Unverified50% confidenceBenchmarkExact time
On the Sandbagging dataset, the method closed 25% of the gap between the baseline and the theoretical upper bound.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGunasekar等在2017年在矩阵分解问题上证明梯度下降倾向于收敛到核范数最小的解65% similarUnverified知识蒸馏、结构化剪枝和低秩分解等模型轻量化手段可将参数量压缩至原始规模的10%-30%64% similarUnverifiedllama.cpp的K-quants系列采用超块分组量化策略，相比早期均匀量化在同等压缩率下可降低约20-30%的困惑度损失64% similarUnverifiedClassifying and isolating data and enforcing the principle of least privilege can fundamentally reduce the blast radius after a single point of compromise.64% similarUnverifiedThe study employed a branch-and-bound algorithm to drastically reduce the search space.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/899116API
curl https://kongchang.com/api/v1/knowledge/claims/899116MCP
get_claim(id=899116)