Reproduce “TRIP-Bench: A Benchmark for Long-Horizon Interactive Agents in Real-World Scenarios” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuanzhe Shen, Zisu Huang, Zhengyuan Wang, Muzhao Tian, Zhengkang Guo, Chenyang Zhang, Shuaiyu Zhou, Zengjie Hu and 9 more
A new benchmark, TRIP-Bench, evaluates long-horizon interactive agents in real-world travel planning scenarios by incorporating complex global constraints and dynamic multi-turn user interactions. The proposed Group Turn-level Policy Optimization (GTPO) method improves agent performance, enabling a Qwen2.5-32B-Instruct model to achieve 40% on mid-difficulty loose tasks and outperform Gemini-3 Pro.
Key insights
- Explicit reasoning ("thinking") significantly improves agent performance on easier tasks, but current mechanisms are insufficient for fully correct and verifiable outcomes in the most challenging scenarios.
- Multi-turn interactions degrade global consistency on complex tasks with strong global constraints, as errors can accumulate, unlike single-turn interactions which maintain higher reliability.
- GTPO's Turn-wise Reward Differencing is critical for boosting complete pass rates by prioritizing incremental gains and preventing reward inheritance issues across turns.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.