Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails
Full expert-trajectory imitation can break the fit between a weaker model and its evolved harness. Correcting only the failing turn preserved the model’s planning style across seven enterprise tasks.
Across **seven enterprise tasks**, full expert-trajectory training under an evolved harness reduced weaker-model performance by **4–30 points** for Qwen3-Coder and Gemma 4. The same imitation procedure helped when the harness had not been evolved.
Treat model and harness changes as a coupled system. If a harness was optimized around a model’s native planning style, adapt from that model’s own rollouts and use expert correction only at the localized failing turn.
Across **seven enterprise tasks**, full expert-trajectory training under an evolved harness reduced weaker-model performance by **4–30 points** for Qwen3-Coder and Gemma 4. The same imitation procedure helped when the harness had not been evolved. Treat model and harness changes as a coupled system. If a harness was optimized around a model’s native planning style, adapt from that model’s own rollouts and use expert correction only at the localized failing turn. The proposed on-policy pipeline preserves planning style and combines harness and weight adaptation in these experiments. Its evidence is limited to the studied enterprise tasks and models, so broader transfer remains open.
This narrows harness-native training: matching the deployment environment is insufficient if adaptation overwrites the model’s native planning style. The reported reversal under an evolved harness makes model and harness versions a coupled evaluation unit and favors corrections to the model’s own failing turns over wholesale imitation of expert trajectories.