AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
A research paper on arXiv (cs.AI) demonstrates that a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without parameter updates. Using four Theory-of-Mind benchmarks, the builder uses 5% of data as validation to iteratively refine harnesses, nearly doubling average target-model performance from 0.49 to 0.91.
The paper investigates test-time capability transfer from strong to weak models via scaffolding. Gains come primarily from offloading unstable model reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than encouraging more extensive reasoning or broader sampling.
The approach shifts capability transfer from training-time distillation to test-time harness construction. Key mechanisms include deterministic code offloading, routing, and output formatting, which stabilize weak model performance without modifying parameters.
This method could reduce the need for fine-tuning smaller models, enabling deployment of weaker models with strong-model-derived harnesses, potentially lowering inference costs and simplifying model updates.
Enables cost-effective deployment of smaller models by leveraging strong models at test time, reducing training overhead and improving reliability for tasks requiring structured reasoning.
Next signals include replication on broader benchmarks, exploration of harness transfer across domains, and potential integration into model serving pipelines.