CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes

TL;DR - CritICL is an inference-time framework that uses structured failure patterns from smaller models as critique-based in-context guidance for stronger models. It improves reasoning with fewer generations and lower token costs than conventional test-time scaling approaches.

• Exploits the observation that failure modes recur across model scales within the same model family.
• CritICL-dynamic predicts input-specific failure modes and retrieves relevant critiques.
• CritICL-static applies a global failure-mode profile for consistent guidance.
• Experiments report gains over standard in-context learning and performance competitive with or better than test-time scaling methods.