Introducing ConstrainedMimic ( — a control framework for #humanoid# robot #safety# that enables real-time constraint enforcement within #RL-based# motion tracking policies by leveraging whole-body kinematics and dynamics. Recent advances in reinforcement learning have unlocked remarkable whole-body agility for humanoid robots. However, ensuring safety and satisfying constraints—especially those introduced after training—remains a significant challenge for deploying safe and reliable systems. ConstrainedMimic addresses this challenge by combining ideas from operational space control and control barrier functions (CBFs). The framework enables enforcement of arbitrary runtime constraints while preserving the ability of the policy to track complex motions. Importantly, constraints can be imposed on both the kinematic reference motion and the underlying robot dynamics, providing a principled approach to safer, more robust, and more controllable humanoid behavior. As #PhysicalAI#, #humanoid# #robotics#, and #embodied# #AI# systems move from research environments into the real world, the ability to guarantee safety and respect operational constraints will become increasingly important — ConstrainedMimic is a step in this direction. 📄 Paper: 💻 Code: Coming soon Great work led by @danielpmorton . #PhysicalAI# #AISafety# #HumanoidRobotics# #EmbodiedAI# #ReinforcementLearning# #Robotics# @StanfordAILab @StanfordEng

Stellar performance from a 3B model. These results were achieved primarily through post-training refinements on Qwen2.5-Coder. The paper doesn't provide many details, but it appears they distill from RL ckpts and then do a final RL-based instruct RL. 🔗

“FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic LMs” This paper argues that scalable agent RL doesn’t need multiple rollouts of the same prompt, but can learn from each trajectory as soon as it finishes. They combine reward centering across independent prompts, sequence-level trust regions for stale trajectories, and sample-mean optimization so long failures don’t dominate training, creating stable critic-free learning from just one rollout per prompt. This moves from group-based RL that spends compute repeatedly sampling the same prompts to asynchronous RL where every rollout covers a new prompt, immediately becomes training data, and lets agents learn from more diverse experiences with less rollout compute.

KernelBench update: i went through a bunch of traces looking for models attempting to use ncu and getting blocked due to the compute provider disabling performance counter permissions (NVreg_RestrictProfilingToAdminUsers=1). this affected rtx pro 6000 (using a compute provider as my own compute was saturated at the time) h100 and b200 runs. im thinking of doing a bunch of reruns but this would cost me an arm and a leg. im very impressed with models ability to optimize kernels without all the information from ncu profiles. i also wonder if kernel-based rl from labs has ran into this issue and solved it before training, or if models just have this strange intuition about what patterns tend to perform well. before i do anything, i want to get some feedback from the community on this.

📺 Xiaomi Is Livestreaming a Production RL Run — Burn Rate Included @XiaomiMiMo is training its MiMo-V2.6 in public — not a demo, a live post-training dashboard where every step's reward, loss and benchmark movement is watchable in real time. One day in, the meter reads roughly $1.2M spent and 60B tokens consumed. As Zhihu contributor Kitt在进化 puts it: this isn't a training site, it's a money-burning site — about ¥4,000 a minute. His bigger point: this is rare, valuable data. Almost nobody publishes what frontier-scale RL actually costs. 1️⃣ What the dashboard actually shows Every training step exposes the internals most labs keep private. His reading guide: 🔹 Loss family: training loss, entropy (are answers diversifying or collapsing into one mode), gradient norm, and train/inference KL divergence. 🔹 Reward: a mix of test-case scoring and rubric-based reward. 🔹 dynsam (dynamic sampling): avg@n performance over repeated tries, plus pass-rate buckets — including samples never solved and always solved. 🔹 DeepSWE v1.1 as the headline probe (mini-swe-agent, avg@3): Pro at 63.7, Flash at 60.7 at the time of writing. 2️⃣ The cost ledger, out in the open From the dashboard he extracts the numbers the industry usually guesses at: 🔹 MiMo-V2.6 Pro: roughly $36 per million tokens of training. 🔹 MiMo-V2.6 Flash: roughly $8 per million tokens. 🔹 His rule of thumb: equivalent inference runs 30-50x cheaper than these training figures. 3️⃣ What's actually being trained The sample distribution is also public: about 1,500 prompts per step, ~70% of them coding tasks. The remaining third is split across general, visual, cybersecurity and chat tasks — which makes this, in effect, an agentic-coding-centric RL run with side dishes. 4️⃣ Why it's worth watching He notes a university lab livestreamed a training run days earlier, but at nowhere near this scale or frontier relevance. For anyone who wants to learn how production RL behaves — rewards, entropy, dynamic sampling, benchmarks moving step by step — this is a rare open classroom. His only complaint, half-joking: a livestream this good deserves a comment section. 🔗 Full Reading: #Xiaomi# #MiMo# #ReinforcementLearning# #PostTraining# #LLM# #OpenScience# #AI#

GLM-5.2’s shift from GRPO-style group-wise optimisation to critic-based PPO for long-horizon agentic RL is worth paying attention to. Over the past year, the common narrative around LLM RL has been fairly simple: PPO is heavy, critics are expensive, and GRPO is lighter because it removes the critic. This view makes sense for short-horizon RLVR tasks, such as maths, coding problems, and verifiable QA. In those tasks, the structure is clean. The same prompt can produce multiple candidate answers, a verifier can judge which ones are better, and the model can learn through group-relative comparison. But agentic RL is different. A coding agent fixing a bug may need to read the issue, search the codebase, edit files, run tests, fail, locate the problem again, and try another solution. Final success only tells us that the whole trajectory worked. It does not clearly tell us which intermediate step was crucial, which step was a detour, or which action had no immediate reward but enabled later success. This is where GLM-5.2’s blog becomes interesting. It mentions that long trajectories are compacted and split into multiple sub-traces. For the same prompt, different rollouts can produce different numbers of trainable traces, and their lengths can be highly uneven. In this setting, group-wise optimisation becomes less natural, while critic-based PPO can estimate token-level advantages and handle credit assignment across long, irregular trajectories more effectively. So this is not simply “PPO is back”, nor does it mean “GRPO has failed”. A better way to put it is: GRPO is better suited for comparing multiple answers, while PPO is better suited for evaluating each step in a long chain of actions. When models move from answer generators to action takers, the key RL question also changes: from “which answer is better?” to “how should credit be assigned across a long sequence of actions?”

Grok 4.5, based on our 1.5T V9 foundation model, with Cursor data added in supplemental training, is now in private beta at SpaceX & Tesla. Early evals show performance close to, perhaps exceeding Opus. RL is continuing to significantly improve the model, and the Grok Build harness gets better every day. Nice work by all those involved! Completely trained from scratch new models will be released by @SpaceX every month this year.

According to @FundaAI, GPT-6 used roughly 10x the training compute of the GPT-5 generation. "Based on our industry discussions, experiments, RL, synthetic data generation and supporting infrastructure can together consume several times as much compute as the main pre-training run, potentially as much as ten times." "Once TPU 8t and Vera Rubin ship in volume, the next pre-training acceleration can begin, and large-scale interconnect and optics benefit most in that cycle." $NVDA $GOOGL

Coding agents can now create their own training environments from existing software! Xiaomi's new CodeMidas system lets agents inspect open-source projects, extract coding tasks, build execution-based tests, and discard broken or exploitable tasks. It generated 5,545 verified RL environments from 3,185 projects across 23 programming languages. According to the paper, training MiMo-V2.5 on these tasks improved issue repair by 11.7%, whole-program construction by 17%, and Terminal-Bench performance by 8.5%. Instead of humans writing every RL task by hand, CodeMidas can continuously mine existing software for new, verifiable training tasks.