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

What it claims

The teacher gets a text hint saying where the relevant objects are. It does not get a better view of the image. That hint is enough for on-policy self-distillation to improve perception broadly, and the training data is nothing but synthetic counting scenes of colored shapes. The method was tested on three models: Qwen3.5-4B, Qwen3.5-9B and Qwen3-VL-4B. Average gains on the real-world perception suite (CVBench, V*, HR-Bench 4K/8K, ZoomBench, MME-RealWorld, BLINK) are +3.23, +1.07 and +1.29 points respectively. The headline 3.23 is the best case, from Qwen3.5-4B. The other two models gain about a point.

The motivation is a tradeoff the authors measured in crop-based methods. Vision-OPD, which gives the teacher a zoomed crop, improves Qwen3.5-4B by +8.90 on V* and +14.56 on ZoomBench. It also drops CountQA by 10.73. Where-OPD aims for gains that are smaller but not lopsided.

How it works

• Scene generation. Scenes are 1024×1024 canvases with colored geometric shapes from seven shape classes. Shapes are placed without overlap on a uniform background. Each scene state records category, radius, center (u, v) and background color for every object.
• Questions and hints. The only training task is counting ("How many green crosses?"). The answer and hint come directly from the scene state through a fixed template. Example hint: Scanning for green crosses: found one near (377, 400), found one near (213, 805). Counting: 2 total. No human labels, region proposals or external MLLM are involved.
• Training step.
• The student sees (image, question) and samples a response.
• A frozen copy of the model sees (image, question, hint) and scores the student's prefixes.
• The loss is the per-token Jensen–Shannon divergence between student and teacher distributions, averaged over response length.
• The teacher side has stop-grad, so only the student is updated.
• The hint is never used at inference.

flowchart LR
  Z[Scene state] --> I[Image + question]
  Z --> H[Spatial hint]
  I --> S[Student samples y]
  I --> T[Frozen teacher]
  H --> T
  S --> T
  S --> L[Token-level JS loss]
  T -->|stop-grad| L
  L -->|update student only| S

Evidence

All numbers below are gains over the base Qwen3.5-4B:

DocVQA is also listed among the improved benchmarks, but the fetched text gives no number for it.

The fetched text does not include a head-to-head table against Vision-OPD or the other OPSD variants. Only the Vision-OPD figures quoted above are given.

Limits

• Truncated text. The fetched text stops partway through the data-generation section. Dataset size, training steps, compute and hyperparameters do not appear, nor do ablations (for example, hint versus no hint, or hint format).
• Uneven transfer. Gains are much smaller on Qwen3.5-9B and Qwen3-VL-4B (about +1 average) than on Qwen3.5-4B. Behavior on other model families or larger scales is untested in what was shown.
• Narrow training signal. Training uses only counting on simple shapes over plain backgrounds. Whether other synthetic tasks help more is not shown here.
• No release details. A "project page" is mentioned, but no code, weights or licence are stated.
• Preprint status. This is a v1 preprint (dated 1 Oct 2026). Several of the baselines it cites are concurrent 2026 work.