Google OPST Self-Trains Gemma 4 From Topic Labels, +10.2 MMLU

Google and Ulm University researchers say a foundation model can post-train itself from nothing but a list of coarse topic labels, with the student generating its own questions, its own answers, and its own reasoning traces. On Gemma 4 E2B, the authors report MMLU accuracy climbing by 10.2 points, VQAv2 climbing by 6.9 points, and 'consistent gains on GSM8K, GPQA, and HellaSwag' alongside.

The method, which the Hugging Face paper page calls On-Policy Self-Training (OPST), reframes on-policy self-distillation so the loop never touches a human-curated prompt or answer. Seeded only by topic descriptors, the student generates problems; a privileged exponential-moving-average teacher produces reasoning rollouts conditioned on the self-generated answer; the student then learns from the teacher's token-level distribution. The abstract puts the broader claim plainly: 'foundation models possess rich latent capabilities that can be amplified without external task supervision or curated problem distributions.'

The authors motivate the setup with the data-labor arithmetic that drove it. Process-level annotation of 'a single mathematics corpus required about 800,000 human step labels', they write, and open instruction datasets 'rest on crowdsourcing efforts spanning over 13,000 volunteers'. They also report that lifts in self-trained domains leave untrained domains 'either remaining stable or also improving'. The paper lands on top of a dense run of post-training research on our fine-tuning tracker, which has logged 65 stories in ninety days.

The Hugging Face page carries no per-round compute figures, no scale for the topic lists, and no results on models smaller or weaker than Gemma 4 E2B.