TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios

Yuanzhe Shen, Zisu Huang, Zhengyuan Wang, Muzhao Tian, Zhengkang Guo, Chenyang Zhang, Shuaiyu Zhou, Zengjie Hu, Dailin Li, Kaimin Wang, Wenhao Liu, Tianlong Li, Feng Hong, Cao Liu, Ke Zeng
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:110995-111034, 2026.

Abstract

As LLM-based agents are deployed in increasingly complex real-world settings, existing benchmarks underrepresent key challenges such as enforcing global constraints, coordinating multi-tool reasoning, and adapting to evolving user behavior over long, multi-turn interactions. To bridge this gap, we introduce TRIP-Bench, a long-horizon benchmark grounded in realistic travel-planning scenarios. TRIP-Bench leverages real-world data, offers 18 curated tools and 40+ travel requirements, and supports automated evaluation. It includes splits of varying difficulty; the hard split emphasizes long and ambiguous interactions, style shifts, feasibility changes, and iterative version revision. Dialogues span up to 15 user turns, can involve 150+ tool calls, and may exceed 200k tokens of context. Experiments show that even advanced models achieve at most 50% success on the easy split, with performance dropping below 10% on hard subsets. We further propose GTPO, an online multi-turn reinforcement learning method with specialized reward normalization and reward differencing. Applied to Qwen2.5-32B-Instruct, GTPO improves constraint satisfaction and interaction robustness, outperforming Gemini-3-Pro in our evaluation. We expect TRIP-Bench to advance practical long-horizon interactive agents, and GTPO to provide an effective online RL recipe for robust long-horizon training.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-shen26j, title = {{TRIP}-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios}, author = {Shen, Yuanzhe and Huang, Zisu and Wang, Zhengyuan and Tian, Muzhao and Guo, Zhengkang and Zhang, Chenyang and Zhou, Shuaiyu and Hu, Zengjie and Li, Dailin and Wang, Kaimin and Liu, Wenhao and Li, Tianlong and Hong, Feng and Liu, Cao and Zeng, Ke}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {110995--111034}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/shen26j/shen26j.pdf}, url = {https://proceedings.mlr.press/v306/shen26j.html}, abstract = {As LLM-based agents are deployed in increasingly complex real-world settings, existing benchmarks underrepresent key challenges such as enforcing global constraints, coordinating multi-tool reasoning, and adapting to evolving user behavior over long, multi-turn interactions. To bridge this gap, we introduce TRIP-Bench, a long-horizon benchmark grounded in realistic travel-planning scenarios. TRIP-Bench leverages real-world data, offers 18 curated tools and 40+ travel requirements, and supports automated evaluation. It includes splits of varying difficulty; the hard split emphasizes long and ambiguous interactions, style shifts, feasibility changes, and iterative version revision. Dialogues span up to 15 user turns, can involve 150+ tool calls, and may exceed 200k tokens of context. Experiments show that even advanced models achieve at most 50% success on the easy split, with performance dropping below 10% on hard subsets. We further propose GTPO, an online multi-turn reinforcement learning method with specialized reward normalization and reward differencing. Applied to Qwen2.5-32B-Instruct, GTPO improves constraint satisfaction and interaction robustness, outperforming Gemini-3-Pro in our evaluation. We expect TRIP-Bench to advance practical long-horizon interactive agents, and GTPO to provide an effective online RL recipe for robust long-horizon training.} }
Endnote
%0 Conference Paper %T TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios %A Yuanzhe Shen %A Zisu Huang %A Zhengyuan Wang %A Muzhao Tian %A Zhengkang Guo %A Chenyang Zhang %A Shuaiyu Zhou %A Zengjie Hu %A Dailin Li %A Kaimin Wang %A Wenhao Liu %A Tianlong Li %A Feng Hong %A Cao Liu %A Ke Zeng %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-shen26j %I PMLR %P 110995--111034 %U https://proceedings.mlr.press/v306/shen26j.html %V 306 %X As LLM-based agents are deployed in increasingly complex real-world settings, existing benchmarks underrepresent key challenges such as enforcing global constraints, coordinating multi-tool reasoning, and adapting to evolving user behavior over long, multi-turn interactions. To bridge this gap, we introduce TRIP-Bench, a long-horizon benchmark grounded in realistic travel-planning scenarios. TRIP-Bench leverages real-world data, offers 18 curated tools and 40+ travel requirements, and supports automated evaluation. It includes splits of varying difficulty; the hard split emphasizes long and ambiguous interactions, style shifts, feasibility changes, and iterative version revision. Dialogues span up to 15 user turns, can involve 150+ tool calls, and may exceed 200k tokens of context. Experiments show that even advanced models achieve at most 50% success on the easy split, with performance dropping below 10% on hard subsets. We further propose GTPO, an online multi-turn reinforcement learning method with specialized reward normalization and reward differencing. Applied to Qwen2.5-32B-Instruct, GTPO improves constraint satisfaction and interaction robustness, outperforming Gemini-3-Pro in our evaluation. We expect TRIP-Bench to advance practical long-horizon interactive agents, and GTPO to provide an effective online RL recipe for robust long-horizon training.
APA
Shen, Y., Huang, Z., Wang, Z., Tian, M., Guo, Z., Zhang, C., Zhou, S., Hu, Z., Li, D., Wang, K., Liu, W., Li, T., Hong, F., Liu, C. & Zeng, K.. (2026). TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:110995-111034 Available from https://proceedings.mlr.press/v306/shen26j.html.

Related Material