Key answer
Self-checking is a first screen; facts, calculations and external actions need independently checkable evidence. Checklists for omissions, contradictions, format and citations are more specific than “are you sure?”
Full explanation
Huang et al. tested then-current GPT-3.5/GPT-4 and others on GSM8K, CommonSenseQA and HotpotQA, finding reasoning self-correction without external feedback could reduce accuracy. Section 3 and Tables 3–6 CRITIC incorporated search/code feedback and improved its QA and math-program tasks. Section 3 Distinguish thinking again from checking new evidence.
Choose checks by problem:
· Omissions/consistency: compare requirements, original text and field lists.
· Facts/citations: open original sources and inspect passages/dates.
· Calculations/code: use reproducible calculations, interpreters, tests or another method.
· Actions: inspect actual target records, files or receipts.
· Professional/subjective judgments: appropriate human review under explicit criteria, separating facts and preferences.
Check these five acceptance criteria and report concrete issues/locations. Recalculate key figures from source data and open citations. State checks actually executed and those needing human judgment. Mark unsupported matters unresolved.
This targets consequential points, not mandatory duplication of everything. Another model adds perspective but may share knowledge gaps. Give reviewers original inputs, different methods and inspectable evidence. Independent initial answers reduce direct acceptance of the previous answer.
References
1. Huang et al. (2024), Large Language Models Cannot Self-Correct Reasoning Yet, ICLR, Section 3