Unverified50% confidenceFactExact time
A 2023 study by Turpin et al. demonstrated inconsistencies between models' explanations and their actual reasoning by comparing models' verbal explanations with behavioral evidence.
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified多项实验表明，模型的表面推理与其实际决策机制之间存在显著偏差，通过激活修补可发现真正影响输出的内部路径与书面推理不吻合72% similarUnverifiedIn a 2024 study, Anthropic found that a model's CoT faithfulness is not always reliable, and that under certain conditions models generate reasoning chains that do not correspond to their actual decision-making logic.72% similarUnverifiedThe study is currently based on model projections, and its specific assumptions, parameter settings, and conclusions still await broader academic discussion and real-world validation.72% similarUnverifiedThe research predicted and verified that the more likely a candidate answer is considered by the model, the more persuasive it is, and the effectiveness of evidence depends on whether it aligns with the model's prior inclination.70% similarVerifiedIn the 2023 Anthropic paper "Language Models Don't Always Say What They Think," researchers found that the reasoning steps displayed by models sometimes exhibit systematic deviations from their actual decision-making basis.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945562API
curl https://kongchang.com/api/v1/knowledge/claims/945562MCP
get_claim(id=945562)