Breaking

NVIDIA Demonstrates AI Coding Agent Autonomously Training a Vision Model to 96.9% Accuracy

July 14, 2026 at 12:03 EDT

AI Agents
Foundation Models
Open Source

NVIDIA has shown a coding agent that, given only a goal and a time budget, built a training environment and taught a small vision-language model to count colored stars, lifting its accuracy from 25% to 96.9% with limited human steering. The demonstration, published on July 14, 2026, positions the company's open-source NeMo tooling as a foundation for more autonomous machine-learning research.

July 14, 2026 · NVIDIA NeMo

A Coding Agent Taught a Model to See — Almost On Its Own

Given only a goal and a time budget, an AI agent built its own training environment and lifted a compact vision-language model's star-counting accuracy from 25% to 96.9% — with limited human steering.

96.9%

Final accuracy on the counting task

Parameters — Qwen3-VL-2B, a compact vision-language model

1,000s

Concurrent environments via Ray parallelism

Baseline vs. after agent-run training

25%

Baseline

96.9%

After training

Nearly 4× the baseline accuracy on counting colored stars in images.

The autoresearch loop

Configure

Build the environment

→

Train

RL post-training

→

Evaluate

Measure results

→

Propose next

Agent's own follow-up

Built on NeMo RL and NeMo Gym — open-source under Apache 2.0, "battle-tested in production Nemotron training."

Why it matters

Ties evaluation directly to RL training, with agent harnesses and sandboxed execution closing the loop between measuring a model and improving it. Open licensing lowers the barrier to reproducing self-directed loops.

The caveats

Counting stars is a deliberately simple proof of concept, not a frontier benchmark. GPU and setup overhead remain, and how well the approach generalizes to complex, real-world tasks is the open question.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.