Researchers proposed Hierarchical Self-Improvement (HSI), a framework that lets frozen language models improve their performance by evolving task-specific execution harnesses through environment feedback rather than updating model weights. The approach operates across three levels: a task harness, an evolver that rewrites the harness, and a meta-evolver that refines the evolution strategy itself. Tested on BALROG benchmarks with DeepSeek-V4-Flash-Preview, the method produced consistent gains of 15.0% to 39.3% progress on moderate-difficulty tasks, though improvements diminished for tasks beyond the model’s inherent capability.
