A new paper on arXiv examines how vision-language models (VLMs) improve their reasoning abilities through post-training with supervision from stronger models. The abstract identifies two common strategies: supervised fine-tuning (SFT), which trains on teacher-generated outputs, and on-policy distillation, which is implied by the title to train on the teacher's behavior rather than its explicit answers.
The title, "Do As I Do, Not As I Say," suggests a key difference: SFT may learn from what the teacher says (its final outputs), while on-policy distillation learns from what the teacher does (its reasoning process). The abstract is truncated, so the full findings are not available, but the framing points to a comparison between these two approaches.
If the title reflects the paper's conclusion, on-policy distillation may better inherit the teacher's underlying reasoning, potentially leading to more robust VLM improvements. However, without the full abstract, the specific results and evidence remain unclear from this source alone.