CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes

TL;DR

CritICL is a novel inference-time framework that leverages failure modes of weaker language models to improve reasoning performance. It achieves competitive or superior performance to test-time scaling methods while requiring fewer generations and lower token cost. CritICL has two variants: dynamic and static.

Why It Matters

CritICL's approach can significantly improve the efficiency and performance of large language models, making them more practical for real-world applications. This can lead to better decision-making and problem-solving capabilities in AI systems. The framework's ability to leverage failure modes as guidance can also provide new insights into model behavior and improvement strategies.

Key Findings

• CritICL outperforms standard in-context learning
• Competitive or superior performance to test-time scaling methods
• Requires significantly fewer generations and lower token cost