A-Evolve-Training: Autonomous Post-Training of a 30B Model

摘要

Post-training a frontier model today is the work of a research team iterating over weeks: proposing data-mix and recipe changes, launching runs, reading evaluations, and deciding what to keep. We report a autonomous post-training system that runs this loop with no human in the loop, performing post-training of a 30B-parameter Nemotron model across four rounds on GPU clusters over multiple weeks. The autonomously produced model reaches a held-out score of $0.86$ against the top human submission’s $0.87$ on the public leaderboard of the NVIDIA Nemotron-Reasoning Challenge,11 1 Public leaderboard standing as of 2026-06-01, the date of this report. Challenge URL: https://www.kaggle.com/competitions/nvidia-nemotron-model-reasoning-challenge/overview. placing $8\text{th}$ of roughly $4000$ entries at the time of writing this report. Beyond the headline number, the loop detected that its own internal development metric had stopped tracking external performance on the visibly weakest reasoning domain—candidates pushed it to record highs without moving the external target—and revised its own search policy in response: it stopped asking for higher dev and instead asked for interventions that lowered the now-misleading proxy while improving the external target. We treat this as direct, auditable evidence that a scaled autonomous loop can produce discovery, not only optimisation: the loop did not merely optimise inside a fixed measurement frame; it detected that the frame had become misleading and changed what counted as evidence. We adopt the operational view that any autonomous research system worth the “recursive self-improvement” label must eventually be able to perform end-to-end post-training of a frontier-class model; the result reported here is one datapoint of that bar being cleared. We deliberately avoid framing this as a “first autonomous match” of human researchers. The claim we do make is narrower and auditable: to our knowledge, this is the first publicly reported autonomous post-training run at this scale —prior public demonstrations of autonomous ML research operate at roughly GPT-2-class ( $\sim$ 124M) budgets. As a separate scale-up, the same system has been applied to post-train the 120B and 550B Nemotron models end-to-end; with no public human baseline at that scale in the NVIDIA Nemotron-Reasoning Challenge, this run evidences only that the autonomous loop closes at 120B and 550B scale, not that its output is competitive with a human-authored recipe at that scale. We report it here as infrastructure evidence; the effectiveness claim is deferred until a comparable human anchor is available.