Unverified50% confidenceFactExact time
This study proposes a set of reconstruction distortion bounds that can be computed prior to training.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic研究员最早在2020年提出预训练损失与计算量之间存在幂律关系（扩展律）76% similarUnverified研究者已证明可以通过修改不到0.1%的训练数据实现LLM后门植入74% similarVerifiedDuring the post-training phase, an excessively low loss may correspond to overfitting, where the model memorizes the training data and loses generalization ability.72% similarUnverifiedMoving data quality inspection to before training can significantly improve experimental efficiency, avoiding the need to trace back to data after tens of hours of training only to discover policy anomalies.72% similarUnverified诊断训练瓶颈的核心效率差异来自于是否提前建立了性能基线（baseline），如已知模型健康状态下的MFU则可快速发现偏离71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1002264API
curl https://kongchang.com/api/v1/knowledge/claims/1002264MCP
get_claim(id=1002264)