Generalize USA beyond the Qwen3 model family

Investigate whether Update-aware Sharpness-Aware Minimization (USA) for cross-domain on-policy distillation and model merging generalizes to backbone architectures other than the Qwen3 family.

Background

The experiments evaluate only Qwen3-based teachers and students. Although this choice supports controlled comparisons across model scales and baselines, the reported evidence does not establish whether USA's update-aware perturbation geometry transfers to other backbone architectures. The paper explicitly leaves changing the model backbone for future work.

References

Every model belongs to the Qwen3 family, which we chose because it is openly released, stable, published at many sizes and widely enough used for our numbers to be placed against existing ones, so that the comparison across ten baselines stays fair; a change of backbone is left to future work.

— USA: Update-aware SAM for Cross-domain On-Policy Disitllation of Language Agents  (2609.34225 - Zhong et al., 28 Sep 2026) in Appendix, Section 1, Limitations (Appendix A)