Researchers have introduced Where-OPD, a method for improving multimodal large language models (MLLMs) through on-policy self-distillation guided by spatial information. The work appears as a paper on Hugging Face and is accompanied by a public project page.
On-policy self-distillation has become a popular way to strengthen language-model reasoning by having a student learn from a frozen or EMA copy of itself that receives privileged information. Applying this idea to MLLMs, however, has remained largely unexplored. Earlier approaches supplied privileged visual information such as image crops tied to a question, but their benefits were limited to tasks that gain from visual zooming and typically required human-annotated grounding data or external teacher models.
Where-OPD takes a different route: instead of privileged images, the teacher receives textual, spatially grounded guidance that identifies which visual elements are relevant to a query. The authors use procedurally generated scenes in which object identities and spatial coordinates are automatically available, making post-training scalable and free of manual annotation.
With this guidance, the teacher can locate and combine evidence from multiple relevant image regions, while the student learns to reproduce that behavior using only the image and the question. According to the paper, the approach consistently improves performance on counting, document, and chart understanding benchmarks across several models.
Crucially, the authors report that gains transfer beyond the synthetic training distribution: post-training on synthetic scenes alone yields a 3.23-point average improvement across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. The team concludes that spatially grounded privileged information can induce broader perceptual capabilities, enabling substantial synthetic-to-real transfer beyond the task and data distribution used during post-training.