Elamur Research
☰
home / papers / 2024 / $\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

$\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2406.12045

Summary

Most agent benchmarks give the model a complete instruction upfront. They rarely test two things real deployments need: holding a multi-turn conversation with a person, and obeying domain-specific rules. τ-bench addresses this gap with two customer-service domains, τ-retail and τ-airline. In each, an agent gets API tools over hidden databases and a policy document, and it must help a user played by a language model. The authors built the domains in three steps: they designed the schemas and APIs by hand, generated database entries with LM assistance, and wrote and checked user scenarios manually. Success is judged by whether the final database state matches the annotated target state. The paper also introduces pass^k, which estimates how likely an agent is to solve the same task in all of k independent runs. Function-calling agents, even gpt-4o, solve well under half of the tasks, and their reliability drops sharply as k grows. These results suggest that current LM agents are not yet consistent or rule-abiding enough for real-world deployment.

Key points

  • Each task is formalized as a POMDP in which the agent works with two parties: databases accessed through deterministic Python API functions, and an LM-simulated user (gpt-4-0613). The simulated user ends the episode by emitting "###STOP###".
  • gpt-4o with function calling reaches about 61% pass^1 on τ-retail and roughly a third on τ-airline. Its pass^8 on τ-retail falls to about 25%.
  • In an ablation that removes the domain policy from the prompt, gpt-4o barely changes on retail (61.2 → 56.8) but collapses on airline (33.2 → 10.8). gpt-3.5-turbo changes little in either domain, which suggests it hardly uses the complex airline rules at all.
  • The main failure types are:
    • faulty reasoning over database contents;
    • misapplying ad-hoc policies;
    • only partly completing compound requests, since tasks needing more database writes are harder.
  • Code and data are released at github.com/sierra-research/tau-bench.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: AI & machine learning

Cited by

116 papers of our library cite it (of the 140673 whose bibliographies we read); 4 of them have a page here, 112 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2406.12045 · arXiv 2406.12045

@misc{yao2024tau,
  title = {{\$tau\$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains}},
  author = {Yao, Shunyu and Shinn, Noah and Razavi, Pedram and Narasimhan, Karthik},
  year = {2024},
  doi = {10.48550/arxiv.2406.12045},
  eprint = {2406.12045},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2406.12045}
}

References

29 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. M. Ahn, A. Brohan, N. Brown, Y. Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausman, et al. Do as I can, not as I say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. URL https://arxiv.org/abs/2204.01691. ↗ our card
  2. J. Andreas, J. Bufe, D. Burkett, C. Chen, J. Clausman, J. Crawford, K. Crim, J. DeLoach, L. Dorner, J. Eisner, et al. Task-oriented dialogue as dataflow synthesis. Transactions of the Association for Computational Linguistics, 8:556–571, 2020. find
  3. P. Budzianowski, T.-H. Wen, B.-H. Tseng, I. Casanueva, S. Ultes, O. Ramadan, and M. Gašić. Multiwoz–a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling. arXiv preprint arXiv:1810.00278, 2018. find
  4. D. Chen, H. Chen, Y. Yang, A. Lin, and Z. Yu. Action-based conversations dataset: A corpus for building more in-depth task-oriented dialogue systems. arXiv preprint arXiv:2104.00783, 2021. find
  5. M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. de Oliveira Pinto, J. Kaplan, H. Edwards, Y. Burda, N. Joseph, G. Brockman, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry, P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Winter, P. Tillet, F. P. Such,… find our card
  6. S. Chen, S. Wiseman, and B. Dhingra. Chatshop: Interactive information seeking with language agents. arXiv preprint arXiv:2404.09911, 2024. — [arXiv:2404.09911](https://arxiv.org/abs/2404.09911) find
  7. S. eun Yoon, Z. He, J. M. Echterhoff, and J. McAuley. Evaluating large language models as generative user simulators for conversational recommendation, 2024. find
  8. I. Gür, D. Hakkani-Tür, G. Tür, and P. Shah. User modeling for task oriented dialogues. In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 900–906, 2018. doi: 10.1109/SLT.2018.8639652. — [doi:10.1109/slt.2018.8639652](https://doi.org/10.1109/slt.2018.8639652) find
  9. H. He, D. Chen, A. Balakrishnan, and P. Liang. Decoupling strategy and generation in negotiation dialogues. arXiv preprint arXiv:1808.09637, 2018. find
  10. Z. Hu, Y. Feng, A. T. Luu, B. Hooi, and A. Lipani. Unlocking the potential of user feedback: Leveraging large language model as user simulators to enhance dialogue system. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM ’23. ACM, Oct. 2023. doi:… — [doi:10.1145/3583780.3615220](https://doi.org/10.1145/3583780.3615220) find
  11. Y. Huang, J. Shi, Y. Li, C. Fan, S. Wu, Q. Zhang, Y. Liu, P. Zhou, Y. Wan, N. Z. Gong, et al. Metatool benchmark for large language models: Deciding whether to use tools and which to use. arXiv preprint arXiv:2310.03128, 2023. find
  12. C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. Narasimhan. Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770, 2023. find our card
  13. M. Kim, Y. Jung, D. Lee, and S.-w. Hwang. Plm-based world models for text-based games. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1324–1341, 2022. find
  14. X. Liu, H. Yu, H. Zhang, Y. Xu, X. Lei, H. Lai, Y. Gu, H. Ding, K. Men, K. Yang, et al. Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688, 2023. find our card
  15. J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442, 2023. find
  16. Y. Ruan, H. Dong, A. Wang, S. Pitis, Y. Zhou, J. Ba, Y. Dubois, C. J. Maddison, and T. Hashimoto. Identifying the risks of lm agents with an lm-emulated sandbox. arXiv preprint arXiv:2309.15817, 2023. find
  17. J. Schatzmann, D. Jurafsky, M. Galley, and D. Trevillian. Evaluating agenda-based user simulation for reinforcement learning of dialogue management. In Speech Communication, volume 47, pages 95–121, 2007. find
  18. T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761, 2023. find our card
  19. N. Shinn, B. Labash, and A. Gopinath. Reflexion: an autonomous agent with dynamic memory and self-reflection, 2023. find
  20. T. R. Sumers, S. Yao, K. Narasimhan, and T. L. Griffiths. Cognitive architectures for language agents. arXiv preprint arXiv:2309.02427, 2023. find
  21. Q. Wu, G. Bansal, J. Zhang, Y. Wu, S. Zhang, E. Zhu, B. Li, L. Jiang, X. Zhang, and C. Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. find our card
  22. Q. Xu, F. Hong, B. Li, C. Hu, Z. Chen, and J. Zhang. On the tool manipulation capability of open-source large language models, 2023. find
  23. F. Yan, H. Mao, C. C.-J. Ji, T. Zhang, S. G. Patil, I. Stoica, and J. E. Gonzalez. Berkeley function calling leaderboard. https://gorilla.cs.berkeley.edu/blogs/8\_berkeley\_ function\_calling\_leaderboard.html, 2024. ↗
  24. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. ReAct: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. find our card
  25. S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y. Cao, and K. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models, 2023. find
  26. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. React: Synergizing reasoning and acting in language models, 2023. find our card
  27. S. Yao, H. Chen, J. Yang, and K. Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. In ArXiv, volume 35, pages 20744–20757, preprint. find our card
  28. E. Zhang, X. Wang, P. Gong, Y. Lin, and J. Mao. Usimagent: Large language models for simulating search users. arXiv preprint arXiv:2403.09142, 2024. — [arXiv:2403.09142](https://arxiv.org/abs/2403.09142) find
  29. S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, Y. Bisk, D. Fried, U. Alon, et al. WebArena: A Realistic Web Environment for Building Autonomous Agents. arXiv preprint arXiv:2307.13854, 2023. find our card