Explainers
Unread — 5
paper · Oct 2025
The Art of Scaling Reinforcement Learning Compute for LLMs
Fits predictable compute-performance curves for RL, separating what raises the ceiling from what only buys efficiency.
arXiv 2510.13786 paper · Dec 2025Training Superintelligent Software Agents through Self-Play SWE-RL
Agents generate and solve their own progressively harder bugs, removing the human-curated task bottleneck.
arXiv 2512.18552 code · Active 2026verl — RL post-training framework
The de facto open RL post-training stack: GRPO and PPO over FSDP, Megatron, vLLM and SGLang.
GitHub — verl-project/verl post · Apr 2025Welcome to the Era of Experience
Silver and Sutton argue the next leap comes from agents that learn from their own experience in the world, not from human text — the roadmap past imitation.
Google DeepMind paper · Jul 2019Superhuman AI for multiplayer poker
The Pluribus paper: self-play plus real-time search beat top human pros at six-player no-limit poker — a landmark in game-theoretic multi-agent AI.
Science 365