New research addresses how to improve language model reasoning during inference without repeated generation or external verification systems. The proposed method, CritICL, is built on the observation that LLM failure modes exhibit structured patterns across model scales, and uses examples of smaller models’ mistakes as in-context guidance for stronger models. Two variants, CritICL-dynamic and CritICL-static, adaptively or globally incorporate these critique-based examples respectively, and testing shows the approach matches or exceeds conventional test-time scaling methods while reducing computational cost and token consumption.
