#reinforcementlearning

October 17, 2026 For years, AI training has been dominated by verifiable tasks. Write code that...

Beyond Blind Spots: Uncertainty-Aware Cooperative Forecasting in Multi-Agent...

Part 7 of the RL series. In Blog 6, Werbos gave us backpropagation and Klopf gave us the idea of a...

My Among Us simulator let players stand on 3,314 tiles of floor. The 2018 game's own colliders allow 1,894, so the floor was 1.75 times too big, and 3.06 times in O2. In a seeded 40-game test the fix changed the crew's wins by 2, within noise; an earlier test saw 10 points.

My Among Us crew bots explain each move by splitting the decision score into one part per input, exact to about 1e-14 over 155,209 decisions. The captions built from it were still wrong: 45 of 61 moves captioned "fixing the reactor" went farther from the reactor.

Ten thousand games into a self-play run, my Among Us impostors found that sabotaging the reactor beat the crew. The crew now weighted its next task 4.1 times as much as the fix, and fixed 0 of 7 sabotages. Paying for each fix flipped that to 7 times toward the fix, and 30 of 34.

An RL environment is a sandbox where an AI agent attempts tasks with tools and a verifier scores each try. Who builds them, who buys them and what they cost.

MiMo‑V2.6: Scaling Reinforcement Learning Turns Omni‑Modal Foundations into Self‑Improving Agents ...

Experience admission in multi-worker RL: P1–P4 and a preregistered falsification protocol. White-box only; black-box out of claim.

A deep dive into the paradigm shift from pre-training scaling laws to test-time compute. We deconstruct the mathematical derivation of DeepSeek-R1's Group Relative Policy Optimization (GRPO), critic-free architecture advantages, emergent self-reflection in long reasoning traces, and a complete, reproducible hands-on implementation.

Qwen3.8-27B, 600 steps, 12 reward channels: CISPO, REPO-R and prompt filtering lifted holdout from −0.47 to +1.80, and an advantage floor silently zeroed 96% of the signal.

Xiaomi's MiMo-V2.6-Pro just topped Artificial Analysis's Intelligence Index among open-weight models, tied with Grok 4.7, at a fraction of the cost of proprietary rivals. Xiaomi also open-sourced the full training report, 7,000+ RL environments, and livestreamed the entire run.