Codex trained a vision model from 25% to 97% accuracy

Signals Inbox·July 14, 2026·AI Agents

Codex took a Qwen vision model from 25% accuracy to 96.9% after building its own training environment and running the reinforcement learning loop. Counting colored stars is a small task, but the bigger signal is that a coding agent can now do much of the slow experimental work that normally keeps machine-learning researchers busy.

Q1What actually happened?

In its official technical post, NVIDIA says Codex with GPT-5.5 received a goal and a time budget, then built a visual counting environment, trained Qwen3-VL-2B-Instruct, and evaluated the results. Accuracy rose from 25.0% to 96.9%. A researcher still reviewed progress and steered the important choices.

Q2Why is a jump from 25% to 97% interesting?

On a four-choice task, 25% can be close to random guessing. Reaching 96.9% means the model went from barely solving the task to getting almost every example right. That is roughly a 72-point gain and nearly four times the original accuracy. The result is narrow, but the improvement is large enough to show that the training loop really worked.

Q3Did Codex train itself?

No. Codex was the researcher and operator, not the model being trained. It worked on Qwen3-VL-2B-Instruct, a separate vision-language model. Codex wrote code, configured NVIDIA’s tools, launched runs, checked metrics, fixed problems, and changed the experiment. Think of it as an AI engineer teaching another AI model.

Q4What did the agent do that researchers normally handle?

Quite a lot. It brought up the NeMo RL and NeMo Gym stack, handled dependencies and GPU resources, created the star-counting environment, produced a baseline, ran reinforcement learning, compared results, and proposed the next experiment. Those jobs are not glamorous, but they consume a large share of real machine-learning research time.

Q5Does counting stars prove AI can automate science?

Not yet. The task was controlled, easy to score, and far simpler than discovering a drug or training a major foundation model. The agent also used reusable skills that explained local rules, metrics, storage, and recovery steps. The meaningful part is not the stars. It is that the same agent completed a long chain of technical work without needing every step written out by a human.

Q6So why does this matter now?

Coding agents are starting to move beyond generating code snippets. They can stay inside a project, use GPUs, run experiments, read the results, and keep working toward a measurable goal. If this becomes reliable on harder tasks, researchers may spend less time fixing environments and babysitting training runs, and more time choosing the questions worth testing.