VAD: Visual evidence attribution for multimodal on-policy distillation¶
Isolated counterfactual target reconstruction. Candidate process-quality scores proxy evidence-present/removed views; no 4B/9B multimodal model reproduction.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Shanghai Jiao Tong University / Xiaohongshu / CUHK / Zhejiang University / Southeast University |
| First public release | 2026-07-30 |
| Author code | Yes: author implementation and model weights. |
| Local adapter or method key | vad |
| Local reproduction code | src/auto_research/post_training/ |
Paper method and evidence¶
Privileged-view matching mixes visual evidence, language priors and teacher bias. VAD runs the same frozen teacher with and without relevant evidence; the centered log-probability difference is a signed visual direction. Project the original correction onto this direction with a one-sided constraint, then rebuild the target around the student's current distribution. The full privileged teacher remains only a weak regularizer.
flowchart LR
A[Teacher with and without visual evidence] --> B[Centered evidence direction]
B[Centered evidence direction] --> C[One-sided projection]
C[One-sided projection] --> D[Student-anchored target]Original paper figure¶
Figure 5 shows the flow of evidence views and target construction. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Across six visual benchmarks, Qwen3.5-4B Full VAD averages 78.32 vs 75.92 for direct privileged teacher matching. At 9B it also beats Vision-OPD and visual-advantage weighting. No production A/B.
Local reproduction¶
Implements two frozen teacher forwards, centered log-probability differences, one-sided projection, supporting/refuting branch budgets, student-anchored targets, symmetric JSD and 0.1-weighted weak teacher regularization. Logs projection, evidence norm, budgets and correction alignment. GSM8K candidate seed 42 accuracy goes 0.1719→0.7969 (+363.64% from pre-training), reward 0.3124→0.8153 and reference KL 0.0092. Over 120 updates projection activates 100% with mean alignment 0.5145. This is candidate-policy mechanism evidence, not visual benchmark lift. Fixed-seed metrics. No Qwen3.5 training, image crops, visual top-100 token support or six official benchmarks.