Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes

Where-OPD. Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes. By Sophia Sirko-Galouchenko and four others, from Valeo AI, Sorbonne University, and two other institutions. Published on the first of October. Its code repository is worked out: about three hundred source files, with a readme and a licence.

On-policy self-distillation has recently become an effective way to improve reasoning in language models. A student is supervised by a frozen or EMA version of itself, and that teacher copy receives privileged information the student does not. Carrying the idea over to multimodal large language models has remained largely unexplored. Where it has been tried, the privileged information has been visual: an image crop corresponding to the question, used to sharpen fine-grained perception. Two limits come with that choice. The gains stay confined to tasks that benefit from visual zooming, and the crops have to come from somewhere, which means either human-annotated grounding data or an external teacher model.

Where-OPD changes what the teacher is handed. Instead of a crop, the teacher receives textual guidance that is spatially grounded: it identifies which visual elements in the image are relevant to the query. The teacher is still the model itself, now told where to look in words. The student sees none of that guidance. It gets the image and the question alone, and learns, on its own rollouts, to reproduce what the guided teacher does.

The guidance needs object identities and positions, and here the authors sidestep annotation entirely. Post-training runs on procedurally generated scenes, where every object's identity and spatial coordinates are available automatically because the generator placed them there. That makes the approach scalable and annotation-free. With the spatial guidance in hand, the teacher locates evidence and integrates it across multiple relevant regions of the image, rather than a single zoomed patch. The student, lacking the guidance, is trained to arrive at that same behaviour from the raw inputs.

Across multiple models, the approach consistently improves counting, document understanding and chart understanding benchmarks. The more telling result concerns transfer. Post-training used only synthetic scenes, yet the improvements carry over to real-world perception benchmarks. Averaged across CVBench, V star, ZoomBench, BLINK, HR-Bench and MME-RealWorld, the gain is three point two three points.

Here is the argument the results are built to support. Spatially grounded privileged information, passed through on-policy self-distillation, can induce perceptual capabilities broader than the task it was trained on. The training scenes were synthetic and the gains showed up on real images, outside both the task and the data distribution used for post-training. Telling a model where the evidence sits, in words and only during training, appears to teach it something about looking that survives the move to photographs.