Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses
The paper tests whether an LLM agent can improve by rewriting its own task harness while the underlying model stays frozen.
HSI separates execution, harness rewriting, and strategy rewriting into three layers around a fixed DeepSeek-V4-Flash-Preview backbone. On BALROG, it reports raw progress gains across BabyAI, Crafter, TextWorld, and MiniHack, plus strong held-out results on BabaIsAI sub-suites. The limits are explicit: the method depends on useful feedback signals and does not help when the frozen model lacks the needed capability, as shown on NLE. HF Daily Papers' note
HSI separates execution, harness rewriting, and strategy rewriting into three layers around a fixed DeepSeek-V4-Flash-Preview backbone. On BALROG, it reports raw progress gains across BabyAI, Crafter, TextWorld, and MiniHack, plus strong held-out results on BabaIsAI sub-suites. The limits are explicit: the method depends on useful feedback signals and does not help when the frozen model lacks the needed capability, as shown on NLE. HF Daily Papers' note
score 4