AgentBench: Evaluating LLMs as Agents

By Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Yuxian Gu, Han Ding, Kai Men and 14 more. International Conference on Learning Representations, 2023

1,063 citations recorded by the source corpus.

The potential of Large Language Model (LLM) as agents has been widely acknowledged recently. Thus, there is an urgent need to quantitatively \textit{evaluate LLMs as agents} on challenging tasks in interactive environments. We present AgentBench, a multi-dimensional benchmark that consists of 8 distinct environments to assess LLM-as-Agent's reasoning and decision-making abilities. Our extensive test over \num API-based and open-sourced (OSS) LLMs shows that, while top commercial LLMs present a strong ability of acting as agents in complex environments, there is a significant disparity in performance between them and many OSS competitors that are no larger than 70B. We identify the typical reasons of failures in environments and LLMs, showing that poor long-term reasoning, decision-making, and instruction following abilities are the main obstacles for developing usable LLM agents. Improving instruction following and training on high quality multi-round alignment data could improve agent performance. And different from existing assumptions, training on code present ambivalent impacts on different agent tasks. Datasets, environments, and an integrated evaluation package for AgentBench are released at https://github.com/THUDM/AgentBench.

Subfields: Artificial Intelligence

View Paper: https://www.semanticscholar.org/paper/5dbf93a68b7fda600521f046dea35ea8ba9e884f

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Graham Neubig
2. Hannaneh Hajishirzi
3. Ji-rong Wen
4. Wayne Xin Zhao
5. Heng Ji
6. Mohit Bansal
7. Hao Peng
8. Huan Sun

Citations in the corpus

• Unifying Large Language Models and Knowledge Graphs: A Roadmap (2023, methodology): “Instead using two encoders to fuse the knowledge, LLMs can also be treated as agents to interact with the KGs to conduct reasoning [246], as illustrated in Fig.”
• VideoAgent: Long-form Video Understanding with Large Language Model as Agent (2024, neutral): “These models have demonstrated great success across various scenarios, such as online search, card game playing, and database management [25, 26, 61].”
• AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges (2025, neutral): “This progression positions Agentic AI as the next stage of AI infrastructure capable not only of executing predefined workflows but also of constructing, revising, and managing complex objectives across agents with minimal human supervision.”
• AgentTuning: Enabling Generalized Agent Abilities for LLMs (2023, neutral): “After conducting a series of experiments (Cf Appendix C.1), we settled on a sampling ratio of 1:4 between GPT-4 and GPT-3.”
• GPT-4V(ision) is a Generalist Web Agent, if Grounded (2024, methodology): “Existing work (Deng et al., 2023; Liu et al., 2023d) primarily uses large language models (LLMs) such as GPT-4 (OpenAI, 2023) on the raw HTML input.”
• A survey on large language model based autonomous agents (2023, positive): “AgentBench [172] provides a comprehensive framework for evaluating LLMs as autonomous agents across diverse environments.”
• The rise and potential of large language model based agents: a survey (2025, neutral): “Effectively and objectively evaluating LLMs agents presents significant challenges, despite their excellent performance in areas such as standalone operation, collective cooperation, and human interaction [70,484].”
• MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback (2023, background): “Existing work on interaction evaluation mostly focuses on a specific task or dimension, like task completion (Liu et al., 2023c), code generation (Yang et al., 2023a), human-LLM collaborative task solving (Lee et al., 2022b; Huang et al., 2023b; Fu et al., 2023), tool manipulation (Tang et al.,…”
• τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (2024, background): “Existing benchmarks [27, 29, 12, 14, 16] for language agents often feature simplified instruction-following setups, where the agent autonomously interacts with an environment (web, code terminal, or APIs) given all the information upfront, without any human-in-the-loop interaction and without the…”
• Understanding the planning of LLM agents: A survey (2024, background): “Popular interactive programming environments involve issues related to operating systems, databases, etc., such as Agent Bench [Liu et al. , 2023c], MiniWoB++ [Kim and others, 2023].”
• OpenHands: An Open Platform for AI Software Developers as Generalist Agents (2024, positive): “AgentBench (Liu et al., 2023) evaluates agents’ reasoning and decision-making abilities in a multi-turn, open-ended generation setting.”
• GAIA: a benchmark for General AI Assistants (2023, background): “ToolQA (Zhuang et al., 2023) or Gentopia (Xu et al., 2023a) for example combine existing datasets with human annotations (MMLU, MATH, etc.) at the risk of contamination during training, and without ensuring tool usage is actually tested.”
• Empowering Biomedical Discovery with AI Agents (2024, neutral): “As a result, model performance degrades on task variants that deviate from the assumptions made during training [175].”
• Improving Alignment and Robustness with Circuit Breakers (2024, neutral): “LLM agents have been benchmarked [35, 45], but their safety and robustness remain unexplored.”
• Training Language Models to Self-Correct via Reinforcement Learning (2024, neutral): “The most prominent settings include problems where external input tokens from an environment are available, such as agentic tasks (Liu et al., 2023), code repair (Jain et al., 2024), and tool use (Chen et al., 2023).”
• τ2-Bench: Evaluating Conversational Agents in a Dual-Control Environment (2025, neutral): “Following a long line of research into language agents [25, 27, 10, 13, 21], LLM tool use [24, 20, 9], and task-oriented dialog [5, 4, 1, 22, 6, 7, 8], τ -bench [26] is a recently introduced benchmark to measure the reliability of language agents in multi-turn task-oriented conversations such as…”
• MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts (2023, background): “…(Methani et al., 2020; Masry et al., 2022; Kahou et al., 2017; Chang et al., 2022; Kafle et al., 2018), documents (text in images) (Singh et al., 2019; Mathew et al., 2022; Liu et al., 2023c), or synthetic images (Dahlgren Lindström & Abraham, 2022; Li et al., 2023; Bitton-Guetta et al., 2023).”
• AI deception: A survey of examples, risks, and potential solutions (2023, neutral): “To illustrate, Liu et al. 61 and Kinniment et al. 62 measured different LLMs’ ability to autonomously carry out various tasks, such as browsing the web, online shopping, making a phone call, and using a computer’s operating system.”