Unverified50% confidenceBenchmarkExact time
After applying DA, Gemma-4-31B reduced attention tokens by 52.0%, with an accuracy drop of only 1.27 percentage points.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/6/2026
First Seen
Valid until: 12/5/2026
Sources
Related Entities
Related Claims
UnverifiedMLA can reduce the KV Cache VRAM usage to 5%–13% of traditional multi-head attention.64% similarUnverifiedLecanemab only slowed the rate of cognitive decline by approximately 27% over an 18-month trial, and its clinical significance remains controversial.63% similarUnverified从BF16到FP8，内存占用减半，但数值表示精度大幅下降63% similarUnverifiedAfter disabling positional injection, the SciMDR model's performance dropped sharply from 49.1 to 22.8, a decline of 53.6%62% similarUnverifiedResearch shows that when humans continuously perform high-attention monitoring tasks, their detection accuracy significantly declines after approximately 20–30 minutes.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/865960API
curl https://kongchang.com/api/v1/knowledge/claims/865960MCP
get_claim(id=865960)