V-Thinker-Lite: Self-Supervised Interactive Reasoning Without Reinforcement Learning

TL;DR: What if we could get V-Thinker-like interactive reasoning—but without expensive reinforcement learning or point-level supervision, using only unlabeled images and clever self-supervised tasks? The experiment would pretrain an interactive vision-language model using transformer-based self-supervised objectives before fine-tuning on downstream reasoning tasks.

Research Question: Can self-supervised transformer objectives, such as masked image-caption or region-prediction tasks, serve as an effective substitute for RL and point-level annotation in training interactive visual reasoners?

Hypothesis: A model pre-trained with self-supervised “interactive” objectives will acquire remarkable visual attention and region-reasoning skills, approaching or exceeding RL-based V-Thinker in sample efficiency and transferability.

Experiment Plan: - Pretrain a V-Thinker-style architecture using only self-supervised tasks (e.g., masked region prediction, pseudo-interactive region queries) adapted from S3L, CTFusion, and related works.

• Fine-tune the resulting model on VTBench and other interactive reasoning tasks.
• Compare accuracy, robustness, and efficiency against reinforcement learning and/or point-supervised baselines.
• Run ablation studies to isolate the contribution of each self-supervised pretext task.

References: ['Qiao, R., Tan, Q., Yang, M., Dong, G., Yang, P., Lang, S., Wan, E., Wang, X., Xu, Y., Yang, L., Sun, C., Li, C., & Zhang, H. (2025). V-Thinker: Interactive Thinking with Images.', 'Guo, H., & Liu, W. (2024). S3L: Spectrum Transformer for Self-Supervised Learning in Hyperspectral Image Classification. Remote Sensing.', 'Du, K., Fang, L., Chen, J., Chen, D., & Lai, H. (2024). CTFusion: CNN-transformer-based self-supervised learning for infrared and visible image fusion. Mathematical Biosciences and Engineering: MBE.']