gentaiscool/test_small

T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains

T1-Bench is a high-fidelity benchmark for evaluating task-completion and role-playing agents across 25 domains, including 11 single-domain and 14 multi-domain settings. It provides 76 tools and extensive human annotations, enabling systematic evaluation of agents in realistic, policy-grounded multi-domain interactions with natural user–assistant role-playing.

T1-Bench is a fully automated benchmark for evaluating the tool-calling capabilities of conversational AI agents across diverse service domains in task-oriented settings. The framework simulates end-to-end user–agent interactions without requiring human intervention at inference time, where a User Agent generates realistic customer utterances conditioned on predefined task goals and a tool-augmented Assistant Agent responds by invoking domain-specific tools/APIs and producing outputs grounded in tool results, prior conversational context, and domain-specific policies. Designed to capture the sequential and interactive nature of real-world service workflows across multi-domain scenarios, T1-Bench requires agents to maintain conversational state, reason over prior tool outputs, and execute multi-step operations such as search, filtering, booking, modification, and cancellation. Because all interactions are grounded in deterministic datasets and executable tools, the benchmark enables reproducible and fine-grained evaluation of agent behavior, tool-use decisions, and task completion performance.

The paper is accepted at EMNLP Findings 2026. Check our ArXiv.

Dataset structure

Each row has the following columns:

• domain_id: the domain name concatenated with a row index (e.g. attraction-0).
• domain: the domain name, matching the source folder name.
• user_template: the persona, user profile, and goal used to drive the user simulator.
• benchmark_success_criteria: a JSON-stringified object describing the exact-match tool calls and LLM-as-judge metrics used to score the conversation.

Domains

11 single-domain configs: attraction, bar, cafe, cruise, dessert, flight, food&dining, hotel, kayak, liveshow, vehicle.

14 multi-domain configs, prefixed with multi_Xdomain_ where X is the number of entities combined: multi_2domain_attraction_cafe, multi_2domain_attraction_dessert, multi_2domain_food&dining_hotel, multi_2domain_hotel_vehicle, multi_2domain_liveshow_food&dining, multi_2domain_vehicle_attraction, multi_3domain_attraction_food&dining_cafe, multi_3domain_flight_hotel_bar, multi_3domain_hotel_attraction_vehicle, multi_3domain_vehicle_food&dining_hotel, multi_4domain_flight_hotel_attraction_vehicle, multi_4domain_hotel_vehicle_food&dining_bar, multi_8domain, multi_11domain.

Results

Code

Please refer to the official implementation repository: https://github.com/CapitalOne-Research/t1-bench

Contact

E-mail: Genta Indra Winata or Amartya Chakraborty.

License

The dataset is licensed under CC-BY-SA 4.0.

Citation

If you find this dataset useful, please cite the following work

@article{winata2026t1,
  title={T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains},
  author={Winata, Genta Indra and Chakraborty, Amartya and Lin, Yuzhen and Rao, Swasthi P and Siingh, Shikhhar and Lu, Houhan and Bathaee, Nadia and Hatwar, Sriharsha and Dashore, Paresh and Jain, Anmol and others},
  journal={arXiv preprint arXiv:2606.11070},
  year={2026}
}