EmergentTwilight/MMCoT/projects/Visual-OPSD/docs - Storage Bucket

Paper figures

This directory contains the figures referenced from the project's README.md. They are the (PNG-rendered) originals used in the paper Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning.

File	Paper figure	Caption (short)
main_results.png	Figure 1	Headline results: per-benchmark accuracy, Δ-vs-teacher, and accuracy–latency Pareto curve. Visual-OPSD (green) matches or beats the VT teacher on 6/9 benchmarks and dominates the Pareto front at 74.03 % / 10.0 s.
pilot_intervention.png	Figure 2	Pilot intervention on the frozen ThinkMorph teacher: removing or corrupting the intermediate VT images barely changes accuracy across all 9 benchmarks, motivating the move from generative VT to distillation.
latency_comparison.png	Figure 3	Per-sample inference latency on a single H800 — Visual-OPSD is 14.3× faster than the VT teacher because it never invokes the diffusion pathway.
attention_analysis.png	Figure 5	V* attention-pattern analysis: once a VT is rendered, downstream reasoning attends almost exclusively to the generated image and largely ignores the original input — evidence that VTs concentrate task-relevant visual signal.

The original LaTeX / PDF sources for these figures live in the paper's supplementary material. They are bundled here as PNGs so the project README.md can render inline on GitHub.