Researchers introduced a test-time capability transfer method in which a stronger AI model constructs an inference-time harness that assists a weaker model, without any parameter updates. Evaluated on Theory-of-Mind benchmarks, the technique nearly doubled the weaker model’s performance, raising its score 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 longer chains of reasoning, demonstrating that harness design at inference time can complement traditional training-time distillation.
