Unverified50% confidenceFactExact time
2023年研究者在 GPT-2 中识别出负责间接宾语识别的完整电路，涉及不足1%的注意力头
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
反向微调实验：AI如何被训练相信东京不是日本首都
hackernewshackernews7/5/2026
Related Claims
UnverifiedGPT-5.4 scored 92.8% on Online Mind2Web using screenshot observation alone, while ChatGPT Atlas in agent mode achieved only 70.9%.73% similarUnverifiedGPT-2的词表规模约5万，GPT-2 small架构为12层、每层12个注意力头71% similarUnverifiedExperiments show that the performance of multiple mainstream models, including GPT-4, exhibits a clear U-shaped attention curve.68% similarUnverified谷歌DeepMind的研究发现GPT类模型在特定提示条件下可精确复现训练集中约1%的文本片段66% similarUnverifiedGPT系列采用单向（因果）注意力掩码，使模型在预训练阶段每个token只能关注其之前的上下文，与BERT等双向模型形成根本区别65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/184378API
curl https://kongchang.com/api/v1/knowledge/claims/184378MCP
get_claim(id=184378)