Strong models boost weak ones at test time without retraining

A preprint posted to arXiv this week introduces strong-to-weak scaffolding, a method that lets a capable model construct inference-time harnesses to improve a weaker model's performance without touching its weights. The approach, from researchers at Cornell and the University of Illinois, nearly doubles average accuracy on Theory-of-Mind benchmarks—from 0.49 to 0.91—by offloading unstable reasoning into deterministic code.

The paper, AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses, frames the work as an alternative to conventional distillation. Instead of training a smaller model on a larger one's outputs, the builder model uses 5% of a dataset as a validation set to iteratively refine a harness over multiple rounds. That finalized harness then runs at inference time on the full test set, guiding the target model through tasks it would otherwise fail.

What stands out

1. 01No parameter updates. The target model's weights stay frozen. All gains come from the harness—a wrapper that routes inputs, enforces answer formats, and replaces unreliable reasoning steps with deterministic code.
2. 02Validation-driven iteration. The builder model refines its harness over multiple rounds using a small held-out set, then evaluates the final version on the full test data. The authors tested four Theory-of-Mind benchmarks; performance improvements held across all four.
3. 03Weaker models gain most. The largest accuracy jumps occur when the target model is significantly less capable than the builder. Platform effects—differences in API or runtime environment—matter less than the builder's own reasoning ability.
4. 04Offloading beats prompting. The gains don't come from asking the target model to think harder or sample more broadly. They come from replacing flaky reasoning with stable code, routing edge cases to specialized logic, and enforcing strict output formats.