RE-TRAC: REcursive TRAjectory Compression for Deep Search Agents

TL;DR

RE-TRAC tackles the limitations of linear ReAct-based deep research agents by introducing a recursive trajectory compression mechanism that summarizes evidence, uncertainties, and future plans after each rollout. This structured state is used to condition subsequent rollouts, enabling cross-trajectory reflection, broader exploration, and reduced redundancy in tool use and token consumption. Empirically, RE-TRAC yields consistent improvements on frontier-model benchmarks and enables strong performance for smaller models via supervised fine-tuning, while also serving as a training-free test-time scaling method. The approach advances autonomous search by turning exploration into a progressively informed, memory-enabled process with practical benefits for efficiency and scalability.

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.

Figures (11)

• Figure 1: Comparison of RE-TRAC with state-of-the-art agentic models. Our 4B and 30B models surpass the performance of significantly larger, state-of-the-art models.
• Figure 2: Performance of Pass@8 serves as the theoretical upper bound of the models.
• Figure 3: ReAct versus RE-TRAC framework. ReAct leads to premature convergence and forgotten branches in long-horizon tasks (left). RE-TRAC compresses experience from previous rollouts to systematically guide exploration in successive rounds (right).
• Figure 4: A comparative overview of the independent ReAct Pass@N (top) and our proposed RE-TRAC framework (bottom). Unlike the traditional ReAct paradigm, where multiple rollouts are executed in isolated silos without experience sharing, RE-TRAC is an iterative, trajectory-level framework. It employs a compression mechanism to distill analytical conclusions, evidence, and uncertainties from previous attempts. This compressed context is then propagated to successive rollouts, enabling the agent to recursively reflect on its trajectory and progressively improve its exploration strategy.
• Figure 5: Relationship between model accuracy and used tokens/tools with different TTS methods. Evaluated on BrowseComp300. Re-TRAC consistently achieves better performance with less resources used.