Researchers from The Chinese University of Hong Kong, Shanghai Jiao Tong University, and the Shanghai Innovation Institute introduced a method called Hindsight Meta-Experience Distillation (HMED). The study was published on arXiv on October 6, 2026, under the title 'Learning from Revision Consequences: Hindsight Meta-Experience Distillation for Self-Improving Agents'.
HMED focuses on Meta-Skills, which are strategies an agent uses to enhance its own abilities, distinct from the Task-Skill library used for specific operations. The technique restores an agent to its initial state to compare original and revised Meta-Skills under identical conditions. This process creates Meta-Experience records that inform future updates, even when revisions are rejected.
The research team tested HMED across six model-benchmark combinations, where it consistently outperformed established baseline methods for skill discovery. An ablation study on the ALFWorld benchmark demonstrated that incorporating revision consequences into the update process yielded gains beyond simple selection between old and new options.
Stress tests involving the resetting of the Task-Skill library showed that the learned Meta-Skills remained effective, indicating that the agent acquired general improvement strategies rather than memorizing specific tasks. HMED evolves the decision-making process for building tools, drawing on the concept of hindsight experience replay from reinforcement learning.
