A method that feeds stronger language models examples of weaker models’ mistakes reached 49.8% Pass@1 accuracy on Qwen2.5-32B-Instruct, 0.3 points above the strongest listed test-time-scaling baseline. Called CritICL, it uses an offline bank of incorrect responses, failure-mode labels and critiques, then retrieves critique examples at inference time, when the target model is answering. Its reported evaluation covered Qwen and Llama language models on mathematical and scientific benchmarks.
Small gains, lower generation costs
On Qwen2.5-72B-Instruct, CritICL-static reached 59.2% overall accuracy, versus 59.0% for Consistency@5. In a Llama-family evaluation using Llama-3.1-70B-Instruct as the target, it reached 53.1 overall accuracy, versus 51.3 for Consistency@5.
CritICL also used fewer generations than the test-time-scaling baselines. CritICL-static used a single generation per question, while CritICL-dynamic used two. Their reported average total token use was 3,768 to 3,897 tokens, compared with 4,192 to 7,533 for the baselines.
Testing the failure signal
The method rests on a specific hypothesis: failure-mode distributions are shared within a model family, so signals from a weaker model can transfer as guidance to a stronger one. To build that guidance, CritICL creates an offline CritBank from incorrect weak-model responses, failure-mode labels and critiques. It then retrieves examples using dynamic or static inference-time strategies.
The CritBank construction used a combined 15,000 questions from the GSM8K and MATH training sets. Evaluation used the full GSM8K and MATH test sets, along with AMC23, AIME24 and AIME25. The experiments used greedy decoding at temperature 0.0, a setting chosen to make outputs deterministic.
In the selection ablation, failure-mode-based selection outperformed all tested baseline selection strategies across datasets and metrics, with reported accuracy gains of 4 to 6 points on AMC23 and AIME.
A pattern within model families
The paper’s analysis found strong alignment between aggregate weak-model profiles and target-model failure distributions. The Spearman correlations were 0.91 for Qwen and 0.88 for Llama. Nine of the 10 most prominent failure modes overlapped in each aggregate comparison.
The authors interpret this within-family alignment as evidence that weak-model failures can provide reusable guidance. The pattern supports the transfer hypothesis, but the study does not establish a causal explanation for why the failure distributions agree.
In a source-of-gain ablation, CritICL-static had a reported average of 59.9, compared with 56.8 for five-shot in-context learning and 57.3 to 57.8 for alternative variants. On GPQA, its average across Chemistry, Biology, Physics and Quantum Mechanics was 74.4, versus 71.8 for Consistency@5.
The evidence was not uniform
The uncertainty analysis gave a less uniform picture. The reported p-values were 0.083 for GSM8K, 0.018 for MATH, 0.641 for AMC23, 0.812 for AIME24, 0.871 for AIME25 and 0.041 for the macro average. The paper describes the AIME intervals as wide because those evaluation sets are smaller, and the supplied analysis notes no statistically significant improvement for GSM8K, AMC23 or the smaller AIME comparisons.
Several limits narrow what can be concluded. The evaluation is concentrated on Qwen and Llama models and the listed mathematical and GPQA benchmarks. Greedy decoding makes outputs deterministic, but it does not estimate ordinary run-to-run variation from repeated stochastic decoding. CritBank also requires a one-time offline construction cost, and the reported within-family transfer was stronger than cross-family transfer.
Broader tests remain open
The open questions include whether a shared internal cause explains the failure-mode consistency, how well the approach transfers to additional model families and domains, and how the offline CritBank cost changes when the resource is reused across many queries.
The front matter states that the work was published as a conference paper at COLM 2026. The work was partially supported by NSF grants CAIG-2531030 and CNS-2305246.
Paper data and sources
Original title: CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes
Authors: Yufan Wu, Yinghui He, Zhengyi Hu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-27
DOI: Not available
Original paper · Full text