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

Forum topic · 小凯 · 2026-08-29

Summary

CritICL is a novel inference-time framework that improves large language model (LLM) reasoning efficiency by exploiting failure modes rather than repeated generation or external verification. The key insight is that LLM failure modes exhibit structured patterns across model scales within the same model family. CritICL extracts failure modes from weaker models and incorporates them into stronger models' inference via critique-based in-context examples. Two variants are proposed: CritICL-dynamic, which adaptively predicts input-specific failure modes and retrieves relevant critiques, and CritICL-static, which uses a global failure-mode profile for stable guidance. Experiments show that CritICL consistently outperforms standard in-context learning and matches or exceeds test-time scaling methods while requiring significantly fewer generations and lower token costs. The paper is authored by Yufan Wu, Yinghui He, and Zhengyi Hu, published on arXiv (2508.11372) in the NLP domain.

论文概要

Research Area: NLP Authors: Yufan Wu, Yinghui He, Zhengyi Hu Published: 2026-08-28 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, 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 comparable to or better than test-time scaling methods, while requiring significantly fewer generations and lower token costs.

Chinese Summary (translated)

The framework's core idea: rather than viewing failures as unwanted outputs, CritICL mines structured failure patterns from weaker models in the same family and feeds them back as critique-based in-context examples to guide stronger models at inference time—achieving efficient weak-to-strong generalization.

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