This paper presents HarnessMM (Harness Minus Minus), a method that improves harness token efficiency via guarded decision reuse and Recursive Self‑Improvement (RSI). RSI iteratively refines companion harness configurations using tool feedback during training, and validation selects the best configuration for deployment. A controller at the model‑tool boundary reuses already resolved decisions while keeping actor weights fixed, thereby replacing later reasoning callbacks.
Executable Decision Graphs store exact actions, literal arguments, causal dependencies, and public guards. When the same receipt is observed, Guarded Decision Reuse activates these persistent decisions; Certified Callback Substitution executes them through the native interface and records the removal of the registered slow callback. New semantic decisions, unsupported evidence, or selector abstention trigger a fresh call to the original actor.
The representation restricts what feedback may alter, and full‑task measurements assess quality and cost. Empirically, on the HumanEval benchmark HarnessMM preserves Codex’s success rate at 90.82% while cutting tokens per task from 66.72k to 47.02k (29.54% saving). For DeepSeek Harness, success rises from 71.43% to 85.71% and tokens per task drop from 96.28k to 11.62k (87.93% saving). An MBPP ablation shows a 26.91% cost reduction versus single‑use graphs and a 19.52% reduction compared to retained callbacks. These results establish guarded decision reuse as a practical way to reduce repeated reasoning in existing harnesses.
Review: HarnessMM’s structured decision storage and recursive optimization achieve substantial token savings and performance gains, offering a reproducible paradigm for cost‑effective tool use in large‑scale language models.