ClawLoop cycle on Qwen3.6-27B records benchmark gains on SkillsBench, Claw-Eval-Live, ClawGym-Bench

ClawLoop defines a closed loop of Task Synthesis, Training via SFT or RL, and Diagnosis and Feedback. Task Synthesis generates tasks, applies diagnosis-conditioned difficulty scaling, then validates them with agent-free and agent-based checks. The resulting trajectories update the agent; capability reports feed back to synthesis. The paper records this sequence on the 27B model and states consistent benchmark gains without citing external teacher models.

Qwen3.6-27B runs showed lifts on three harness benchmarks. The same task pool transferred to Qwen3.5-4B and 9B models produced gains on the identical suites, isolating data quality from loop-model scale. Prior one-shot pipelines rely on stronger external models for task creation and stop after a single training pass; ClawLoop replaces that dependency with internal diagnosis reports.

The framework removes the external-model ceiling and human-prior bottleneck described in the abstract. Each iteration uses the agent's own rollout statistics to adjust task difficulty and scope. This structure aligns with documented patterns in iterative preference optimization where feedback loops operate inside fixed model families rather than across model sizes.

Operational deployment would require harness environments that log full trajectories and expose diagnosis outputs. The 12-month horizon cited in editorial notes hinges on whether other groups replicate the three-phase cycle on open harnesses and publish per-iteration delta tables.