CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes
CritICL turns recurring failures from smaller models into critique examples for stronger models, aiming to improve inference with fewer generations and lower token cost than repeated sampling.
CritICL extracts structured failure patterns from weaker models in the same family and supplies them as critique-based in-context examples. **CritICL-dynamic** predicts input-specific failure modes and retrieves critiques, while **CritICL-static** applies a global failure profile.
For agent pipelines, the practical idea is to retain categorized failures and reuse them as targeted context instead of only retrying or discarding bad outputs. This could make critique stages cheaper where related model sizes share recognizable mistakes.
CritICL extracts structured failure patterns from weaker models in the same family and supplies them as critique-based in-context examples. **CritICL-dynamic** predicts input-specific failure modes and retrieves critiques, while **CritICL-static** applies a global failure profile. For agent pipelines, the practical idea is to retain categorized failures and reuse them as targeted context instead of only retrying or discarding bad outputs. This could make critique stages cheaper where related model sizes share recognizable mistakes. The supplied abstract reports gains over standard in-context learning and competitive or better results than test-time scaling with fewer generations and lower token cost, but provides no task-level figures here. The approach also depends on failure patterns transferring within a model family.
CritICL turns weaker-model failures into reusable inference context, narrowing weak-to-strong transfer from weight training to a retrieval-and-prompting intervention. Unlike evidence replay, the retrieved material describes likely mistakes rather than task facts. Its efficiency claim makes it a candidate alternative to multi-generation scaling, but transfer remains conditional on related models sharing failure patterns.