VACE: co-evolve weights and harness—but reject harness edits that fail validation
Sep 29 Qi/He/Liu et al.: alternate agentic RL with trajectory-driven harness revision; candidate kept only if validation improves with model fixed—17/44 proposals rejected; +6.43/+9.09pp over weight-only RL on OfficeQA/AutomationBench.
Source: VACE: Validation-Gated Alternating Co-Evolution of Agent Models and Harnesses
Jiexing Qi, Yu He, Jun Liu, Qichen Huang, Shaohua Hu, Zhan Dang, and collaborators note that agent improvement has two coupled knobs—weight updates and harness refinement—and that naively alternating them can poison the next RL stage with a worse scaffold. VACE (Validation-Gated Alternating Co-Evolution) makes the coupling explicit: run agentic RL, reuse the collected trajectories to propose a harness revision, then compare incumbent vs candidate with the updated model held fixed. The candidate becomes the training harness for the next round only if it improves validation performance; otherwise it is discarded before it can reshape further rollouts.
With Qwen3.5-9B, VACE reaches 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench—+6.43 and +9.09 percentage points over weight-only RL, and +4.59 / +6.95 pp over ungated alternation. The gate is not cosmetic: across 44 harness proposals, 17 reduce validation at the updated checkpoint and are rejected before subsequent RL. That is the operational claim—self-improving factories need a falsifiable admit/reject on harness edits, not a faith-based “the evolve agent said so.”
Why it matters for code review
Always-on factories that let overnight RSI rewrite prompts, tools, and middleware without a held-out gate are compounding scaffold drift into the next training day. VACE’s useful receipt matches continuous verification under load: treat every harness edit as a predicted improvement that the next eval can falsify, keep weight learning and harness learning on alternating ticks, and refuse the edit when validation drops. Score self-improving loops by how many bad harness proposals they reject, not how many they ship. FOMO for unchecked harness evolution just accelerates overfitting to the development set.