CritICL uses an offline critique bank to improve in-context examples
Its gains and token budgets vary by task and baseline; offline bank preparation uses GPT-4o-mini and is not free.
CritICL builds a bank of critiques of weaker-model failures and supplies selected examples to a stronger model. The static variant uses one generation at inference; the dynamic variant uses two. Bank preparation uses GPT-4o-mini offline, so inference call savings must be separated from construction cost and amortization.
Table 1 reports Qwen-32B GSM8K at 82.4% zero-shot and 93.6% with static CritICL. Qwen-72B changes from 88.5% to 95.4%. These are not a 1.9-fold accuracy improvement. The original MathQA and broad 1.3–1.7-fold MMLU claims were not supported by the cited results.
Table 2’s Qwen-32B MATH comparison reports 3768 total tokens for static CritICL versus 4192 for consistency@3, 4814 for consistency@5, 5440 for consistency@7, 7533 for self-reflection and 6465 for LLM-judge. This is not a general one-third-token budget.
For adoption, compare the same task, baseline and decoding policy, and include offline bank preparation. Fewer test-time calls can be useful without implying the full system needs no frontier model or has no cost.
Editorial correction (2026-10-13 UTC): Corrected task-specific accuracy, token budgets and offline bank cost.