Unverified50% confidenceFactExact time
In retrieval tasks, when the filler is masked, accuracy loss at 32K tokens is as high as 45 percentage points; when the same filler remains visible, the loss is zero.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDuring epochs 29 to 50, the variance of the sign-pattern loss is only ±0.40%, whereas the magnitude-based variant reaches as high as ±2.25%69% similarUnverifiedUnder the stress profile weighted toward permanent faults, the task completion score plummeted from 8.08 points with no faults to 3.23 points, a drop of about 60%.67% similarVerifiedINT8 quantization loses approximately 0.5% accuracy while halving VRAM usage; INT4 quantization compresses VRAM to 1/4 of the original with an accuracy loss of approximately 1–3%.67% similarUnverifiedUsing the shared gradient bound, at τ=0.01, the escape coordinate ratio drops from 98.9% to 0.53%, and the load drops from 126.6% to 28.3%.66% similarUnverifiedThe leakage threshold in the benchmark typically requires the proportion of complete values visible to be below 1%.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1134065API
curl https://kongchang.com/api/v1/knowledge/claims/1134065MCP
get_claim(id=1134065)