LoongReflect: Boosting Long-Horizon Reflection in Search Agents via Global Perspective Distillation

TL;DR - LoongReflect trains search agents to make better long-horizon reflection and backtracking decisions by combining global teacher supervision with outcome-based reinforcement learning. It improves multi-hop retrieval and mathematical reasoning over outcome-only RL and self-distillation baselines.

• Models reflection as memory control over a reversible trajectory tree.
• Reflection updates working memory with verified facts, missing evidence, and branch risks.
• Backtracking removes unreliable branches while retaining concise corrective lessons.
• Coordinates targeted teacher distillation with trajectory-level GRPO using a look-ahead, extragradient-style mechanism.