Paper summary
Paper summary: InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency
InternVL3.5 adds Cascade RL (offline then online RL), a Visual Resolution Router and decoupled vision/language deployment, reporting up to +16% reasoning and 4.05× faster inference than InternVL3.
- Cascade RL runs offline RL first for stable convergence, then online RL for refinement.
- The Visual Resolution Router lowers the resolution of visual tokens where it doesn't hurt, for faster inference.