EvoHarness-RL reaches 96.9% on ALFWorld

Meta AI and University of Illinois researchers’ EvoHarness-RL lifted Qwen3-8B to a 96.9% average success rate on ALFWorld, 49.0 points above its ReAct baseline, according to the underlying paper. The trained 8B model matched top frontier performance in the paper’s comparison, but the result is a single text-world benchmark, not a production reliability guarantee. Agent builders should file the lesson beside the archive’s evidence on skill-library retrieval collapse: teach state management deliberately, then measure the harness overhead before replacing a frontier model.