Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document Question Answering via Reasoning-Free Alignment

arXiv cs.CL · 6 hours ago

Researchers developed Perception-RFT, a training framework using Group Relative Policy Optimization for multimodal document question answering that skips intermediate reasoning tokens and directly aligns visual features with answer locations. The approach reduces per-query inference token length by more than 60% compared to reasoning-enabled models and achieves comparable grounding precision with 65% less training data. Direct perception-based training outperforms reasoning-centric approaches, suggesting that intermediate reasoning steps are unnecessary for this task at the 4B parameter scale.