简体中文 · English

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

Core agent mechanism mini-suite. Runs a standalone planning, memory, credit assignment or topology path and records process counts on deterministic tasks, not the paper's full external environment.

Paper information¶

Field	Details
Paper	arXiv 2608.11967
Institution	Authors did not disclose affiliation
First public release	2026-08-12 (arXiv v1)
Author code	No public author code found as of 2026-08-24.
Local adapter or method key	`loongreflect`
Local reproduction code	`src/auto_research/agent_research/historical_b10_b11.py`
First author	Zhixin Zhang

Original paper summary¶

Background and main changes¶

Treat reflect and backtrack as memory-control actions on a reversible trajectory tree. Coordinate fast privileged-teacher supervision with slower outcome GRPO.

flowchart LR
  O[Task and observations] --> P[Planning and state management]
  P --> M[LoongReflect mechanism]
  M --> A[Actions and tools]
  A --> V[Verification and feedback]
  V --> P

Key figure from the paper¶

Figure 2: Core architecture, modules and connections. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ \mathcal L=\mathcal L_{reflect\text{-}OPD}+\lambda\mathcal L_{trajectory\text{-}GRPO} \]

Offline and online results¶

The paper consistently beats outcome-only RL and self-distillation on multi-hop RAG and math reasoning. The local mini-suite does not establish those benchmark or online results.

Local reproduction¶

On 120 PlanBench-mini episodes, loongreflect has joint success 1.0000 and average cost 0.7700; the long-context control has 1.0000 and 64.5000. With saturated accuracy, verify traces, sources and paper-specific counts instead of claiming gains. See per-paper metrics and batch index.

auto-research agent-research --method loongreflect --benchmark planbench-mini --episodes 120 --seed 42

Scope¶

The deterministic mini-suite does not access browsers, AppWorld, SWE-bench or external tool services. It checks mechanism branches, process state and reporting. Full environment results require the paper's resources. Runs go in runs/agent-research/; only metrics are committed. See experiment records.