Topic
Capability gains obtained by extending RL, task environments and post-training compute on an existing base model, using stacks such as IndexShare, SAO and slime.
NVIDIA's LoGRA cuts average RL post-training memory by up to 45.7% by storing gradients as low-rank sketches, according to a write-up of its paper. With that saving, a 27B model trained stably for over 1,100 steps on one eight-GPU node where dense Adam runs out of memory.
Reality
A dev.to walkthrough of DeepSeek's GRPO puts a 70B PPO training loop over 600GB of VRAM before any sharding. The critic-free version deletes the most expensive network and pays for it in sampled completions.
Reality
A LessWrong experiment tapped a Qwen3-1.7B after block 23 and fed the result into layer 5 of the next token, then trained it to solve math while hiding the work from an LLM judge. The control failed the same objective.
Reality
Nvidia's Nemotron scored 30 of 42 at IMO 2026 writing proofs in ordinary prose, and the September 9 paper publishes the data, the code and the acceptance rule that decided which candidate proofs survived.
Reality
Z.ai's August 14 post claims post-training gains for coding agents, but the company's release notes still stop at GLM-5.1 and there is no API endpoint, model identifier or weight download.
Reality