This is the intermediate Visual-Action SFT checkpoint used in Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration? It is provided to reproduce the rebuttal ablation that initializes offline Single-turn GRPO from the same SFT policy as Multi-turn GRPO.

Backbone: Qwen3.5-9B
Training: full-parameter Visual-Action SFT, epoch 5 / step 250
TVRBench canonical evaluation: 254/500 successful episodes (50.8%)
Output format: Action: <action name>
Thinking mode: disabled
Visual memory: concatenated multi-turn current/target image pairs with action responses

The benchmark code and final VA-SFT + Multi-turn GRPO model are available under the TVRBench organization.

TVRBench/tvr-qwen3.5-9b-va-sft

TVRBench Qwen3.5-9B VA-SFT

TVRBench/
tvr-qwen3.5-9b-va-sft