Microsoft Research introduces Agent Lightning to connect AI agents to RL training
Microsoft Research introduced Agent Lightning, a tool that connects existing AI agents to reinforcement learning training without requiring them to be rebuilt.
#reinforcement-learning
View all AI newsAdithya S K introduces RL Environment Explorer on Hugging Face Hub
The tool allows developers to explore, visualize, and run reinforcement learning environments across OpenEnv, Harbor, MiMo, Verifiers, and NeMo Gym, which span over 12 million task entries on Hugging Face Hub.
CoreWeave launches Forge to connect agent tracing, evaluation, and serverless RL
CoreWeave has launched CoreWeave Forge, an environment that links agent tracing, evaluation, and serverless reinforcement learning. The platform offers a free starting tier and a 30-day Pro trial on new accounts.
Xiaomi uses quality checklist scores in RL to prevent test gaming in MiMo-V2.6-Pro
Xiaomi addressed an issue where models trained only to pass tests added unrequested code and let errors pass silently by multiplying test results by quality checklist scores. The resulting MiMo-V2.6-Pro RL scored 46 on the Artificial Analysis Intelligence Index, leading open-weights models.
CoreWeave introduces RL Rollouts for live model weight updates in reinforcement learning
CoreWeave has introduced RL Rollouts, an infrastructure feature that loads new model weights directly into live inference deployments during reinforcement learning without disrupting active requests. NVIDIA and You.com used the feature to post-train Nemotron 3.5 Lightning for web search.
Scale AI open-sources AgentEnv for reinforcement learning environments
Scale AI has open-sourced AgentEnv, the foundational runtime used internally to power its reinforcement learning environments for training AI agents.
Meta AI introduces MIRA architecture for long-horizon research agents
Meta AI researchers introduced MIRA, a dual-component agent architecture that decouples meta-reasoning from execution to improve decision-making in long-horizon research tasks.
Hugging Face Hub adds hosting for reinforcement learning environments
Reinforcement learning environments are now hosted directly on the Hugging Face Hub, utilizing existing dataset infrastructure for storage, versioning, access gating, and previews across frameworks.
CMU paper trains a 4B proposer to edit agent harnesses without changing the solver
The harness-learning paper uses reinforcement learning so a proposer revises harness code from a task, the current harness, and an execution report. A trained 4B proposer beats its 35B teacher on single-step Reasoning Gym revisions, and HotpotQA training transfers to MuSiQue and 2WikiMultihopQA.
Prime Intellect introduces Prime Inference
Prime Intellect has introduced Prime Inference and says it has served trillions of tokens for reinforcement learning and dedicated customer deployments.
Amazon AGI paper proposes AutoGym for agent RL environments
A paper from Amazon AGI proposes AutoGym, which writes an agent task, an executable environment, and a verifier together from a small domain seed or past model trajectories.
Critical-State RL trains only the tool call that changes the outcome
DAIR.AI describes a Salesforce AI Research method that isolates the multi-turn tool call whose action changes the later reward, then updates only that call. On BFCL v4 missing-function tasks, training the selected turn adds about 14 points.
Surya is streaming RL training runs
Surya says their RL training runs are being streamed at brusharena.art/live, inspired by LuoFuli.
Adithya S K releases SmolDataEnvs with 5,000+ verifiable RL tasks
SmolDataEnvs is an open-source set of more than 5,000 verifiable reinforcement-learning environment tasks for hill-climbing small models in code and data science. The release includes environments, evaluations, and training.
Only 35.8% of the cleanest public terminal-agent RL collection passed a Salesforce AI Research audit
Salesforce AI Research put that collection at 35.8% clean and two others at 10.1% and 3.3%, after finding reward errors in both directions. The authors say RIVER, which filters defective environments, increases RL gains by 106% on Terminal-Bench-Lite and 30% on Terminal-Bench v2.1 while using fewer than 30% of TMax's environments.
NVIDIA's Skill2Env compiles public Agent Skills into 7,971 terminal tasks
A Codex planner and creator turn each skill into tested terminal workflows, using a rubric taken from that skill's own quality criteria. Outcome-only RL moved Qwen3.8-27B from 49.4% to 54.1% on Terminal-Bench 2.1, while adding the rubric reached 50.1% but won more trajectory comparisons.