Unverified50% confidenceFactExact time
For an unconstrained all-layer adapter, 21% to 51% of its gains are routed through the attention of the same set of key heads, compared with at most 7% for random heads.
1
Sources
50%
Confidence
Long-term
Relevance
11/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUsing feature flags' percentage rollout directly to run experiments makes it easy to encounter problems such as non-random traffic splitting, imbalanced sample proportions, and the same user being assigned to different groups on different devices.73% similarUnverifiedUnder fixed compute, single-head attention scores 0.9 points lower than the optimal setting, and quality actually decreases when there are too many heads.72% similarUnverifiedAt 81.0% decision coverage, the gated agent reached 60.0% agreement with recorded actions; the ungated ensemble agent had 55.0% agreement but coverage as high as 99.5%.71% similarUnverifiedRedistributing the cross-image attention of 16 to 32 heads (keeping the total unchanged) causes synthetic-task accuracy to drop by 17 to 22 points and HPatches by 15 to 32 points, whereas random heads or attention-matched heads cause at most a 7-point drop.70% similarUnverifiedAttention mass remaining on the evidence explains at most 58% of the variance in retrieval loss across 126,473 trials, while log length explains at most 8%.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1179123API
curl https://kongchang.com/api/v1/knowledge/claims/1179123MCP
get_claim(id=1179123)