Multimodal Large Language Models often stumble on visual‑mathematical reasoning because their perception is too coarse, leading to visual hallucinations that cascade into reasoning errors.
Standard end‑to‑end reinforcement learning relies on sparse rewards, which cannot separate perceptual mistakes from logical missteps, making targeted perception improvement difficult. Fine‑tuning with perception‑enhanced Chain‑of‑Thought data, on the other hand, requires costly annotations and still suffers from hallucinations.
We introduce UniCAR‑RL, an annotation‑free RL framework that explicitly decouples perception optimization from reasoning optimization during training, allowing each capability to be improved in isolation.
UniCAR‑RL consists of three synergistic branches:
- Caption‑RL optimizes perception by using a verifier‑guided reasoning check;
- Reasoning‑RL performs logical inference based on a gold image description to stop error propagation;
- QA‑RL retains the native end‑to‑end alignment to ensure robust question‑answering performance.
Experiments show that, using only raw short‑answer data, UniCAR‑RL markedly boosts mathematical and visual reasoning accuracy and generalizes strongly across various architectures and scales.
This work demonstrates that decoupling perception and reasoning is a viable path toward reliable visual mathematics, offering a new direction for multimodal model optimization.
Review: UniCAR‑RL’s structured RL design effectively curtails the chain reaction from perceptual hallucinations to reasoning failures, achieving a balance between data efficiency and model versatility, and marking a significant advance in visual‑mathematical AI.