BREAKING
NVIDIA AI Agent Trains Vision Model
0
%
Baseline
0
%
After training
The Autoresearch Loop
1
Configure environment
↓
2
Train model
↓
3
Evaluate results
↓
4
Propose next test
NeMo Gym vs NeMo RL
NeMo Gym
Apache 2.0
●
Evaluate and RL-train models
●
Scales via Ray to thousands of envs
NeMo RL
Apache 2.0
●
Post-training RL library
●
Supports GRPO and Qwen3
A proof of concept, not a benchmark
Toward agentic RL research
AI NEWS BLITZ
NVIDIA shows a coding agent that trained a vision model almost on its own.