Source-linked AI summary
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
TL;DR
Deep research agents built on linear ReAct workflows struggle to revisit states and explore alternative trajectories. Re-TRAC recursively compresses each trajectory into structured state representations for subsequent planning, achieving 15–20% absolute gains over ReAct on BrowseComp with frontier LLMs and state-of-the-art performance at comparable smaller-model scales.
Problem
Linear ReAct workflows make revisiting earlier reasoning states and branching into alternative search trajectories challenging, especially under long-context settings.
Method
Re-TRAC recursively compresses each trajectory into structured state representations that summarize evidence, uncertainties, failures, and plans for subsequent trajectories.
Results
15–20% absolute gains over ReAct on BrowseComp with frontier LLMs, while Re-TRAC-aware SFT achieves state-of-the-art performance at comparable smaller-model scales.
Takeaways & Limitations
Re-TRAC supports more precise, lower-overhead exploration through cross-trajectory knowledge consolidation and conditioned planning.
Takeaways & Limitations
A reported limitation is that standard ReAct execution can neglect early planned branches.
Abstract
from arXiv · showhide
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.
1. Introduction
Re-TRAC addresses the limitations of linear ReAct workflows by conditioning successive search trajectories on structured summaries of evidence, uncertainties, failures, and plans. It improves BrowseComp performance over ReAct and extends strong results to smaller models through supervised fine-tuning.
- Limitations of ReAct: ReAct-based agents append reasoning and tool calls linearly, limiting their ability to revisit earlier states and explore alternative search directions.The paper analyzes these inherent limitations despite trained behaviors such as backtracking and self-reflection.
- Motivation: Higher pass@k than pass@1 in existing deep research models motivates explicitly guiding agents toward previously unexplored search trajectories.Repeated inference produces diverse reasoning trajectories, indicating that independent exploration can expose solutions missed by a single trajectory.
- Re-TRAC framework: Re-TRAC recursively constructs a structured state representation after each trajectory and uses it to condition subsequent trajectories.Each representation summarizes accumulated evidence, unresolved uncertainties, identified failure modes, and a forward-looking research plan.
- Empirical results: 15–20% absolute gains over ReAct are achieved on BrowseComp with frontier LLMs.These results demonstrate the benefit of cross-trajectory exploration on the benchmark.
- Empirical results: 53% accuracy is achieved by the fine-tuned 30B model on BrowseComp, while the 4B model reaches 30%.The models establish state-of-the-art performance among models of comparable scale.
2. Related Work
Related work situates Deep Research Agents as autonomous systems for long-horizon reasoning, strategic planning, and persistent tool use. It emphasizes context management and test-time compute scaling as key directions for improving such agents.
- Deep Research Agents: Deep Research Agents extend information retrieval with long-horizon reasoning, strategic planning, and persistent tool utilization.The passage frames this development as a transition toward autonomous systems and cites proprietary-model agents as examples.
- Context Management: Research on context management divides into intrinsic context optimization and external memory mechanisms for state maintenance.The passage identifies these as the two general categories for managing context in long-horizon tasks.
- Test-Time Compute Scaling: Recent scaling paradigms increasingly use test-time compute rather than only larger models and training datasets.The passage identifies expanded internal reasoning, including Chain-of-Thought extensions and reasoning models, as the dominant test-time scaling approach.
3. Motivation
The motivation identifies incomplete exploration and inefficient use of multiple trajectories as key limitations of linear deep-research agents. ReAct’s sequential execution and lack of communication can cause premature convergence, redundant exploration, and failure to synthesize global solutions.
- Incomplete Branch Exploration: Up to 93% of failed trajectories contain branches that the model planned but ultimately failed to explore.This incomplete branch exploration is identified as a common failure phenomenon in advanced deep research agents.
- Incomplete Branch Exploration: Linear ReAct execution inhibits strategic branching, backtracking, and re-evaluation across long trajectories with interdependent tool calls.The paper links this constraint to catastrophic forgetting within trajectories spanning hundreds of thousands of tokens.
- The Potential from Multiple Trials: The gap between Pass@1 and Pass@8 reveals substantial performance potential that single trajectories fail to reach.Multiple stochastic trials expose an untapped exploration ceiling under the Pass@K metric.
- The Potential from Multiple Trials: Independent trials such as Majority Voting and Best-of-N repeat exploration and prevent cross-trajectory experience sharing for global synthesis.These inefficiencies motivate a trajectory-level recursive framework that compresses previous trajectories instead of restarting each attempt from scratch.
4. Method: Re-TRAC Framework
Re-TRAC recursively compresses each research trajectory into a structured state that carries conclusions, verified evidence, uncertainties, failures, and candidate branches into subsequent rollouts. This cross-trajectory context broadens search coverage while reducing redundant exploration and preserving branching diversity.
- Core framework: Re-TRAC compresses each rollout into a structured state representation and propagates it across successive rollouts.The state is updated after each trajectory according to a fixed compression specification and conditions the next rollout.
- Core framework: The state records analytical conclusions, evidence provenance and verification status, plus unresolved hypotheses, candidate branches, failures, and discarded directions.These facets provide reusable reasoning anchors, prevent repeated checking, and expose unexplored search space.
- Recursive search: Re-TRAC preserves multiple unresolved candidates and actionable options, preventing exploration from collapsing into a single path while filtering low-level trace details.This balances focused guidance with controlled branching diversity across recursive rounds.
- Recursive search: Re-TRAC improves coverage by preserving incomplete branches for later execution and reduces redundancy by avoiding repeated tool calls for verified facts.Independent ReAct rounds may instead re-explore the same path and neglect early planned branches.
- Execution: The method is training-free at inference time, beginning with a standard ReAct rollout and using its compressed state as the next round’s initial user message.The default maximum is N = 8 rounds, and performance is benchmarked against Single Run, Best-of-N, Majority Voting, and Weighted Voting.
5. Experiments
RE-TRAC achieves state-of-the-art or competitive results across multiple deep-research benchmarks, including gains over same-size baselines and competitiveness with much larger and closed-source models. Test-time scaling, SFT, and summary-use strategies further improve performance while reducing resource use or expanding smaller-model capabilities.
- Benchmark Performance: RE-TRAC-30B-A3B improves 8%∼10% over Tongyi-DeepResearch on BrowseComp, BrowseComp-ZH, GAIA, and XBench.Tongyi-DeepResearch is the previous strongest 30B-parameter baseline.
- Benchmark Performance: RE-TRAC-4B achieves the best performance across benchmarks among baselines with fewer than 15B parameters.The result establishes state-of-the-art performance among same-size baselines.
- Competitive Performance Against Larger Models: The 30B model beats MiniMAX-M2-229B on every benchmark except HLE and reaches 53% accuracy on BrowseComp versus 52% for GLM-4.7-358B.These results demonstrate competitiveness against substantially larger models.
- Exceed Closed-Source Models: RE-TRAC-30B-A3B beats all closed-source baselines on GAIA and ranks second on BrowseComp and BrowseComp-ZH.The finding shows that the framework can make a small open-source model competitive with advanced closed-source systems.
- Test-Time Scaling: Re-TRAC achieves the best or competitive performance across test-time scaling methods, and all evaluated models benefit from Re-TRAC.Advanced models also gain from Majority Voting, Weighted Voting, and Best-Of-N, whereas GLM-4.7 does not obtain comparable gains from Majority Voting and Best-Of-N.
- Efficiency and Training: Re-TRAC uses only 50% resources while achieving better performance as scaling increases, and SFT enables a 4B model to achieve comparable or better results than large-scale reinforcement-learning agents.State inheritance and search-space convergence reduce redundant exploration; SFT teaches basic search and state-representation strategies.
- Ablations: Free-use summary instructions improve performance in every round, while a stronger summarizer improves the 4B model but not the 30B model.The findings address search-path stagnation and indicate weaker summarization ability in the 4B model.
6. Conclusion
Re-TRAC extends ReAct with recursive, experience-based exploration that consolidates knowledge across trajectories for more precise, lower-overhead deep research. Its gains in frontier models and successful smaller-model SFT underscore the value of structured memory and conditioned planning.
- Re-TRAC moves beyond linear ReAct through recursive, experience-based exploration for deep research agents.The framework is designed around cross-trajectory exploration rather than a single linear trajectory.
- Cross-trajectory knowledge consolidation helps agents navigate complex search spaces with higher precision and lower computational overhead.
- Consistent gains in frontier models and successful supervised fine-tuning for smaller models highlight structured memory and conditioned planning.
A. Analysis Details of Incomplete Branch Exploration
Section 3 analyzes incomplete branch exploration, finding that most deep research agents fail to explore all proposed branches. The analysis collects trajectories with incorrect final answers and uses GPT-5 to classify them.
- A. Analysis Details of Incomplete Branch Exploration: Most deep research agents fail to explore all proposed branches.The analysis collects trajectories with incorrect final answers and prompts GPT-5 to classify them; the prompt appears in Figure 6.
B. Training Details
The training procedure distills two smaller models from GLM-4.7 using four-round Re-TRAC trajectories on 33k InfoAgent questions, yielding 132k raw training samples before filtering.
- Distillation setup: Qwen3-4B-Instruct and Tongyi-DeepResearch-30B-A3B are distilled from GLM-4.7.The distillation uses GLM-4.7-generated solutions as supervision.
- Data generation: 33k InfoAgent questions are solved with Re-TRAC using 4 rounds.Each question is processed through the framework’s four-round trajectory procedure.
- Sample construction: 132k raw samples result because each four-round trajectory is flattened into 4 independent training samples.The rounds have independent contexts, enabling one sample per round.
C. Evaluation Details … C.3.4. MODEL-SPECIFIC HYPER-PARAMETERS
The evaluation specifies model-adapted web tools, answer verification, and Re-TRAC’s structured cross-trajectory state mechanism. It also defines state variants, continuation conditioning, round limits, and model-specific inference settings.
- C.1. Tools: The evaluation uses search and visit tools whose interfaces are adapted to each model’s native function-calling conventions.Adaptation covers parameter names, input types, and output formats.
- C.1.1. SEARCH TOOL: Search returns five relevant Google Search Web API results per query, including each page’s title, URL, and text snippet.The evaluation applies no additional processing or filtering to search results.
- C.1.2. VISIT TOOL: The visit tool fetches URLs, extracts page text, and uses GPT-4o-mini to produce goal-relevant structured summaries.HTML extraction uses Trafilatura, while PDF content is extracted page by page.
- C.1.2. VISIT TOOL: For Tongyi and GLM models, visit-tool interfaces use model-specific names, parameters, input cardinality, and condensed output formats.Tongyi uses “url” and explicit evidence and summary sections; GLM uses “open,” single-string “url,” and “pattern.”
- C.2. Verifier: Final-answer correctness is assessed following BrowseComp by comparing each model answer with ground truth using OpenAI o4-mini.The verifier receives the question, ground truth, and model’s final answer.
- C.3.1. STRUCTURED STATE REPRESENTATION: Re-TRAC appends a structured state representation after each trajectory, capturing answers, evidence, conclusions, sources, and uncertainties.The base version contains five complementary facets, with source annotations and explicitly linked conclusions.
- C.3.1. STRUCTURED STATE REPRESENTATION: Frontier LLMs receive three additional audit facets: failed attempts, uncompleted proposals, and discarded possibilities.The prompt is applied when the model either provides a final answer or reaches the context limit.
- C.3.2. CONTINUATION PROMPT: Each subsequent round is conditioned on the previous structured state representation, prepended alongside a continuation prompt.This prompt guides the model’s use of consolidated cross-trajectory information.
C.4. Evaluation Results
This section evaluates Re-TRAC using test-time scaling metrics and the additional Accuracy Prefix metric. AP@N measures early-round correctness and bounds achievable performance under answer selection.
- Metrics: The evaluation uses RT@N, MV@N, WV@N, Best@N, and Pass@N as test-time scaling metrics.These methods are defined in Section 5.2.
- Metrics: AP@N measures whether at least one correct answer appears among the first N rounds.It is similar to Pass@N but focuses on the first N rounds.
- Metrics: AP@N serves as an upper bound for Re-TRAC performance and quantifies remaining room for answer-selection improvement.The metric is specific to evaluating Re-TRAC in this appendix.
C.4.1. PER-MODEL DETAILED RESULTS
This section presents detailed eight-round BrowseComp300 evaluations for five models, reporting per-round accuracy alongside cumulative and aggregation-based metrics. Results are organized by model across Tables 9–13.
- Evaluation setup: Eight-round evaluations report Acc%, Pass@N, RT@N, AP@N, MV@N, WV@N, and Best@N for each model.The reported metrics cover individual-round accuracy, cumulative Re-TRAC performance, prefix accuracy, voting methods, and Best-of-N.
- Model-specific results: o4-mini, o3, GPT-5, DeepSeek-V3.2, and GLM-4.7 are presented in Tables 9, 10, 11, 12, and 13, respectively.All five model-specific result tables evaluate performance on BrowseComp300.