V-Zero: Answer-Label-Free On-Policy Distillation with Contrastive Evidence Gating for Fine-Grained Visual Reasoning
V-Zero improves fine-grained visual reasoning without annotated answer labels. The student model samples on-policy reasoning trajectories from the full image, while a teacher model replays the same trajectories with paired positive and negative visual evidence views. By contrasting teacher support under the task-relevant crop and an irrelevant crop, V-Zero estimates how well each trajectory is grounded in visual evidence and uses this signal to gate dense token-level distillation. The resulting training objective keeps standard full-image inference unchanged while providing answer-label-free supervision for localized visual reasoning.
The V-Zero training implementation is included under verl/, with the public launcher in scripts/run_vzero_qwen35_vl_fsdp.sh. It configures the on-policy distillation recipe with teacher replay, positive/negative evidence crops, and evidence-gated token distillation.
Install the training package from the repository root:
uv pip install -e .See scripts/README.md for data schema, environment variables, and launch examples.
- Release training code
- Release data preparation scripts
- Release evaluation scripts
- Release model checkpoints
- Add detailed reproduction instructions
This project is released under the Apache License 2.0.