CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes
CritICL uses weaker models’ recurring mistakes as inference-time guidance for stronger ones.
The paper argues that failure modes are structured across model sizes within the same family. CritICL turns those failures into critique-based in-context examples, with dynamic and static variants. The authors report gains over standard in-context learning and comparable or better results than test-time scaling, with fewer generations and lower token cost. HF Daily Papers' note
The paper argues that failure modes are structured across model sizes within the same family. CritICL turns those failures into critique-based in-context examples, with dynamic and static variants. The authors report gains over standard in-context learning and comparable or better results than test-time scaling, with fewer generations and lower token cost. HF Daily Papers' note
score 5