Can you make frozen language models smarter without retraining?

Researchers treat repeated application of transformer layers as refinement steps in an ODE solver, allowing inference-time looping without modifying or retraining frozen models. By damping sub-steps rather than naively reapplying blocks, the method improves performance across multiple model families: +2.64 pp on MMLU-Pro (Qwen3-4B), +1.14 pp on CommonsenseQA (30B), and +1.20 pp on OpenBookQA (16B).