Думаю стоило бы закрепить

Meta-Reinforcement Learning with Zero-Shot Reinforcement Learning https://openreview.net/forum?id=XyGJJ4FPoX

https://www.kimi.com/blog/kimi-k3 https://github.com/MoonshotAI/Kimi-K3/blob/main/k3techreport.pdf https://huggingface.co/moonshotai/Kimi-K3

наверное самый важный релиз со времен qwen 3.5/3.6 https://t.me/researchim/3/6693

Understanding Reasoning from Pretraining to Post-Training https://arxiv.org/abs/2607.16097 https://www.alphaxiv.org/ru/overview/2607.16097

https://alexzhang13.github.io/blog/2026/harness/