Recursive self‑improvement is a promising route toward superintelligence. This work studies a weaker, currently measurable form: frontier agents post‑training a small raw base model with public data to test research potential. We built Athanor, an open agent‑native platform that runs on real GPUs and supports autonomous post‑training (AutoPostTrain). The platform gives agents full research autonomy, low‑friction engineering tools, and efficient starter code. Agents read datasets and evaluation results, construct training data, write and modify training algorithms, tune hyper‑parameters, and analyze outcomes to decide the next experiment. All actions occur in a sandbox without external teacher models. Athanor supplies frozen evaluation, managed jobs, decontaminated data, a persistent research log, and a held‑out suite unknown to the agents.
In a 104‑hour campaign, four Claude Opus 5.5 agents, each on an eight‑GPU node and receiving no research advice, improved Qwen3‑4B‑Base from 15.9 to 67.3 on a six‑task development suite and from 9.2 to 48.2 on a held‑out suite, approaching the official Qwen3‑4B‑Instruct‑2507 scores of 66.5 and 47.9. The full record shows agents reached this level by recombining known methods and each other's models: SFT on public reasoning data, RL with a rollout change introduced by one agent, weight merges across agents, and rounds of self‑distillation. With execution handled, today’s agents already possess the engineering and research capability to conquer this task. We hope Athanor provides a foundation for scaling autonomous AI research.
Review