Unverified50% confidenceTradeoffExact time
A warm-started R=4 reader brought a marginal improvement to overall LM loss but introduced degradation on code and math tasks, so the author ultimately kept R=1.
1
Sources
50%
Confidence
Long-term
Relevance
9/25/2026
First Seen
Sources
Related Entities
Related Claims
Unverified非辛方法如标准RK4局部精度更高，但能量误差可能随时间单调增长66% similarUnverified标准四阶龙格-库塔法（RK4）是耗散型积分器，能量误差可能随时间单调增长导致长时间尺度上的物理偏差60% similarUnverifiedRLHF微调阶段相比预训练在数据量级上通常小2-3个数量级，使安全行为编码为低秩覆写而非底层重塑60% similarUnverified测量协方差矩阵 R 越大滤波器越不信任该传感器读数，越小则越快响应观测变化59% similarUnverified4-bit quantization introduces some accuracy loss, but in many practical tasks its performance is very close to that of the original model.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/949991API
curl https://kongchang.com/api/v1/knowledge/claims/949991MCP
get_claim(id=949991)