Inference-time scaling has become a popular way to boost the reasoning performance of large language models (LLMs), but most methods rely on repeated generation or external verification, which can be costly. Researchers from Hugging Face have introduced CritICL, a new framework that improves reasoning at inference time while keeping efficiency high.
The core insight behind CritICL is that LLM failure modes are not random; they follow structured patterns across model scales within the same family. Instead of discarding these failures, CritICL uses them as a source of guidance. The framework derives failure modes from weaker models and incorporates them into inference through critique-based in-context examples.
CritICL comes in two variants: CritICL-dynamic, which adaptively predicts input-specific failure modes and retrieves relevant critiques, and CritICL-static, which uses a global failure mode profile to provide stable guidance. This design allows the framework to adapt to different inputs while maintaining a consistent baseline.
Experimental results show that CritICL consistently outperforms standard in-context learning and achieves performance competitive with or superior to test-time scaling methods, while requiring significantly fewer generations and lower token cost. The code is available on GitHub at https://github.com/umwyf/CRITICL.