CoEvolve
Collection Summary
Agent rollout에서 forgetting and uncertainty signal을 추출하고 failure-prone interaction pattern을 바탕으로 새 task를 합성해 agent와 data distribution을 함께 진화시킨다.
Evolution History Store
- **Uses rollout history**: yes, rollout trajectories are the feedback substrate.
- **Stored form**: interaction trajectories plus forgetting/uncertainty signals and validated synthesized tasks.
- **Durable buffer status**: an explicit replay-buffer retention policy is not established in the abstract; the evolving training distribution acts as the persistent derivative artifact.
- **Security relevance**: uncertainty spoofing, forgetting-signal manipulation, task synthesis poisoning, environment-validation compromise, feedback-loop amplification.