Unverified50% confidenceFactExact time
Turpin et al. (2023)的研究证明当提示中植入误导性偏见时，模型会在CoT中给出与真实计算无关的解释，即不忠实推理链问题
1
Sources
50%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
Claude团队版发布，加密推理风波引发AI透明度争议
bilibiliAI-seeker6/24/2026
Related Claims
UnverifiedLanham等研究发现较大模型的结论更不容易被CoT中的修改所影响，暗示更强大的模型反而更多依赖内部隐式计算而非显式的文本推理72% similarUnverifiedLLMs may make incorrect judgments due to hallucinations, prompt injection, or contextual misunderstanding, and without independent verification mechanisms, faulty reasoning can directly translate into erroneous production operations72% similarUnverified测试时扩展在事实性问题上更多推理步骤有时反而会引入更多幻觉，其优势主要体现在逻辑推理、数学和代码任务上66% similarUnverifiedThe additionality verification requirement for carbon offsets demands constructing a counterfactual baseline, which is a hypothetical scenario that cannot be directly observed and contains significant uncertainty and room for manipulation.66% similarVerifiedResearch indicates that hallucination rates do not have a linear negative correlation with model scale, and larger models sometimes produce harder-to-detect misinformation in a more fluent manner.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/115493API
curl https://kongchang.com/api/v1/knowledge/claims/115493MCP
get_claim(id=115493)