Unverified50% confidenceFactExact time
修复前模型每约29次下一token预测中约有1次的选择与数学上应有结果不同
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
First Seen
Sources
三行代码修复：Tesla P100在llama.cpp中精度提升2300倍
redditr/LocalLLaMA7/12/2026
Related Claims
Unverified马尔可夫头仅需记录前一个token状态即可推导下一token修正概率，参数量通常不超过原模型的0.1%71% similarUnverifiedMultiple independent studies in 2023 showed that some top models scored extremely high on standard mathematical reasoning tests but performed significantly worse on equivalent rephrased versions, indicating that the models learned test patterns rather than genuine reasoning ability.67% similarUnverified该研究发现同一模型不同部署时间点在代码生成和逻辑推理任务上通过率可能出现10%至20%的显著波动，且方向不总是单向改善67% similarUnverified同一个基座模型在不同系统提示词下的表现差异可以高达30-50个百分点65% similarUnverified在8个模型均随机猜测的假设下，单次二元预测中至少有一个押中的概率约99.6%65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/613155API
curl https://kongchang.com/api/v1/knowledge/claims/613155MCP
get_claim(id=613155)