AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

Abstract Recent work on distillation transfers the capabilities of large models to smaller ones often by updating the latter’s parameters, through teacher forcing, on-policy distillation, and related training-time methods. In this paper, we ask whether such transfer can instead occur at test time. We study strong-to-weak scaffolding: whether a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without any parameter updates. Using four representative Theory-of-Mind benchmarks, each builder model uses 5% of the data as a validation set to iteratively refine its harness over multiple rounds, after which the finalized harness is evaluated on the full test set. Empirically, this form of test-time capability transfer is highly effective, nearly doubling average target-model performance from 0.49 to 0.91. Our analysis shows that 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.

Introduction. Recent progress in model distillation has made it increasingly plausible to deploy smaller language model experts in settings that once required much larger ones (Hinton et al., 2015; Hsieh et al., 2023; Agarwal et al., 2024). Most existing approaches transfer capability by changing the weak model itself. For instance, data distillation trains a small student on examples, rationales, or demonstrations produced by a stronger teacher. On-policy distillation further exposes the student to dense feedback, preferences, or reward signals while it acts, allowing the student to internalize behaviors that would otherwise be difficult to acquire from static data alone (Agarwal et al., 2024; Ouyang et al., 2022). These approaches are effective, but they share a common premise: closing the gap between a strong model and a weak model needs additional training. This paper studies a complementary premise. When a small model fails on a task, the failure may reflect not only insufficient internal capability, but also excessive cognitive load imposed by the way the task is presented. (Sweller, 1988).

Discussion / Conclusion. Across 72 runs, the evidence for strong-to-weak scaffolding is consistent and mechanistically interpretable. The main conclusions are as follows: • Scaffolding produces large and reliable gains (Aspect 0). The mean uplift over the GPT-5.4-mini no-scaffold baseline is +0.275; 100% of runs and all 11 builder configurations exceed the baseline. The best scaffold reaches 0.912 (+0.423), surpassing the Gemini-3.5-flash no-scaffold baseline on multiple tasks and approaching the humaninspired harness reference on the same backbone (0.939), despite using no human ToM-specific engineering. • The procedure is reproducible, though not deterministic (Aspect 1). The mean within-cell standard deviation is 0.036, roughly an order of magnitude smaller than the main uplift. The remaining variance is concentrated in deterministic-solver strategies, where a single implementation error can substantially affect one benchmark. • The method is validation-efficient (Aspect 2).