# Where-OPD

The work introduces Where-OPD, an on-policy self-distillation scheme in which the teacher additionally receives textual spatial hints naming question-relevant visual elements and their coordinates while the student sees only the image and question; training data is generated procedurally from geometric counting scenes with no human annotation or external teacher, and across Qwen3.5-4B, Qwen3.5-9B and Qwen3-VL-4B it improves counting, document and chart benchmarks and lifts the average over CVBench, V*, ZoomBench, BLINK, HR-Bench and MME-RealWorld by 3.23, 1.07 and 1.29 points despite training only on synthetic scenes.

The work introduces Where-OPD, an on-policy self-distillation scheme in which the teacher additionally receives textual spatial hints naming question-relevant visual elements and their coordinates while the student sees only the image and question; training data is generated procedurally from geometric counting scenes with no human annotation or external teacher, and across Qwen3.5-4B, Qwen3.5-9B and Qwen3-VL-4B it improves counting, document and chart benchmarks and lifts the average over CVBench, V*, ZoomBench, BLINK, HR-Bench and MME-RealWorld by 3.23, 1.07 and 1.29 points despite training only on synthetic scenes.