Before a chatbot can answer questions, follow instructions or write code, a lot of work happens after its initial training. AI labs call this post-training: teams of researchers choose data, run training experiments, check the results and decide what to try next, often for weeks.
A new preprint asks whether AI agents can now do that work themselves.
In one experiment, they could. Four AI agents, working for 104 hours without research advice from humans, turned a raw model into a general assistant that matched the developer’s own official version on average.
Four AI agents post-trained a raw model in 104 hours
The team, led by Yufa Zhou with colleagues at ByteDance, Duke University, the University of Chicago and Virginia Tech, built a system called Athanor. The paper was posted on arXiv on 15 October 2026 and hasn’t been peer reviewed.
They started with Qwen3-4B-Base, a raw 4-billion-parameter model from Alibaba’s Qwen family. A raw model has read huge amounts of text but hasn’t been taught to act as an assistant. Alibaba’s own polished version, Qwen3-4B-Instruct-2507, served as the benchmark to match.
Four copies of Claude Opus 5.5, Anthropic’s AI model, each got its own server with eight GPUs. Their instruction was a single sentence: raise the model’s score on a six-task test suite as high as possible. The tasks covered maths, logic puzzles, coding, general knowledge, following instructions and completing multi-step tasks in a simulated household.
The agents chose their own training data, wrote and changed their own training code, ran experiments and analysed the results. They could use public datasets and read each other’s notes, but they couldn’t call or download any other AI model to learn from.
Several safeguards aimed to keep the result honest:
- A frozen scorer the agents couldn’t change
- Training data cleaned of test questions
- A hidden second test suite measuring the same skills on different questions, which the agents never knew existed
- A full record of every experiment, note and result
The agents matched the developer’s official assistant
The agents matched the official model. On the six tasks they optimised, the model rose from 15.9 to 67.3, against 66.5 for Alibaba’s version. On the hidden test suite, it rose from 9.2 to 48.2, against 47.9. The differences from Alibaba’s model weren’t statistically meaningful, so the fair summary is that the two were level.
The strengths were different. The agents’ model did better on multi-step tasks and coding, but worse on maths and logic. Without the multi-step tasks, it trailed the official model by about four points.
It cost about $12 000. Most of that was GPU time. The model reached a score of about 60 after roughly $2 400; the last six points cost around $8 000.
The agents recombined known methods
The agents didn’t invent a new technique. They recombined known methods:
- Fine-tuning on a public collection of step-by-step reasoning examples
- Reinforcement learning, rewarding correct answers
- Merging models trained by different agents by averaging their weights
- Self-distillation, training the model on its own best answers
The most striking moment came from one agent. It noticed that its model often got stuck repeating itself in endless loops, and that standard reinforcement learning didn’t fix this. It worked out why and changed the training so that a quarter of the practice answers were generated the same way the test generated them. Loops dropped sharply, and its score jumped from about 50 to 57. Within hours, the other three agents had copied the fix.
The agents converged too fast
They converged too fast. Within the first hour, all four agents chose the same public dataset as their main training data, and never properly tested alternatives. The authors see this lack of diverse ideas as the clearest limit on progress.
They checked format, not correctness. The agents caught data that was too long or couldn’t be scored, but rarely checked whether the answers in their training data were actually right.
They needed a nudge to cooperate. The agents only began sharing notes and models after the human running the experiment told them, at hour 16, that the team’s best model would be the one scored. After that, every new best model combined work from several agents.
What the study doesn’t show
- It’s one run. One team of four identical agents, one small model, one experiment. It hasn’t been repeated.
- Not entirely from scratch. Much of the public data the agents trained on was originally generated by stronger AI models.
- Not the best possible model. Alibaba’s reasoning versions of the same model scored considerably higher.
- Small scale. Whether this works for the much larger models at the frontier of AI is untested.
- It’s a preprint, not yet peer reviewed.
AI is starting to do part of its own development
A long-running question in AI is whether systems will eventually speed up their own development, by doing the research that produces better AI. This study shows a narrow but concrete piece of that: agents doing work that AI lab teams do, end to end, at a modest cost.
Whether it counts as “research” is debatable. The agents mostly assembled known parts well, rather than discovering new ones. But that’s also what much real research looks like.
AI agents can now carry out a real part of AI development on their own, at least for small models.
They’re good engineers, but genuinely new ideas are still beyond them.