#reinforcement-learning

#…

11 posts

Microsoft Turned Database Grading Into RL Rewards—And It Worked

Microsoft's ThinkingBox converted database correctness checks into RL rewards, tripling repeatable success for agent tasks—but sharpening came at a cost.

#agents #reinforcement-learning #benchmarks #grounding #reliability

Hugging Face Just Solved the RL Environment Registry Problem

Hugging Face turned its dataset hub into a universal RL environment registry. No new infra, no sign-ups—just tags. Harbor, Verifiers, OpenEnv, and NeMo Gym environments all land in one filter.

#reinforcement-learning #agents #huggingface #infrastructure #benchmarks

LoRA, a Storage Bucket, and a Proxy: How HuggingFace Jobs Got Async GRPO Working Across Machines

HuggingFace shows how to split RL training and inference across separate Jobs with no shared NCCL—just a rank-1 adapter, a FUSE-mounted bucket, and smart routing. Training time drops from 3h27m to 53m.

#reinforcement-learning #lora #distributed-training #vllm #infrastructure

Granite 4.2: IBM's Open Reasoning Models With a Five-Phase Pre-Training Curriculum and Multi-Stage Agentic RL

IBM just dropped Granite 4.2—3B, 8B, and 30B reasoning models trained from scratch on 15T tokens, with thinking/non-thinking modes and a staged RL pipeline that teaches agents to code, search, and use terminals.

#llms #reasoning-models #reinforcement-learning #open-source #agents

LFM2.5-2.6B: Liquid AI ships the fastest local agent yet

Liquid AI's new 2.6B model beats competitors 4x its size on tool use and instruction following, runs at 220 tok/s on a laptop, and was trained inside real agent harnesses with RL.

#agents #edge-ai #local-models #tool-use #reinforcement-learning

The Simulation Stack is Now the Embodied AI Stack

NVIDIA, DeepMind, and Disney just open-sourced Newton—a GPU-accelerated physics engine that treats simulation as infrastructure. If you're building physical AI, your bottleneck just shifted.

#physical-ai #simulation #robotics #reinforcement-learning #infrastructure

LeRobot v0.6.0 closes the loop: world models, reward signals, and six new benchmarks

LeRobot's biggest release yet ships world-model policies that imagine before acting, reward models that know when robots succeed, and a unified eval suite across six simulation benchmarks.

#robotics #world-models #vla #reinforcement-learning #datasets

North Mini Code: Cohere's First Real Agent-Coding Model

Cohere just released a 30B MoE model trained specifically for agentic software engineering. It's Apache 2.0, beats models 4× its size, and actually works across multiple agent harnesses.

#agents #code-models #reinforcement-learning #moe #cohere

OpenEnv Goes Full Open: Why the RL Training Layer Just Got a Governance Committee

Hugging Face, Meta PyTorch, Nvidia, and a dozen others just formed a committee to govern OpenEnv—the protocol layer trying to make agentic RL training actually interoperable.

#agents #reinforcement-learning #open-source #infrastructure #governance

vLLM V1 and the Correctness-First Paradigm in RL

ServiceNow AI's deep dive into vLLM's V1 upgrade reveals why getting base correctness right matters more than chasing incremental RL gains—a lesson in engineering priorities.

#vllm #reinforcement-learning #inference #ml-infrastructure #correctness

Ecom-RLVE: Training Conversational Agents in Verifiable E-Commerce Sandboxes

Hugging Face just dropped Ecom-RLVE, a reinforcement learning framework that trains e-commerce agents in realistic but controllable environments. This is how we move from chatbots to actually useful shopping assistants.

#reinforcement-learning #agents #ecommerce #evaluation #huggingface

Loading…