RE-TRAC: REcursive TRAjectory Compression for Deep Search Agents
Jialiang Zhu, Gongrui Zhang, Xiaolong Ma, Lin Xu, Miaosen Zhang, Ruiqi Yang, Song Wang, Kai Qiu, Zhirong Wu, Qi Dai, Ruichun Ma, Bei Liu, Yifan Yang, Chong Luo, Zhengyuan Yang, Linjie Li, Lijuan Wang, Weizhu Chen, Xin Geng, Baining Guo
2026-02-03
Summary
This paper introduces a new way to build AI research assistants powered by large language models, called Re-TRAC, that aims to make them more efficient and effective at complex tasks.
What's the problem?
Current AI research assistants, often built using a method called ReAct, struggle with long and complicated research projects. They tend to get stuck focusing on one path, repeat the same searches, and have trouble remembering what they've already learned or considering different approaches. Essentially, they lack a good way to step back, reflect on their progress, and plan strategically.
What's the solution?
The researchers developed Re-TRAC, which improves on ReAct by having the AI create a summary after each attempt at research. This summary doesn't just record *what* happened, but also what the AI was unsure about, where it failed, and what it plans to try next. This 'state representation' then guides the AI's next steps, allowing it to learn from past attempts and explore more intelligently. They also showed how to train smaller AI models to work even better with Re-TRAC.
Why it matters?
This is important because it makes AI research assistants significantly better at tackling complex problems. Re-TRAC uses fewer resources – fewer searches and less processing power – while achieving better results. This means we can build AI that can help with research more effectively and affordably, and it represents a step towards AI that can truly think and plan like a researcher.
Abstract
LLM-based deep research agents are largely built on the ReAct framework. This linear design makes it difficult to revisit earlier states, branch into alternative search directions, or maintain global awareness under long contexts, often leading to local optima, redundant exploration, and inefficient search. We propose Re-TRAC, an agentic framework that performs cross-trajectory exploration by generating a structured state representation after each trajectory to summarize evidence, uncertainties, failures, and future plans, and conditioning subsequent trajectories on this state representation. This enables iterative reflection and globally informed planning, reframing research as a progressive process. Empirical results show that Re-TRAC consistently outperforms ReAct by 15-20% on BrowseComp with frontier LLMs. For smaller models, we introduce Re-TRAC-aware supervised fine-tuning, achieving state-of-the-art performance at comparable scales. Notably, Re-TRAC shows a monotonic reduction in tool calls and token usage across rounds, indicating progressively targeted exploration driven by cross-trajectory reflection rather than redundant search.