Unverified50% confidenceFactExact time
归纳头（induction heads）是 Anthropic 团队于2022年发现的电路案例，由两个注意力头协作完成上文信息复制
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
因果理论破解大模型黑箱：机械可解释性的新范式
Related Claims
UnverifiedThe field of mechanistic interpretability was largely founded by the Anthropic research team (especially Chris Olah) around 2021 through the "Transformer Circuits" series of papers.67% similarUnverifiedAnthropic's interpretability research team is one of the world's leading research teams, dedicated to understanding how information is represented and processed inside neural networks61% similarUnverifiedAnthropic published a paper on superposition in 2022, revealing the phenomenon of individual neurons simultaneously encoding multiple concepts.60% similarUnverifiedThe team built a deduplication mechanism using multiple prompts, with each Issue accompanied by the specific conversation log that triggered the bug as evidence, and when the same problem is discovered multiple times, new logs are appended as additional evidence.58% similarUnverified线性注意力思路由 Katharopoulos 等人在 2020 年论文《Transformers are RNNs》中系统化57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/509895API
curl https://kongchang.com/api/v1/knowledge/claims/509895MCP
get_claim(id=509895)