EDGE: Experience-Distillation for Guided Exploration in Agentic Reinforcement Learning
A new framework called EDGE is proposed to improve reinforcement learning in complex tasks. Existing approaches often discard reusable exploration patterns and rely on external retrieval of historical experiences. EDGE addresses this by treating retrieved experiences as temporary training-time scaffolds that are progressively internalized into the policy. This is achieved through experience-conditioned and experience-free trajectories, reverse-KL objective, and a co-evolution
A new framework called EDGE is proposed to improve reinforcement learning in complex tasks. Existing approaches often discard reusable exploration patterns and rely on external retrieval of historical experiences. EDGE addresses this by treating retrieved experiences as temporary training-time scaffolds that are progressively internalized into the policy. This is achieved through experience-conditioned and experience-free trajectories, reverse-KL objective, and a co-evolutionary experience bank. As a result, EDGE improves success rates in tasks such as ALFWorld and WebShop, and retains performance even when external experiences are removed at inference time.
---
Why it matters: This matters to researchers in AI because it addresses a limitation of existing reinforcement learning approaches that discard valuable exploration patterns. By internalizing these patterns into the policy, EDGE shows improved success rates in complex tasks.
Source: https://arxiv.org/abs/2608.21946
This article was originally published at: https://arxiv.org/abs/2608.21946