A paper published on September 29, 2026 tests on-policy distillation with one query across mathematics, coding, instruction following, and agentic tool use.
Published on September 29, 2026, the paper tests on-policy distillation (OPD) using a single training query. It reports results in mathematics, coding, instruction following, and agentic tool use; the available summary does not quantify those results.
The work argues that algorithm improvements may benefit OPD more than adding training data, and suggests reducing the data required while retaining much of the reported gain. To assess the method, metrics, and limitations, consult the original paper and check its experiments and results; the available summary gives no figures or methodological details.