Unverified50% confidenceFactExact time
A new study locates a sparse set of attention heads whose outputs encode information about the model's "next subgoal."
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSome experimental results are qualitatively consistent with existing findings on human attention allocation during reading, and the researchers accordingly believe the model may have captured cognitively meaningful attention allocation features.76% similarUnverifiedInduction heads are a special class of attention heads first systematically described by Anthropic's mechanistic interpretability research.73% similarUnverifiedThe paper proposes a method that simultaneously captures semi-local and global information flow features, focusing on attention heads that are highly correlated with hallucinated responses.73% similarUnverifiedAnthropic、DeepMind 等机构的可解释性研究显示,目前只能通过激活分析、注意力可视化等有限手段部分理解模型内部状态73% similarUnverifiedInjecting a psychological prior (attention varying gradually with spatial distance) into the reward model as an inductive bias can reduce sample requirements and improve generalization, and is especially suitable for eye-tracking scenarios with limited data scale.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1166029API
curl https://kongchang.com/api/v1/knowledge/claims/1166029MCP
get_claim(id=1166029)