---
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.5-9B
tags:
  - vision-language
  - embodied-ai
  - navigation
  - tvrbench
---

# TVRBench Qwen3.5-9B VA-SFT

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](https://huggingface.co/TVRBench).