VASPO: Self-Distillation with Privileged Value Estimation
arXiv cs.CL · 15 Oct 2026, 03:17 AEST
VASPO introduces a method for self-distillation that enhances value-prediction accuracy in reasoning models. The approach combines privileged information with a self-teacher to improve performance on mathematical reasoning benchmarks.
Key facts
- On-policy self-distillation provides dense supervision for reasoning models via a privileged self-teacher.
- VASPO improves mean accuracy over PPO by 2.59 and 1.81 percentage points on Qwen3-4B and OLMo3-7B variants, respectively.
- VASPO achieves the highest mean accuracy among evaluated methods on Knights and Knaves.
Read aloud
VASPO introduces a method for self-distillation that enhances value-prediction accuracy in reasoning models. The approach combines privileged information with a self-teacher to improve performance on mathematical reasoning benchmarks. On-policy self-distillation provides dense supervision for reasoning models via a privileged self-teacher. VASPO improves mean accuracy over PPO by two point five nine and one point eight one percentage points on Qwen3-4B and OLMo3-7B variants, respectively. VASPO achieves the highest mean accuracy among evaluated methods on Knights and Knaves.
Sources in this cluster
- VASPO: Self-Distillation with Privileged Value Estimation
arXiv cs.CL · 15 Oct 2026, 03:17 AEST