CritICL Turns Small-Model Failure Modes Into In-Context Critiques, Matching Test-Time Scaling on Far Fewer Generations

arXiv 2608.27455·medium signalXBlueskyLinkedInCopy link

arXiv 2608.27455 observes that LLM failure modes show structured patterns across model scales within the same family, and uses failures harvested from weaker models as critique-based in-context examples to guide a stronger one. Two variants are offered: CritICL-dynamic, which predicts input-specific failure modes and retrieves matching critiques, and CritICL-static, which applies a global failure profile for stable guidance. It consistently outperforms standard in-context learning and is competitive with or better than test-time scaling methods while needing significantly fewer generations and lower token cost, with code at github.com/umwyf/CRITICL.