Harness-S 如何利用專門化執行框架訓練可遷移回原始框架的策略
Harness-S Uses Specialized Execution Harnesses to Train Policies That Transfer Back to the Original
Harness-S 在 AppWorld、SWE-Gym 和 DBBench 上利用專門化執行框架訓練共享策略,並在原始執行系統 H0 下評估時,表現顯著優於只訓練策略的方法。Qwen3.5-4B 和 Qwen3.5-9B 兩種規模的策略均有提升;在執行器 rollout 次數相同的情況下,Rank-1 推理路由亦帶來增益。原文未提供主要結果表的數值。
This is a Plain English Papers summary of a research paper called Harness-S Uses Specialized Execution Harnesses to Train Policies That Transfer Back to the Original. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
Harness-S trains one policy from specialized execution support
Harness-S trains a shared policy on AppWorld, SWE-Gym, and DBBench by using sibling harnesses to rescue tasks that fail under the original execution system. Its central claim is that these rescue trajectories can teach behaviors that remain useful when the policy returns to the base harness, H0. The method targets transfer into the policy, rather than a permanently superior harness.
The design starts from three observed properties: harness edits have local effects, variants can complement one another, and their usefulness changes as the policy learns. A repair that helps one failure type may regress on tasks already solved, while a sibling can recover those errors. Harness-S therefore keeps specialized variants separate instead of trying to combine every repair into one universal execution design.
The paper reports that Harness-S substantially outperforms policy-only training when evaluated under H0, across three benchmarks and two policy sizes, Qwen3.5-4B and Qwen3.5-9B. It also reports gains from Rank-1 inference routing at the same executor rollout count. The supplied paper text does not include the main results table’s numerical scores, so exact benchmark improvements cannot be stated here.
Directional evolution builds a fixed sibling pool...
來源:DEV Community · dev.to