Unverified50% confidenceOpinionExact time
After pre-training scaling laws encountered diminishing returns, RL took over and extended the new performance curve.
1
Sources
50%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
Unverified强化学习训练比监督学习更不稳定，将NVFP4激进量化引入RL训练会进一步放大固有不稳定性70% similarUnverifiedThere is still a gap between validating 100+ steps and supporting full production-level RL training, and stability, long-term convergence performance, and scalability to larger clusters require continued validation going forward.70% similarUnverifiedThe essence of choosing a simulation platform is a trade-off between RL training efficiency and physical fidelity70% similarUnverifiedAfter adding the DINO objective, CAPI's originally smooth training loss curve became noticeably more erratic, and training stability decreased.69% similarUnverified近期研究表明RL（尤其是RLVR）并没有给模型注入大量新知识，而更像是激发和对齐预训练阶段已具备的能力69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/929303API
curl https://kongchang.com/api/v1/knowledge/claims/929303MCP
get_claim(id=929303)