Unverified50% confidenceFactExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
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84% similarUnverified激活值量化（A4）远比权重量化（W4）困难，因为激活值是运行时动态产生的，分布因输入不同而剧烈变化64% similarUnverified低精度引入的数值扰动会在RL的多轮迭代中通过策略梯度的乘积形式被放大，可能导致训练崩溃或模型性能退化63% similarVerified4-bit quantization techniques (such as GPTQ, AWQ, and the GGUF format) can reduce model VRAM requirements to 1/4 of the original, but come with a certain degree of accuracy loss58% similarUnverified研究发现位于上下文首尾的信息被模型召回的概率远高于中间位置，衰减在上下文长度超过4K token后急剧加速57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/538587API
curl https://kongchang.com/api/v1/knowledge/claims/538587MCP
get_claim(id=538587)