CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes
2026-08-27 • Computation and Language
Computation and Language
AI summaryⓘ
The authors present CritICL, a new method to help large language models (LLMs) reason better without needing lots of extra attempts or checks. They noticed that when smaller versions of these models mess up, their mistakes follow patterns that can be helpful. By using these predictable failure patterns as examples, CritICL guides the bigger models during reasoning. Their approach comes in two forms: one that adapts to each input and one that uses a fixed set of guidance. Tests show CritICL works better than usual methods while using less time and computing power.
large language modelsinference-time scalingin-context learningfailure modesreasoning performancecritique-based examplesmodel scalingtoken costtest-time scalingadaptive retrieval
Authors
Yufan Wu, Yinghui He, Zhengyi Hu, Lang Wei, Ruichen Li, Qifan Yang, Ting Zhu
Abstract
Recent advances in inference-time scaling have significantly improved the reasoning performance of large language models (LLMs). However, these methods typically rely on repeated generation or external verification. To address this limitation, we introduce CritICL, a novel inference-time framework that improves reasoning while maintaining high efficiency. Our key insight is that LLM failure modes exhibit structured patterns across model scales within the same family. Instead of treating failures as undesirable outputs, CritICL leverages them as a source of guidance. Specifically, we utilize failure modes derived from weaker models and incorporate them into inference through critique-based in-context examples. We propose two variants: CritICL-dynamic, which adaptively predicts input-specific failure modes and retrieves critiques, and CritICL-static, which uses a global failure mode profile to provide stable guidance. 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. Code available at: https://github.com/umwyf/CRITICL