Split training improves vision and language reasoning accuracy
SPLIT-RL: Staged Perception-Language Reasoning Training with Claim-Level Advantages
Computer Vision and Pattern Recognition
Summary
Understanding questions about pictures needs two key skills: seeing the important parts in the image, and using language to figure out answers. The authors show that training these skills separately with tailored rewards, instead of all at once, makes the system better at both seeing and reasoning. Their method, called SPLIT-RL, breaks down visual claims to give clearer feedback during training. This results in more accurate answers when the model is tested.
What this means in practice
- •For ai application developers: Build more accurate AI systems that answer questions about images by separately training visual and language reasoning components.
- •For multimodal system engineers: Improve joint vision-language models by employing staged training to enhance perception and reasoning abilities independently.
Authors
Raja Kumar, Rajat Koner, Ritwick Chaudhry, Zhuowei Li, Nishant Sankaran, Yifan Xing
Abstract
Vision-Language (VL) reasoning requires a model to both extract relevant and accurate information from an image (visual reasoning, VR), and to infer the answer from it (language reasoning, LR). Reinforcement learning with verifiable rewards typically trains both through a single chain-of-thought with a final-answer reward. This gives every CoT token the same sequence-level advantage, failing to distinguish capability specific errors. We propose SPLIT-RL, a staged post-training approach that trains VR and LR in disjoint phases. Because a group's rollouts differ along one capability at a time, the group-relative advantage isolates it, and each phase is optimized using phase-specific reward. We further introduce Claim-Level Advantage (CLA-GRPO), which decomposes VR-phase rollouts into atomic visual claims and provides a fine-grained advantage at claim level based on visual-type group formation. Although trained in two phases, trained policy is evaluated like GRPO model, with a single CoT call at inference time. Under this protocol, SPLIT-RL improves average accuracy over GRPO by 1.4-6.1 points across Qwen3-VL models from 2B to 30B-A3B and InternVL3.5-8B. Evaluating each capability using an oracle based diagnostic shows that answer-only GRPO leaves perception unchanged, whereas SPLIT-RL improves both VR and LR.