Feed Entities Podcasts Deep Dives Developers About

[KongchangAI]

Real-time AI knowledge base. Every claim is traceable to its source, with data freshness under 24 hours.

Powered by 240+ first-hand AI sources

Knowledge

Feed
Hot Entities
Podcasts

Developers

CLI
MCP Server
REST API
About

© 2026 KongchangAI. The freshest, most accurate AI knowledge

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% similar Unverified2023年多项独立研究表明部分顶级模型在标准数学推理测试中分数极高，但在等价的重新表述版本上成绩大幅下滑，表明模型学习的是测试模式而非真正的推理能力67% similar Unverified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% similar Unverified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% similar Unverified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/613155

API

curl https://kongchang.com/api/v1/knowledge/claims/613155

MCP

get_claim(id=613155)