What's this about?
Bilevel Coordinated Reflection models an orchestrator and workers as a coordination game, then studies how textual reflections become reusable memory. Its key result is an impossibility: a gate looking only at the transcript cannot always tell whether a proposed memory helps; evaluation must check effects in the environment.
What exactly is an “environment-grounded gate”?
It is a test that accepts a new lesson only after rerunning the task or measuring an external outcome. Think of a team retrospective: eloquent notes are not proof of improvement; the next delivery has to get better.
Why is this worth noting?
The proposed system reaches 72.2% on 500 SWE-bench instances versus 70.8% for a public mini-SWE-agent reference. More importantly, it offers a disciplined answer to a common agent-memory failure: confident but wrong reflections accumulating as permanent instructions.