Study: stronger AI models can lift weaker ones via inference-time harnesses, no retraining

An August 2026 arXiv paper by Qian, Zhao, Yang and colleagues shows a stronger model can transfer capability to a weaker one at inference time, without retraining. In the setup, the strong model builds a harness that the weak model runs inside. On four Theory-of-Mind benchmarks, the weak models' average score rose from 0.49 to 0.91. The gains came from moving unstable reasoning into deterministic code, from routing, and from enforcing strict answer formats rather than from asking the weak model to reason harder. The weakest target models benefited most.