ACL 2025 论文列表

In recent research, large language models (LLMs) have been increasingly used to investigate public opinions. This study investigates the algorithmic fidelity of LLMs, i.e., the ability to replicate the socio-cultural context and nuanced opinions of human participants. Using open-ended survey data from the German Longitudinal Election Studies (GLES), we prompt different LLMs to generate synthetic public opinions reflective of German subpopulations by incorporating demographic features into the persona prompts. Our results show that Llama performs better than other LLMs at representing subpopulations, particularly when there is lower opinion diversity within those groups. Our findings further reveal that the LLM performs better for supporters of left-leaning parties like The Greens and The Left compared to other parties, and matches the least with the right-party AfD. Additionally, the inclusion or exclusion of specific variables in the prompts can significantly impact the models’ predictions. These findings underscore the importance of aligning LLMs to more effectively model diverse public opinions while minimizing political biases and enhancing robustness in representativeness.

In this paper, we introduce the Akan Cinematic Emotions (AkaCE) dataset, the first multimodal emotion dialogue dataset for an African language, addressing the significant lack of resources for low-resource languages in emotion recognition research. AkaCE, developed for the Akan language, contains 385 emotion-labeled dialogues and 6162 utterances across audio, visual, and textual modalities, along with word-level prosodic prominence annotations. The presence of prosodic labels in this dataset also makes it the first prosodically annotated African language dataset. We demonstrate the quality and utility of AkaCE through experiments using state-of-the-art emotion recognition methods, establishing solid baselines for future research. We hope AkaCE inspires further work on inclusive, linguistically and culturally diverse NLP resources.

Multi-modal Large Language Models (MLLMs) integrating images, text, and speech can provide farmers with accurate diagnoses and treatment of pests and diseases, enhancing agricultural efficiency and sustainability. However, existing benchmarks lack comprehensive evaluations, particularly in multi-level reasoning, making it challenging to identify model limitations. To address this issue, we introduce Agri-CM3, an expert-validated benchmark assessing MLLMs’ understanding and reasoning in agricultural management. It includes 3,939 images and 15,901 multi-level multiple-choice questions with detailed explanations. Evaluations of 45 MLLMs reveal significant gaps. Even GPT-4o achieves only 63.64% accuracy, falling short in fine-grained reasoning tasks. Analysis across three reasoning levels and seven compositional abilities highlights key challenges in accuracy and cognitive understanding. Our study provides insights for advancing MLLMs in agricultural management, driving their development and application. Code and data are available at https://github.com/HIT-Kwoo/Agri-CM3.

We study abstraction in an emergent communication paradigm. In emergent communication, two artificial neural network agents develop a language while solving a communicative task. In this study, the agents play a concept-level reference game. This means that the speaker agent has to describe a concept to a listener agent, who has to pick the correct target objects that satisfy the concept. Concepts consist of multiple objects and can be either more specific, i.e. the target objects share many attributes, or more generic, i.e. the target objects share fewer attributes. We tested two directions of zero-shot generalization to novel levels of abstraction: When generalizing from more generic to very specific concepts, agents utilized a compositional strategy. When generalizing from more specific to very generic concepts, agents utilized a more flexible linguistic strategy that involves reusing many messages from training. Our results provide evidence that neural network agents can learn robust concepts based on which they can generalize using adaptive linguistic strategies. We discuss how this research provides new hypotheses on abstraction and informs linguistic theories on efficient communication.

Most discussions about Large Language Model (LLM) safety have focused on single-agent settings but multi-agent LLM systems now create novel adversarial risks because their behavior depends on communication between agents and decentralized reasoning. In this work, we innovatively focus on attacking pragmatic systems that have constrains such as limited token bandwidth, latency between message delivery, and defense mechanisms. We design a permutation-invariant adversarial attack that optimizes prompt distribution across latency and bandwidth-constraint network topologies to bypass distributed safety mechanisms within the system. Formulating the attack path as a problem of maximum-flow minimum-cost, coupled with the novel Permutation-Invariant Evasion Loss (PIEL), we leverage graph-based optimization to maximize attack success rate while minimizing detection risk. Evaluating across models including Llama, Mistral, Gemma, DeepSeek and other variants on various datasets like JailBreakBench and AdversarialBench, our method outperforms conventional attacks by up to 7×, exposing critical vulnerabilities in multi-agent systems. Moreover, we demonstrate that existing defenses, including variants of Llama-Guard and PromptGuard, fail to prohibit our attack, emphasizing the urgent need for multi-agent specific safety mechanisms.

Reward models (RMs) are crucial for the training and inference-time scaling up of large language models (LLMs). However, existing reward models primarily focus on human preferences, neglecting verifiable correctness signals which have shown strong potential in training LLMs. In this paper, we propose agentic reward modeling, a reward system that combines reward models with verifiable correctness signals from different aspects to provide reliable rewards. We empirically implement a reward agent, named RewardAgent, that combines human preference rewards with two verifiable signals: factuality and instruction following, to provide more reliable rewards. We conduct comprehensive experiments on existing reward model benchmarks and inference-time best-of-n searches on real-world downstream tasks. RewardAgent significantly outperforms vanilla reward models, demonstrating its effectiveness. We further construct training preference pairs using RewardAgent and train an LLM with the DPO objective, achieving superior performance on various NLP benchmarks compared to conventional reward models. Our codes are publicly released to facilitate further research.

We introduce Agentic Reasoning, a framework that enhances large language model (LLM) reasoning by integrating external tool-using agents. Agentic Reasoning dynamically leverages web search, code execution, and structured memory to address complex problems requiring deep research. A key innovation in our framework is the Mind-Map agent, which constructs a structured knowledge graph to store reasoning context and track logical relationships, ensuring coherence in long reasoning chains with extensive tool usage. Additionally, we conduct a comprehensive exploration of the Web-Search agent, leading to a highly effective search mechanism that surpasses all prior approaches. When deployed on DeepSeek-R1, our method achieves a new state-of-the-art (SOTA) among public models and delivers performance comparable to OpenAI Deep Research, the leading proprietary model in this domain. Extensive ablation studies validate the optimal selection of agentic tools and confirm the effectiveness of our Mind-Map and Web-Search agents in enhancing LLM reasoning. Our code and data are publicly available.

Large Language Models (LLMs) have achieved considerable performance across various agentic planning tasks. However, traditional approaches adopt a “flood irrigation” methodology that indiscriminately injects gold trajectories, external feedback, and domain knowledge into agent models. This practice overlooks the fundamental human cognitive principle of self-awareness - the ability to dynamically assess situational demands and strategically employ resources during decision-making. We propose Agentic Knowledgeable Self-awareness to address this gap, a novel paradigm enabling LLM-based agents to autonomously regulate knowledge utilization. Specifically, we propose KnowSelf, a data-centric approach that applies agents with knowledgeable self-awareness like humans. Concretely, we devise a heuristic situation judgement criterion to mark special tokens on the agent’s self-explored trajectories for collecting training data. Through a two-stage training process, the agent model can switch between different situations by generating specific special tokens, achieving optimal planning effects with minimal costs. Our experiments demonstrate that can outperform various strong baselines on different tasks and models with minimal use of external knowledge.

Digital agents capable of automating complex computer tasks have attracted considerable attention. However, existing agent methods exhibit deficiencies in their generalization and specialization capabilities, especially in handling open-ended computer tasks in real-world environments. Inspired by the rich functionality of the App store, we present AgentStore, a scalable platform designed to dynamically integrate heterogeneous agents for automating computer tasks. AgentStore allows the system to continuously enrich its capabilities and adapt to rapidly evolving operating systems. Additionally, we propose a novel core MetaAgent with the AgentToken strategy to efficiently manage diverse agents and utilize their specialized and generalist abilities for both domain-specific and system-wide tasks. Extensive experiments on three interactive real-world benchmarks demonstrate that AgentStore significantly expands the capability boundaries of agent systems in both generalization and specialization, underscoring its potential for developing the specialized generalist computer assistant.

Existing LLM-based agents have achieved strong performance on held-in tasks, but their generalizability to unseen tasks remains poor. Hence, some recent work focus on fine-tuning the policy model with more diverse tasks to improve the generalizability. In this work, we find that finetuning a reward model to guide the policy model is more robust than directly finetuning the policy model.Based on this finding, we propose AgentRM, a 8B generalizable reward model, to guide the policy model for effective test-time search.We comprehensively investigate three approaches to construct the reward model, including explicit reward modeling, implicit reward modeling and LLM-as-a-judge.We then use AgentRM to guide the answer generation with Best-of-N sampling and beam search.We show that AgentRM is robust to paraphrasings of task instructions and can generalize to unseen tasks that require novel optimal behavior.Through extensive evaluation across nine tasks spanning four categories, AgentRM enhances the non-finetuned 8B policy model by 8.8 points on average, surpassing the top general agent by 4.0.Moreover, it demonstrates weak-to-strong generalization, yielding greater improvement on more powerful policy models.As for the specializability, AgentRM can also boost a finetuned policy model and outperform the top specialized agent by 11.4 on three held-in tasks.Further analysis verifies its effectiveness in test-time scaling.We release the code and data at https://github.com/thunlp/AgentRM.

Large language models (LLMs) have emerged as a promising foundation to build generally-capable agents (LLM-based agents) that can handle multi-turn decision-making tasks across various environments. However, the community lacks a unified interactive framework that covers diverse environments for comprehensive evaluation of agents, and enables exploration and learning for their self-improvement. To address this, we propose AgentGym, a framework featuring 7 real-world scenarios, 14 environments, and 89 tasks for unified, real-time, and concurrent agent interaction. We construct expanded instruction set, high-quality trajectories, and comprehensive benchmarking suite for developing LLM-based agents. Moreover, AgentGym supports interactive exploration and learning for agents through multi-turn interactions and real-time feedback. Based on AgentGym, we take the initial step to develop LLM-based agents that can handle diverse tasks via methods like self-improvement or reinforcement learning. Experimental results show that the trained agents can achieve results comparable to commercial models. We hope our work can help the community develop more advanced LLM-based agents. We release the code, dataset, benchmark, and checkpoints at https://agentgym.github.io/.

Multi-agent systems (MAS) based on large language models (LLMs) have demonstrated significant potential in collaborative problem-solving. However, they still face substantial challenges of low communication efficiency and suboptimal task performance, making the careful design of the agents’ communication topologies particularly important. Inspired by the management theory that roles in an efficient team are often dynamically adjusted, we propose AgentDropout, which identifies redundant agents and communication across different communication rounds by optimizing the adjacency matrices of the communication graphs and eliminates them to enhance both token efficiency and task performance. Compared to state-of-the-art methods, AgentDropout achieves an average reduction of 21.6% in prompt token consumption and 18.4% in completion token consumption, along with a performance improvement of 1.14 on the tasks. Furthermore, the extended experiments demonstrate that AgentDropout achieves notable domain transferability and structure robustness, revealing its reliability and effectiveness. We release our code at https://github.com/wangzx1219/AgentDropout.

Current research in LLM-based simulation systems lacks comprehensive solutions for modeling real-world court proceedings, while existing legal language models struggle with dynamic courtroom interactions. We present **AgentCourt**, a comprehensive legal simulation framework that addresses these challenges through adversarial evolution of LLM-based agents. Our AgentCourt introduces a new adversarial evolutionary approach for agents called **AdvEvol**, which performs dynamic knowledge learning and evolution through structured adversarial interactions in a simulated courtroom program, breaking the limitations of the traditional reliance on static knowledge bases or manual annotations. By simulating 1,000 civil cases, we construct an evolving knowledge base that enhances the agents’ legal reasoning abilities. The evolved lawyer agents demonstrated outstanding performance on our newly introduced **CourtBench** benchmark, achieving a 12.1% improvement in performance compared to the original lawyer agents. Evaluations by professional lawyers confirm the effectiveness of our approach across three critical dimensions: cognitive agility, professional knowledge, and logical rigor. Beyond outperforming specialized legal models in interactive reasoning tasks, our findings emphasize the importance of adversarial learning in legal AI and suggest promising directions for extending simulation-based legal reasoning to broader judicial and regulatory contexts.

As Multimodal Large Language Models (MLLMs) advance, multimodal agents show promise in real-world tasks like web navigation and embodied intelligence. However, due to limitations in a lack of external feedback, these agents struggle with self-correction and generalization. A promising approach is to use reward models as external feedback, but there is no clear on how to select reward models for agents. Thus, there is an urgent need to build a reward bench targeted at agents. To address these challenges, we propose Agent-RewardBench, a benchmark designed to evaluate reward modeling ability in MLLMs. The benchmark is characterized by three key features: (1) Multiple dimensions and real-world agent scenarios evaluation. It covers perception, planning, and safety with 7 scenarios; (2) Step-level reward evaluation. It allows for the assessment of agent capabilities at the individual steps of a task, providing a more granular view of performance during the planning process; and (3) Appropriately difficulty and high-quality. We carefully sample from 10 diverse models, difficulty control to maintain task challenges, and manual verification to ensure the integrity of the data. Experiments demonstrate that even state-of-the-art multimodal models show limited performance, highlighting the need for specialized training in agent reward modeling. Code is available at github.

Code-switching is prevalent in multilingual communities but lacks adequate high-quality data for model development, especially for African languages. To address this, we present AfroCS-xs, a small human-validated synthetic code-switched dataset for four African languages (Afrikaans, Sesotho, Yoruba, isiZulu) and English within a specific domain—agriculture. Using large language models (LLMs), we generate code-switched sentences, including English translations, that are rigorously validated and corrected by native speakers. As a downstream evaluation task, we use this dataset to fine-tune different instruction-tuned LLMs for code-switched translation and compare their performance against machine translation (MT) models. Our results demonstrate that LLMs consistently improve in translation accuracy when fine-tuned on the high-quality AfroCS-xs dataset, highlighting that substantial gains can still be made with a low volume of data. We also observe improvements on natural code-switched and out-of-domain (personal finance) test sets. Overall, regardless of data size and prior exposure to a language, LLMs benefit from higher quality training data when translating code-switched texts in under-represented languages.

Large-scale multilingual evaluations, such as MEGA, often include only a handful of African languages due to the scarcity of high-qualityevaluation data and the limited discoverability of existing African datasets. This lack of representation hinders comprehensive LLM evaluation across a diverse range of languages and tasks. To address these challenges, we introduce AFROBENCH—a multi-task benchmark for evaluating the performance of LLMs across 64 African languages, 15 tasks and 22 datasets. AFROBENCH consists of nine natural language understanding datasets, six text generation datasets, six knowledge and question answering tasks, and one mathematical reasoning task. We present results comparing the performance of prompting LLMs to fine-tuned baselines based on BERT and T5-style models. Our results suggest large gaps in performance between high-resource languages, such as English, and African languages across most tasks; but performance also varies based on the availability of monolingual data resources. Our findings confirm that performance on African languages continues to remain a hurdle for current LLMs, underscoring the need for additional efforts to close this gap.

Recent advancements in large language model (LLM) performance on medical multiplechoice question (MCQ) benchmarks have stimulated interest from healthcare providers and patients globally. Particularly in low-andmiddle-income countries (LMICs) facing acute physician shortages and lack of specialists, LLMs offer a potentially scalable pathway to enhance healthcare access and reduce costs. However, their effectiveness in the Global South, especially across the African continent, remains to be established. In this work, we introduce AfriMed-QA , the first largescale Pan-African English multi-specialty medical Question-Answering (QA) dataset, 15,000 questions (open and closed-ended) sourced from over 60 medical schools across 16 countries, covering 32 medical specialties. We further evaluate 30 LLMs across multiple axes including correctness and demographic bias. Our findings show significant performance variation across specialties and geographies, MCQ performance clearly lags USMLE (MedQA). We find that biomedical LLMs underperform general models and smaller edge-friendly LLMs struggle to achieve a passing score. Interestingly, human evaluations show a consistent consumer preference for LLM answers and explanations when compared with clinician answers.

In this work, we present a manually annotated corpus for Adverse Event (AE) extraction from discharge summaries of elderly patients, a population often underrepresented in clinical NLP resources. The dataset includes 14 clinically significant AEs—such as falls, delirium, and intracranial haemorrhage, along with contextual attributes like negation, diagnosis type, and in-hospital occurrence. Uniquely, the annotation schema supports both discontinuous and overlapping entities, addressing challenges rarely tackled in prior work. We evaluate multiple models using FlairNLP across three annotation granularities: fine-grained, coarse-grained, and coarse-grained with negation. While transformer-based models (e.g., BERT-cased) achieve strong performance on document-level coarse-grained extraction (F1 = 0.943), performance drops notably for fine-grained entity-level tasks (e.g., F1 = 0.675), particularly for rare events and complex attributes. These results demonstrate that despite high-level scores, significant challenges remain in detecting underrepresented AEs and capturing nuanced clinical language. Developed within a Trusted Research Environment (TRE), the dataset is available upon request via DataLoch and serves as a robust benchmark for evaluating AE extraction methods and supporting future cross-dataset generalisation.

Current LLM pipelines account for only one possible tokenization for a given string, ignoring exponentially many alternative tokenizations during training and inference. For example, the Llama3 standard tokenization of penguin is [p,enguin], yet [peng,uin] is another perfectly valid alternative. In this paper, we show that despite LLMs being trained solely on one tokenization, they still retain semantic understanding of other tokenizations, raising questions about their implications in LLM safety. Put succinctly, we answer the following question: can we adversarially tokenize an obviously malicious string to evade safety and alignment restrictions? We show that not only is adversarial tokenization an effective yet previously neglected axis of attack, but it is also competitive against existing state-of-the-art adversarial approaches without changing the text of the harmful request. We empirically validate this exploit across three state-of-the-art LLMs and adversarial datasets, revealing a previously unknown vulnerability in subword models.