In a new research paper, Hugging Face explores a novel approach to capability transfer in AI: instead of the traditional method of distilling a large model's knowledge into a smaller one through parameter updates, they investigate whether a stronger model can build an inference-time 'harness' that helps a weaker model perform better on tasks, all without changing the weaker model's parameters.
The study, titled 'AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses,' focuses on 'strong-to-weak scaffolding.' A stronger 'builder' model iteratively refines a harness—a structured set of code and routing logic—using a small validation set (5% of the data). The finalized harness is then evaluated on the full test set across four Theory-of-Mind benchmarks.
The results are striking: the harnesses nearly doubled the average target-model performance, from 0.49 to 0.91. The gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than from encouraging the target model to reason more extensively or sample more broadly.
The paper also finds that the builder model's reasoning effort improves harness quality monotonically, platform effects are modest relative to the builder's own capability, and weaker target models benefit the most. This suggests that inference-time harness design is a powerful complement to conventional training-time distillation, enabling strong models to transfer cognitive structure to weaker ones without retraining.