Verified65% confidenceFactExact time
Research by Google DeepMind and other institutions in 2023 confirmed that a model's memorization behavior is closely related to the number of repetitions of training data.
3
Sources
65%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
纽约时报指控OpenAI隐匿证据，ChatGPT版权诉讼再升级
rss7/9/2026
Related Claims
UnverifiedResearch shows that under certain conditions (such as a particular piece of text appearing at very high frequency in training data), models may reproduce verbatim content from training data, and in 2023 researchers from Google DeepMind and the University of California, Berkeley successfully extracted original fragments of training data from ChatGPT.79% similarUnverified卡内基梅隆大学等机构研究证明大型扩散模型存在「训练数据记忆」现象，可在特定提示词下高度还原训练集中的版权图像77% similarUnverifiedSWE-bench上已有多个研究团队记录到「patch记忆」现象，即模型通过识别问题描述特征检索训练时见过的补丁71% similarUnverifiedAdaptive reasoning technical approaches include difficulty-aware routing, early-stopping mechanisms, budget-controlled training, and Mixture-of-Depths, with institutions such as Meta and Google DeepMind publishing related work between 2024 and 2025.71% similarUnverifiedUsing real pre-trained PLE performs significantly better than control groups using random memory and shuffled memory.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/494400API
curl https://kongchang.com/api/v1/knowledge/claims/494400MCP
get_claim(id=494400)