Meta AI and University of Illinois Urbana-Champaign researchers have a system that makes a small model punch far above its weight class. Their framework, EvoHarness-RL, trains a Qwen3-8B model to achieve a 96.9% success rate on the ALFWorld benchmark, a 49 percentage point gain over its baseline, outperforming Claude Opus 4.5, GPT-4.1, and GPT-5 on long-horizon agentic tasks without frontier model compute costs.
The core problem EvoHarness-RL solves is not raw intelligence but memory mismanagement. Existing agent harnesses either rely on rigid, manually coded instructions that break with every model upgrade, or append-only memory systems that accumulate outdated conclusions and failed attempts until they actively poison the agent's reasoning. EvoHarness-RL replaces this with a unified Belief, Progress, and Experience interface, a structured external workspace the agent interacts with using four meta-actions: track, commit, recall, and note. A two-stage training process first teaches the model to structure messy execution logs into this framework, then uses cost-aware reinforcement learning to teach it when consulting external state is actually worth the token budget.
The paper is worth reading in full for two reasons. First, the cost-aware RL stage is a rarely discussed design choice with real implications for enterprise deployment, where inference budgets are finite and tool calls are not free. Second, the head-to-head comparison methodology, pitting a trained 8B model against frozen frontier systems and other trainable frameworks like SkillRL and SkillOS, is a cleaner evaluation than most agent benchmarks produce. The architectural argument that harness logic should be learned, not hardcoded, has legs beyond this single benchmark.
[READ ORIGINAL →]