ReTree Rolls Back the Search Tree When Evidence Conflicts, Raising Qwen3-8B’s Overall Answer Accuracy From 30.1% to 44.0%

A team at Shanghai Jiao Tong University introduced [ReTree](https://arxiv.org/abs/2608.10676) on August 11 to address two related problems in long-running search agents: retaining complete ReAct trajectories causes the context to grow continuously, while repeated summarization saves tokens but may preserve only conclusions and lose the passages from which they were derived. Once a later search overturns an earlier fact, agents typically replace only that fact, leaving queries and conclusions based on the outdated premise in memory.

ReTree saves each state update as a tree node. Each node contains a task summary of up to 140 characters, atomic evidence with stable identifiers and URLs, and a revision log. At each step, the model sees only the current summary and the five pieces of evidence with the highest lexical relevance. If new information conflicts with existing evidence concerning the same entity, attribute, scope, and time period, a second adjudication step examines the sources and revision history. Once the conflict is confirmed, the system returns to the node where the erroneous information was first introduced, replaces the evidence, rebuilds the summary, deletes all descendants, and resumes searching from the corrected state.

Using Qwen3-8B, a maximum of eight Google searches, and the same question order, the team evaluated 2,149 questions across Bamboogle, 2WikiMultiHopQA, [HotpotQA](https://hotpotqa.github.io/), and FRAMES. ReTree achieved an overall accuracy of 44.0% according to a GPT-5 judge, compared with 30.1% for full-trajectory ReAct; exact-match accuracy was 28.0% versus 20.6%. The average maximum strategy-memory context per question was 1,190 characters, compared with 1,677 for the baseline. HotpotQA specifically requires finding answers and supporting facts across multiple documents, making it well suited to assessing whether provenance chains are preserved during multi-hop reasoning.

This design upgrades agent memory from a mutable block of text to rollback-capable dependency state, making it suitable for research agents that must handle updated information, contradictory sources, or auditable citations. However, the current system relies on an LLM to identify conflicts and on GPT-5 for evaluation, while all primary runs use only one model and a single seed. ReTree also treats ancestor relationships between nodes as semantic dependencies: pruning an entire subtree is safer, but may repeat unrelated searches. Before real-world deployment, the system still needs fine-grained dependency tracking, conflict-detection tests, and comprehensive reporting on API costs and latency.