RL Data ScalingBinfeng Xu
Agent Research · NVIDIA
I’m a researcher at Nvidia, co-optimizing RL systems for language agents – infra, algorithms, harness and data/env.
I built Polar, the first open RL framework bridging post-training and harness engineering. To scale RL data in the open, Skill2Env synthesizes environments with public Agent Skills. FlashREINFORCE presents an elegant RL recipe to reduce instability and false credit-assignment in long, async RL.
See below for recent publications on terminal/computer-use agent, post-training infrastructure and algorithm, vertical harness (eg. autoresearch) and continual learning.
Research
RL Data Scaling
Async RL Recipe
Agent RL Infra
Agent RL Infra
Agent Harness
Memory Efficient RL
Agent Harness / AutoResearch
Computer-use Agent
Multimodal LLM
RL Recipe
Agent HarnessBlog
Misc
I play all games by Hidetaka Miyazaki. I enjoy art, training large neural nets and petting cats.
