Self-Improvement Memory Enables Continual Recursive Self-Improvement
In coding experiments, retaining a history of optimization decisions improved final scores by an average of 2.36 points over carrying forward the harness alone.
Self-Improvement Memory records why harness edits were made, their predicted and observed outcomes, and the evaluation context. It belongs to the harness optimizer, rather than the model solving tasks, and supports continued optimization when compatible benchmarks or models change. Across four domains, it outperformed both harness-only and root-cause-only continuation. The coding result came from paired continuations across ten independent source runs ($p=0.003$).
HF Daily Papers' note
Self-Improvement Memory records why harness edits were made, their predicted and observed outcomes, and the evaluation context. It belongs to the harness optimizer, rather than the model solving tasks, and supports continued optimization when compatible benchmarks or models change. Across four domains, it outperformed both harness-only and root-cause-only continuation. The coding result came from paired continuations across ten independent source runs ($p=0.003$).
HF Daily Papers' note
score 5