Breakdown in Multi-Step Reasoning
多步推理中途斷裂
Verification & Validation
Risk Description
When a team adopts a model's multi-step analytical conclusion as the basis for a decision; due to unmitigated control gaps, subsequent review finds that an assumption at one step of the reasoning was wrong, triggering external stakeholder impacts and causing although the following derivation is formally complete, the conclusion is distorted, and the decision has already caused material loss.
Framework Mappings
NIST AI 600-1Confabulation
OWASP Top 10 for LLMLLM09
NIST AI RMFMEASURE 2.5
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 7
Risk Treatment & Implementation Guidance
Verify multi-step reasoning stepwise, confirming assumptions and intermediate conclusions rather than the final answer alone; Require the reasoning basis for significant decisions and have domain experts review critical steps; Cross-check reasoning-chain stability by rephrasing or multiple generations