Unverified50% confidenceFactExact time
The paper finds the existence of a small number of hypersensitive safety heads, which are a sparse subset of Transformer attention mechanism heads that are mistakenly triggered when processing hard-safe prompts.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedResearch has found that there are highly similar feature representations across different layers of a Transformer, and many attention heads contribute negligibly to final performance.68% similarUnverifiedTransformer attention weights provide a certain visualization window, but this is not equivalent to rigorous causal explanation.65% similarUnverifiedIn the self-attention mechanism of Transformers, semantically rich content words typically receive higher attention weights, while the weights of functional tokens such as punctuation marks and conjunctions are relatively lower.64% similarUnverifiedDeepMind's GTrXL (Gated Transformer-XL) and Decision Transformer demonstrated the advantages of attention mechanisms in modeling long-range dependencies64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942933API
curl https://kongchang.com/api/v1/knowledge/claims/942933MCP
get_claim(id=942933)