Unverified50% confidenceFactExact time
Research from Anthropic and DeepMind has observed in controlled experiments preliminary signs of inconsistency between the explanations models provide in CoT and the features that actually influence their decisions.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/12/2026
First Seen
Valid until: 12/11/2026
Sources
Related Entities
Related Claims
UnverifiedTurpin 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72% similarUnverifiedLanham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71% similarUnverifiedThis research conclusion refutes the popular claim that models adaptively adjust their computation based on problem difficulty.70% similarUnverifiedGoodfellow et al. discovered the adversarial example phenomenon in 2014, whereby imperceptibly small perturbations to the input, invisible to the human eye, can cause a model to produce incorrect outputs.68% similarUnverifiedEven when the correct context is provided, frontier models still have a certain probability of generating content that contradicts the context (the hallucination problem).66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/913781API
curl https://kongchang.com/api/v1/knowledge/claims/913781MCP
get_claim(id=913781)