TL;DR: A new pipeline automatically builds 5,545 RL training tasks for coding agents using only source code itself, no issues or commit history needed, and it prioritizes quality over quantity. Title: CodeMidas: Scaling Agentic Coding RL Environments from Code Itself URL: Key points 🏗️ Auto-builds 5,545 tasks from 3,185 repos across 23 languages and 15 domains 🧪 Generates verifiers via execution-grounded tests, running the reference solution to record expected outputs 🛡️ Three-stage filtering: leakage checks, agent-solution agreement, and rollout difficulty filtering 📈 Big gains after RL training: DeepSWE +11.7pt, ProgramBench +17.0pt, Terminal-Bench +8.5pt 🔍 5k filtered tasks consistently beat 8k unfiltered tasks, proving quality beats quantity 🧠 Trained agents explore more and self-verify more, and these behaviors transfer to external benchmarks I like how simple and practical the core idea is: you don't need issue trackers or dev history to build RL environments, just the code. #CodingAgent# #ReinforcementLearning#

AGIBOT officially open-sourced AGIBOT WORLD 2026 Theme 3: Reinforcement Learning, a real-world dataset designed to advance reinforcement learning for embodied AI. Built for real-world robot learning, the dataset captures more than expert demonstrations — it captures successes, failures, and human corrections across the path from demonstration to deployment. • 14 real-world tasks • 11,430 trajectories • Fine-grained annotations for progress, mistakes & interference • Human-in-the-loop correction trajectories Learn from success. Learn from failure. Learn how to improve. 📖 Explore the project: 🤗 Access the dataset: #AGIBOT# #EmbodiedAI# #ReinforcementLearning# #PhysicalAI# #Robotics#

Today we're releasing @nvidia AlpaGym, our new open-source reinforcement learning (RL) framework for end-to-end autonomous driving. A key challenge for #Physical# #AI# is enabling policies to learn from the consequences of their actions. While supervised learning can teach a model to imitate behavior, robust autonomy ultimately requires learning through interaction with the environment. AlpaGym enables exactly that. Built on top of: - AlpaSim: our high-fidelity closed-loop autonomous driving simulator - Cosmos-RL: NVIDIA's distributed RL training and rollout infrastructure AlpaGym provides the glue that connects simulation, training, and driving policies into a scalable framework for post-training autonomous vehicle models in closed loop. With AlpaGym, researchers and developers can: ✅ Train end-to-end driving policies using reinforcement learning ✅ Run large-scale closed-loop simulations ✅ Experiment with new reward functions, policy architectures, and training strategies ✅ Benchmark models on public leaderboards 📖 Learn how it works: 💻 GitHub: 🏆 Open Challenges: - AlpaSim Closed-Loop E2E Driving Challenge: - Physical AI AV Reasoning Challenge: Learn more about the #Alpamayo# open platform: #PhysicalAI# #AutonomousDriving# #ReinforcementLearning# #Robotics# #OpenSource# #NVIDIA# #MachineLearning# @NVIDIADRIVE @NVIDIAAI

[𝐏𝐑𝐎𝐃𝐔𝐂𝐓 𝐔𝐏𝐃𝐀𝐓𝐄] 🇮🇹 🤖 #GENE01# designed from scratch and sent to batch production. Two scalable lower bodies. #PhysicalAI# deployed for motor control and world-action modeling. All in three months. @G_Bionics is running. 🎥 Enjoy the video below: blind #walking# and #running#, enabled by #reinforcementlearning# deployed across different robot platforms. 🙏 Kudos to the entire company for making these extraordinary results a reality. Stay tuned. The best is yet to come. #GenerativeBionics# #HumanoidRobotics# #AI# #Europe# #Innovation# #Startup# #DeepTech# #AMD# #PhysicalAI#

📺 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#

Most “Self-Evolving AI” Is Not Recursive Self-Improvement Yet AI may generate its own data, rewards, skills, and code. But if humans still define what counts as better and approve deployment, the loop is not closed. Zhihu contributor 一口鸟 (@alsaceym) argues that RSI’s hardest bottleneck is reliable verification. 1️⃣ Moving humans out of the loop RSI progresses through three stages: 🔹 Human in the loop: AI proposes changes; people approve them. 🔹 Human on the loop: data, rewards, and verifiers are automated; people supervise deployment. 🔹 Closed loop: the system generates, verifies, and applies improvements itself. Most “self-evolving” systems remain in the second stage. 2️⃣ What actually improves? Self-refine changes the current answer. Test-time training writes experience into weights. Agent evolution modifies prompts, tools, memory, skills, workflows, or Agent code across tasks. output → weights → the Agent itself Training-time RSI follows another ladder: 🔹 Zero-label: AI generates supervision. 🔹 Zero-data: AI also generates problems and curricula. 🔹 Auto research: AI chooses hypotheses, training recipes, and experiments. The system gradually takes over how to learn, what to learn, and finally how to improve learning itself. 3️⃣ Self-improvement can amplify mistakes A generator and verifier may share the same biases. Wrong outputs can produce biased evaluations, biased learning signals, and stronger errors. Even correct rewards do not guarantee stability: training can improve and later collapse. Self-play may also lose diversity or favor problems that are easy to reward rather than genuinely useful. Automation makes grounding more important, not less. 4️⃣ Verification is the real bottleneck Math and code are RSI-friendly because proofs, unit tests, and execution feedback provide clear signals. Open-ended Agent work is harder. A verifier must judge not only correctness, but novelty, usefulness, importance, and research taste. The next step may be evolving the verifier itself. But if the policy and evaluator change together, what keeps both aligned with reality? ✅ The boundary of true RSI The loop is expanding: answer → experience → learning signal → problem and curriculum → verifier True RSI requires improvement across every layer without losing contact with real objectives. Until evolving verifiers remain reliable without constant human grounding, recursive self-improvement is still an aspiration rather than an achieved capability. 🔗 Full analysis: #RecursiveSelfImprovement# #RSI# #SelfEvolvingAI# #AIAgents# #ReinforcementLearning# #AISafety#

📘 Introducing the Zhihu AI Cookbook (CN) Saved dozens of AI articles but still not sure where to begin? We’ve organized some of Zhihu’s best technical discussions into clear, structured learning paths. Over the past few years, Zhihu has become home to a wealth of in-depth AI writing from researchers, engineers and practitioners. The Zhihu AI Cookbook brings together some of the most useful and enduring pieces in one place. The first edition brings together 80+ selected posts from 68 contributors across three tracks: 🖥️ AI Infra 🧠 Reinforcement Learning for LLMs 🤖 Embodied AI It’s not another textbook. Think of it as a map connecting explainers, source-code walkthroughs, project retrospectives, interview insights and open-source tutorials. 🧭 How to use it: 1️⃣ Start with the introductory guide to understand the key concepts and prerequisites. 2️⃣ Choose the chapter that matches what you need right now. 3️⃣ Read each selection with a guiding question, then use the self-check prompts to test your understanding. 4️⃣ Follow the links to the original Zhihu posts for full derivations, implementation details and community discussions. 5️⃣ Put the ideas into practice with hands-on projects and open-source tutorials. Whether you’re exploring a new research direction, working through a project or preparing for technical interviews, the Zhihu AI Cookbook can help you find the right place to start and build a more complete understanding. 🤝 With open-source tutorials contributed by our content partner, @datawhale2018. 🌏 This edition is currently available in Chinese. An English version is coming soon. 🚀This is only the first edition. More topics and learning paths are on the way… 🔗 Explore the Zhihu AI Cookbook (CN): 💻 View it on GitHub: If you find it useful, give the repo a star and let us know what topic we should cover next. 🌟 #AI# #LLM# #EmbodiedAI# #AIInfrastructure# #ReinforcementLearning# #MachineLearning#

🤔 Filter by solve rate, reweight by advantage, adaptively tune domain mixtures — the RLVR literature is full of data policies promising better training. But most were introduced inside large training stacks where the base model, verifier, and optimizer settings all changed at once, making it hard to tell whether the data policy itself was doing the work. So the authors froze the GRPO recipe completely and built DataFlex-RL, an evaluation platform that isolates three intervention families — selection, reweighting, and mixture adaptation — and ran 156 training runs across 13 configurations and 12 seeds on Qwen2.5-7B. Uniform sampling alone lifts Overall accuracy from 42.01 to 49.77, a 7.76-point gain from GRPO training. Yet none of the eight selection methods, eight reweighting methods, or three mixture methods beat uniform sampling by a statistically meaningful margin — every confidence interval crossed zero. Title: DataFlex-RL: An Evaluation Platform for RLVR Data Policies URL: An honest negative result that cares more about "does it actually work" than pitching a flashy new method. #ReinforcementLearning# #MachineLearning#

🎮 Code agents thrive because code is executable and verifiable. This paper brings the same idea to game development to accelerate world model training. Title: Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models URL: 🧩 Overview The paper identifies a "verifiability bottleneck": fuzzy proxies like CLIP scores are noisy and gameable as reward signals. It proposes tapping game engines' collision, physics, and navmesh checks as a dense, trustworthy reward source instead. ❗ Problem it solves Spatial generation models (video, 3D, world models) lack reliable reward signals, so scaling compute doesn't reliably translate into real capability — what the authors call the "unverifiability tax." 🛠 Method RLHEV combines dense engine-verified rewards with sparse human acceptance judgments. AWoMo runs a Propose-Render-Verify-Repair-Review loop as a developer agent. UWDP unifies development traces into a single structured protocol for training data. 📊 Results On UnitySceneBench, full RLHEV beats the fuzzy-proxy baseline by +0.098. Cross-engine generalization (Unity to Godot) rises from 0.15 to 0.35. Downstream, it boosts D4RL Gym-MuJoCo by +48.43% and R2R navigation by +0.79%. #WorldModels# #ReinforcementLearning#

When you add annotations as rollouts in RL training, something counterintuitive happens — good policy actions start receiving negative advantages. OraRL names the problem and fixes it. Annotations as Rollouts: Efficient and Scalable Reinforcement Learning for Video MLLMs ❓ What goes wrong when you add oracle rollouts to GRPO training? 💡 "Advantage Inversion" occurs. Standard GRPO normalizes advantages across the rollout group using group mean and variance. Injecting a high-reward oracle rollout raises the group mean, which in turn causes policy rollouts that outperform the current average to receive negative advantages — a sign flip. Measuring 92,024 rollouts, naive oracle mixing inverted advantages in 42.5% of groups and 22.4% of rollouts. ❓ How does OraRL solve this? 💡 Five decoupled components. First, an oracle-free on-policy baseline is constructed without variance normalization (inversion is impossible by construction). A directional gain from the oracle-policy gap amplifies above-mean rollouts. The oracle advantage itself is calibrated against the strongest on-policy signal so it can't dominate the update. Sign-balanced advantage pruning retains equal numbers of positive and negative rollouts, compressing the residual flip rate to 0.3% and delivering a 1.48× speedup. ❓ How does it perform? 💡 Video-ORA-9B scores 73.1 on VSI-Bench, surpassing GPT-5 (55.0) and Gemini-3-Pro (55.1) by 18+ points. ReasonVOS segmentation improves +42.2 J&F and VideoHolmes gains +15.2 points over backbone. Training cost is just 2.2× SFT — under half the 4.9× overhead of GRPO with chain-of-thought. ❓ What about inference efficiency? 💡 No chain-of-thought is required, so P90 inference latency is 25.15 seconds vs 62.67 seconds for the CoT backbone. On data scaling, OraRL gains +5.2 points at 100k prompts versus +2.8 for GRPO, showing better sample efficiency throughout. #VideoMLLM# #ReinforcementLearning#