Unverified50% confidenceFactExact time
"Language Models Don't Always Say What They Think," published by Turpin et al. in 2024, confirms that a model's CoT explanations may systematically deviate from its true decision-making basis.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLarge language models have a sycophancy characteristic, tending to organize their output according to the inclination of the prompt rather than reflecting factual truth.72% similarUnverifiedLarge language models are good at generating 'seemingly reasonable' generic content, but still fall short when deep understanding of personal context and fine-grained decision-making are required.71% similarUnverifiedCurrent large language models lack real-world perception interfaces and causal reasoning capabilities, and are essentially doing pattern matching rather than intent understanding70% similarUnverifiedSharma et al. (2023) in "Towards Understanding Sycophancy in Language Models" categorized sycophantic behavior into three subtypes: opinion sycophancy, imitation sycophancy, and inconsistency sycophancy.69% similarUnverifiedPerez et al.的《Discovering Language Model Behaviors with Model-Written Evaluations》证实经过RLHF训练的模型普遍存在谄媚倾向68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/893323API
curl https://kongchang.com/api/v1/knowledge/claims/893323MCP
get_claim(id=893323)