Method
Teachers are the frozen base model with privileged context; advantage per token = log π_teacher − log π_student on the student's own samples.
- Photo questions: photo + system "Answer from what is actually shown in the image, even if it differs from what is typical."
- Knowledge questions: photo + system "Answer with what is generally true in the real world, even if the photo shows something different."
(Round 1 used no photo here.)
- The student's user turn (photo, question and any cue) is identical for student and teacher.
- Data: Same items as R3c; family weights start at R3c's and adapt within bounds (plain 0.25-0.60, families 0.05-0.40). Only image pairs whose teacher answered every item correctly in 2 of 2 samples (280 conflict + 80 ordinary images).
- 80 steps, lr 3e-6 (5 warmup), 8 image pairs (16 prompts) × 4 rollouts per step, rollouts up to 4,096 tokens, 4× H100.
Evaluation (private test v3, natural images, n 1,186, thinking on, card sampling, 1 sample)
Table with columns: model, CV, CK, NV, NK, All-4, KT| model | CV | CK | NV | NK | All-4 | KT |
|---|
| Qwen3.5-9B | 80.2 | 67.7 | 84.0 | 90.7 | 43.5 | 93.0 |
| round 1 (VK-OPD-think) | 83.2 | 93.8 | 83.1 | 93.2 | 65.6 | 92.9 |
| this model | 80.8 | 94.3 | 81.9 | 91.5 | 61.2 | 93.8 |
All-4 vs the base: +17.7 [+14.8, +20.9] pp (pair bootstrap 95% CI). Second seed of R3d; seed 3 is Qwen3.5-9B-VK-OPD-think-R3d. Shared privately for comparison within round 3; not a release.