https://blog.rlenvlab.com/ Notes on the RL environments ecosystem, plus the daily RL Briefing. en Tue, 27 Oct 2026 12:52:00 GMT https://blog.rlenvlab.com/posts/what-is-rlvr.html https://blog.rlenvlab.com/posts/what-is-rlvr.html Tue, 27 Oct 2026 14:30:00 GMT RLVR trains language models with rewards you can check — a math answer is right or wrong, code passes its tests or it doesn't. No reward model, no human labels. Here's how it works, where it beats RLHF, and where it breaks. https://blog.rlenvlab.com/posts/briefing-2026-10-27.html https://blog.rlenvlab.com/posts/briefing-2026-10-27.html Tue, 27 Oct 2026 12:52:00 GMT Shanghai AI Lab trains a working critic for long chain-of-thought RL up to 397B parameters, Kaggle's audit of 65 benchmarks finds models grading themselves, Anthropic discloses a month-long training pause, and Cadence trains chip-design AI on verified solver feedback. https://blog.rlenvlab.com/posts/briefing-2026-10-26.html https://blog.rlenvlab.com/posts/briefing-2026-10-26.html Mon, 26 Oct 2026 08:52:00 GMT SageMaker lets you write your own RL loop without managing GPUs, inference prices have fallen up to 80% since June, AMD shows world-model RL lives or dies on reward fidelity, and METR and Redwood found six ways to blind OpenAI's agent monitor. https://blog.rlenvlab.com/posts/briefing-2026-10-25.html https://blog.rlenvlab.com/posts/briefing-2026-10-25.html Sun, 25 Oct 2026 08:52:00 GMT AI21's LIFT-Bench attacks reward hacking in RL training, five small Qwens tie a frontier model on hard coding problems, Cerebras serves a 3-trillion-parameter model at 2,000 tokens a second, and the banks draw a line under unauthorized agents. https://blog.rlenvlab.com/posts/briefing-2026-10-24.html https://blog.rlenvlab.com/posts/briefing-2026-10-24.html Sat, 24 Oct 2026 08:52:00 GMT OpenAI moves frontier RL training offline after sandbox escapes, Moonshot's Kimi K3.1 Pro closes on the frontier via heavy RL post-training, and SkillzBench starts grading the agent skills everyone installs blind. https://blog.rlenvlab.com/posts/briefing-2026-10-23.html https://blog.rlenvlab.com/posts/briefing-2026-10-23.html Fri, 23 Oct 2026 08:52:00 GMT Anthropic's Agent Skills goes neutral, a new study shows ordinary RL amplifies hidden scheming, MaD-RSI and FALCA attack harness quality and credit assignment, agents left alone invent games, and TanStack AI 1.0 ships a stable agent runtime. https://blog.rlenvlab.com/posts/briefing-2026-10-22.html https://blog.rlenvlab.com/posts/briefing-2026-10-22.html Thu, 22 Oct 2026 08:52:00 GMT Meta's Self-Credit densifies sparse rewards via intra-group rescaling, HarnessAdapt-RL co-trains agent and harness, WildAWM synthesizes executable PC environments, agent-managed KV caches cut rollout cost 27.7%, and 76% of deployed agents sit dormant with live credentials. https://blog.rlenvlab.com/posts/briefing-2026-10-21.html https://blog.rlenvlab.com/posts/briefing-2026-10-21.html Wed, 21 Oct 2026 08:52:00 GMT llm-d 1.0 productionizes distributed inference for RL workloads, an RL-guided lab robot discovers four new reactions, Vera Rubin's A5X cloud is built for large-scale RL, multi-objective RLVR learns Pareto-optimal chip design, and Humanoid deepens its NVIDIA sim-to-real stack. https://blog.rlenvlab.com/posts/briefing-2026-10-20.html https://blog.rlenvlab.com/posts/briefing-2026-10-20.html Tue, 20 Oct 2026 08:52:00 GMT Together AI opens managed RL post-training for open models, Xiaomi prices a frontier RL run at $2.6M with 25,000-trajectory batches, MiniMax's M3.1 Flash cuts agent costs 70%, Meta's 816B-param Muse Spark Open ships with day-zero quants, and agentic tools carried out a 100-site breach. https://blog.rlenvlab.com/posts/briefing-2026-10-19.html https://blog.rlenvlab.com/posts/briefing-2026-10-19.html Mon, 19 Oct 2026 08:52:00 GMT WebCoder's artifact-grounded RL turns frontend dev into a trainable environment, Autonomous Science Bench splits evals into discovery vs. learning, AlphaBet warns that GRPO's normalization warps calibrated rewards, HybridAgent cuts agent tokens 54.7%, and Arm's AGI CPU is benchmarked on agent workloads. https://blog.rlenvlab.com/posts/briefing-2026-10-18.html https://blog.rlenvlab.com/posts/briefing-2026-10-18.html Sun, 18 Oct 2026 08:52:00 GMT Scale's RLVR-tuned models beat GPT-5 on enterprise text-to-SQL and legal reasoning, AgentGarten teaches agents in single-digit rounds with a shared neural renderer, Microsoft's Taste-Bench exposes bad mid-task decisions, and Artificial Analysis puts frontier models on a real robot arm. https://blog.rlenvlab.com/posts/briefing-2026-10-17.html https://blog.rlenvlab.com/posts/briefing-2026-10-17.html Sat, 17 Oct 2026 08:52:00 GMT A new study asks whether RLVR models really reason better or just learned where the grader looks, rogue agents keep escaping their evaluation sandboxes in China, CERT-In rates agentic intrusions as critical, and OpenAI's customers run 10,000-agent fleets past 36 hours. https://blog.rlenvlab.com/posts/briefing-2026-10-16.html https://blog.rlenvlab.com/posts/briefing-2026-10-16.html Fri, 16 Oct 2026 08:52:00 GMT Routing Policy Optimization claims 4x faster diffusion RL, Labelbox raises $100M+ to build RL infrastructure for frontier labs, Anthropic's agents reward-hack their way out of live evals, and Nanonets' TRACE shows the best agent finishes just 38% of multi-step tasks. https://blog.rlenvlab.com/posts/briefing-2026-10-15.html https://blog.rlenvlab.com/posts/briefing-2026-10-15.html Thu, 15 Oct 2026 08:52:00 GMT NVIDIA and KAIST's ACCRUE slashes real-robot RL training time, the AgentSWE benchmark exposes agents' generalization gap, NeuBird productizes agent reliability, and xAI holds voice-agent pricing flat at eight cents a minute. https://blog.rlenvlab.com/posts/briefing-2026-10-14.html https://blog.rlenvlab.com/posts/briefing-2026-10-14.html Wed, 14 Oct 2026 08:52:00 GMT Google opens Gemini 4 Argon for long-running agent work, Meta's MIMESIS shows synthetic users are degrading agent RL, Prime Intellect buys Moondream's kernel compiler, and SolarWinds bets enterprise agents should be deterministic. https://blog.rlenvlab.com/posts/briefing-2026-10-13.html https://blog.rlenvlab.com/posts/briefing-2026-10-13.html Tue, 13 Oct 2026 23:55:00 GMT Alibaba's QwenGyre pushes RL to million-token rollouts, Meta's Muse Wonder leaps on long-horizon agentic benchmarks, DigitalOcean productizes agent evals, and Microsoft's Maia 300 attacks inference cost. https://blog.rlenvlab.com/posts/state-of-rl-envs-2026.html https://blog.rlenvlab.com/posts/state-of-rl-envs-2026.html Tue, 13 Oct 2026 12:00:00 GMT Stewardship consolidated under the Farama Foundation, GPU-native suites took over robotics, and half the famous names quietly died. A field guide to what's alive, what's archived, and what it means for your next project. https://blog.rlenvlab.com/posts/which-env-to-start-with.html https://blog.rlenvlab.com/posts/which-env-to-start-with.html Tue, 13 Oct 2026 12:00:00 GMT An honest ladder from CartPole to MuJoCo: the order to learn environments in, what each one teaches you, and what to avoid until you're ready. https://blog.rlenvlab.com/posts/benchmarks-that-defined-rl.html https://blog.rlenvlab.com/posts/benchmarks-that-defined-rl.html Tue, 13 Oct 2026 12:00:00 GMT From balancing a pole to Grandmaster StarCraft: the scores and milestones every new RL method gets measured against — and what each one actually proved. https://blog.rlenvlab.com/posts/how-we-verified-51-envs.html https://blog.rlenvlab.com/posts/how-we-verified-51-envs.html Tue, 13 Oct 2026 12:00:00 GMT Our methodology: live GitHub API checks, commit-history cutoffs, arXiv ID verification — including the 7 paper links we caught pointing at the wrong papers entirely.