Unverified50% confidenceFactExact time
After RL training, a model's output diversity decreases noticeably and fixed small habits (ticks) form.
1
Sources
50%
Confidence
Long-term
Relevance
9/15/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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73% similarVerifiedThe marginal returns of simply increasing model parameters and training data are diminishing, a phenomenon known as 'Scaling Law slowdown.'70% similarUnverifiedExperiments show that after only a few generations of self-training, the variance of model outputs drops sharply and the text tends toward homogeneity69% similarVerifiedDuring LoRA fine-tuning, the original model parameters are frozen and only the inserted low-rank matrices are updated, significantly reducing VRAM requirements and training costs.69% similarUnverifiedDuring model collapse, the model successively experiences a reduction in output diversity, loss of minority categories and edge knowledge, and ultimately produces meaningless or highly repetitive outputs.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/923720API
curl https://kongchang.com/api/v1/knowledge/claims/923720MCP
get_claim(id=923720)