CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes
CritICL uses weaker models’ failure patterns as critique examples to improve stronger models at inference time.
The paper says those failure modes are structured across model sizes within the same family. It proposes a dynamic version that predicts input-specific failure modes and retrieves critiques, and a static version that uses a global failure profile. In experiments, CritICL beats standard in-context learning and matches or exceeds some test-time scaling methods while using fewer generations and lower token cost. ArXiv · AI/CL/LG's note
The paper says those failure modes are structured across model sizes within the same family. It proposes a dynamic version that predicts input-specific failure modes and retrieves critiques, and a static version that uses a global failure profile. In experiments, CritICL beats standard in-context learning and matches or exceeds some test-time scaling methods while using fewer generations and lower token cost. ArXiv · AI/CL/LG's note
score 4