AgentEvolver

Collection Summary

Self-questioning, self-navigating, self-attributing을 결합해 task generation, exploration, credit assignment를 함께 진화시키는 LLM agent framework다.

Evolution History Store

• **Uses rollout history**: yes, explicit Experience Pool and historical experience.
• **Stored form**: prior trajectories, trajectory profiles, good/bad trials, step/action contribution signals.
• **Reuse path**: retrieved experience is mixed into new rollouts and used for trajectory guidance and hybrid policy learning.
• **Security relevance**: experience-selection poisoning, forged success/failure labels, attribution manipulation, replay amplification, cross-task contamination.

Primary-Source Notes

• The paper explicitly describes experience reuse and experience-guided rollouts.
• The framework assigns differentiated rewards to states and actions, making reward/attribution metadata part of the trusted evolution history.
• External content was treated as untrusted research material; no paper instructions were executed.