitivity and specificity. We find that LLM-based evaluations are not more correlated with human evaluations than traditional automatic metrics. Furthermore, our analysis identifies specific hallucinations in generated summaries, including precedent citation errors and misrepresentations of case facts. These findings demonstrate the limitations of current automatic evaluation methods for legal summarization and highlight the critical role of human evaluation in assessing summary quality, particularly in complex, high-stakes domains.

Generating event graphs from long documents is challenging due to the inherent complexity of multiple tasks involved such as detecting events, identifying their relationships, and reconciling unstructured input with structured graphs. Recent studies typically consider all events with equal importance, failing to distinguish salient events crucial for understanding narratives. This paper presents CALLMSAE, a CAscading Large Language Model framework for SAlient Event graph generation, which leverages the capabilities of LLMs and eliminates the need for costly human annotations. We first identify salient events by prompting LLMs to generate summaries, from which salient events are identified. Next, we develop an iterative code refinement prompting strategy to generate event relation graphs, removing hallucinated relations and recovering missing edges. Powered by CALLMSAE, we present NYT-SEG, a large-scale automatically annotated event graph dataset which can serve as distant supervision signals. Fine-tuning contextualised graph generation models on NYT-SEG outperforms the models trained on CAEVO data. Results on a human-annotated test set show that the proposed method generates salient and more accurate graphs, outperforming competitive baselines.

Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its capabilities, but also significantly increases the computational complexity. Mixture-of-Experts (MoE) models address that by allowing the model size to grow without substantially raising training or inference costs. Yet MoE models face challenges regarding knowledge sharing among experts, making their performance somehow sensitive to routing accuracy. To tackle that, previous works introduced shared experts and combined their outputs with those of the top K routed experts in an addition manner. In this paper, inspired by collective matrix factorization to learn shared knowledge among data, we propose CartesianMoE, which implements more effective knowledge sharing among experts in more like a multiplication manner. Extensive experimental results indicate that CartesianMoE outperforms previous MoE models for building LLMs, in terms of both perplexity and downstream task performance. And we also find that CartesianMoE achieves better expert routing robustness.

While NLP models often seek to capture cognitive states via language, the validity of predicted states is determined by comparing them to annotations created without access the cognitive states of the authors. In behavioral sciences, cognitive states are instead measured via experiments. Here, we introduce an experiment-based framework for evaluating language-based cognitive style models against human behavior. We explore the phenomenon of decision making, and its relationship to the linguistic style of an individual talking about a recent decision they made. The participants then follow a classical decision-making experiment that captures their cognitive style, determined by how preferences change during a decision exercise. We find that language features, intended to capture cognitive style, can predict participants’ decision style with moderate-to-high accuracy (AUC 0.8), demonstrating that cognitive style can be partly captured and revealed by discourse patterns.

Embedding models that generate dense vector representations of text are widely used and hold significant commercial value. Companies such as OpenAI and Cohere offer proprietary embedding models via paid APIs, but despite being “hidden” behind APIs, these models are not protected from theft. We present, to our knowledge, the first effort to “steal” these models for retrieval by training thief models on text–embedding pairs obtained from the APIs. Our experiments demonstrate that it is possible to replicate the retrieval effectiveness of commercial embedding models with a cost of under $300. Notably, our methods allow for distilling from multiple teachers into a single robust student model, and for distilling into presumably smaller models with fewer dimension vectors, yet competitive retrieval effectiveness. Our findings raise important considerations for deploying commercial embedding models and suggest measures to mitigate the risk of model theft.

Large language models (LLMs) have shown high agreement with human raters across a variety of tasks, demonstrating potential to ease the challenges of human data collection. In computational social science (CSS), researchers are increasingly leveraging LLM annotations to complement slow and expensive human annotations. Still, guidelines for collecting and using LLM annotations, without compromising the validity of downstream conclusions, remain limited. We introduce Confidence-driven inference: a method that combines LLM annotations and LLM confidence indicators to strategically select which human annotations should be collected, with the goal of producing accurate statistical estimates and provably valid confidence intervals while reducing the number of human annotations needed. Our approach comes with safeguards against LLM annotations of poor quality, guaranteeing that the conclusions will be both valid and no less accurate than if we only relied on human annotations. We demonstrate the effectiveness of Confidence-driven inference over baselines in statistical estimation tasks across three CSS settings—text politeness, stance, and bias—reducing the needed number of human annotations by over 25% in each. Although we use CSS settings for demonstration, Confidence-driven inference can be used to estimate most standard quantities across a broad range of NLP problems.