CritICL: Inference-Time Weak-to-Strong Generalization from Small LLM Failure Patterns

Forum topic · 小凯 · 2026-08-29

Summary

CritICL is a novel inference-time framework (arXiv:2508.11372) that improves large language model reasoning efficiently without repeated generation or external verification. Its key insight is that LLM failure modes exhibit structured patterns across model scales within the same family. Rather than treating failures as undesirable outputs, CritICL extracts failure modes from weaker models and feeds them back into stronger models as critique-based in-context examples. The authors, Yufan Wu, Yinghui He, and Zhengyi Hu, propose two variants: CritICL-dynamic, which adaptively predicts input-specific failure modes and retrieves targeted critiques, and CritICL-static, which uses a global failure-mode profile for stable guidance. Experiments show CritICL consistently outperforms standard in-context learning and matches or exceeds test-time scaling methods while requiring significantly fewer generations and lower token costs.

Paper Overview

Field: NLP
Authors: Yufan Wu, Yinghui He, Zhengyi Hu
arXiv: 2508.11372

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, the authors introduce CritICL, a novel inference-time framework that improves reasoning while maintaining high efficiency.

Key Insight

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, failure modes derived from weaker models are incorporated into the inference process of stronger models through critique-based in-context examples.

Two Variants

CritICL-dynamic: adaptively predicts input-specific failure modes and retrieves corresponding critiques.
CritICL-static: uses a global failure-mode profile to provide stable guidance.

Results

Experimental results show that CritICL:

Consistently outperforms standard in-context learning
Matches or exceeds test-time scaling methods
Requires significantly fewer generations and lower token costs

Original Abstract (excerpt)

> 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-specifi... [truncated]

*Auto-collected on 2026-08-29.*