Athanor: An Agent-Native System for Autonomous Post-Training Research

Athanor. An Agent-Native System for Autonomous Post-Training Research. By Yufa Zhou and seven others, from ByteDance, Duke University, and two other institutions. Published on the fifteenth of October. No code repository is linked to it.

Recursive self-improvement, AI systems building stronger AI systems, is one proposed route to superintelligence. It is hard to measure directly. The authors study a weaker form that can be measured today: frontier agents post-training a small, raw base model using public data. How well they do is treated as a test of their research potential. The test is run on real GPUs, with the agents doing the research rather than following a recipe.

Athanor is an open, agent-native system for autonomous post-training research, which the authors call AutoPostTrain. The agents get full research autonomy, low-friction engineering tools and efficient starter code, and then the research is theirs. They read datasets and evaluation results, build their own training data, write and modify training algorithms, tune hyperparameters, and analyse what happened to decide what to try next. Nobody hands them a plan. Each step of the research loop, from choosing data to judging a result, is the agents' own.

They work inside a sandbox, and they cannot use external teacher models. Around them, Athanor's services supply the rest. Evaluation is frozen. Jobs are managed. Data is decontaminated. A persistent research record keeps what has been done. And there is a held-out suite of tasks that the agents are unaware of, so the final score is not only on tasks they could see. The agents get freedom over the research. The system keeps fixed the things that decide whether the result means anything: how models are scored, what data is clean, and what is kept from view.

The campaign ran for one hundred and four hours. A team of four Claude Opus 5.5 agents, each on its own eight-GPU node and given no research advice, worked on Qwen3-4B-Base. On a six-task development suite, the base model went from fifteen point nine to sixty-seven point three.

On the held-out suite it went from nine point two to forty-eight point two. For comparison, the official Qwen3-4B-Instruct-2507 scores sixty-six point five on the development suite, and forty-seven point nine on the held-out one. To the authors' knowledge, this is the first demonstration that a team of frontier agents can autonomously post-train a raw base model into a cross-domain generalist that matches or surpasses the official instruct version on average.

The campaign's full record shows how they got there, and it was not by inventing a new paradigm. The agents recombined known methods and each other's models. They used supervised fine-tuning on public reasoning data, reinforcement learning with a rollout change that one of the agents introduced, merges of weights across agents, and rounds of self-distillation. Each ingredient is known. What the record shows is agents assembling them, and building on one another's models along the way.

The authors' reading is that, once execution is handled, today's agents already have the engineering and research ability to conquer this task. The weight of that sentence sits on its first clause. Athanor is the part that handles execution: the frozen evaluation, the managed jobs, the clean data, the record and the hidden test. With that in place, four agents and one campaign were enough to match an official instruct model on average. The authors offer the system as a foundation for scaling autonomous AI research.