Perception-RFT Improves Multimodal QA by Bypassing Reasoning

Achieving efficient multimodal document question answering (QA) with precise visual grounding, where the exact document region supporting an answer is identified, remains a challenge. Current methods either rely on Supervised Fine-Tuning (SFT), which requires extensive annotated data, or reasoning-centric Reinforcement Learning (RL), which inflates inference costs with verbose intermediate traces without clear performance benefits. Researchers introduce Perception-RFT, a novel training framework that applies Group Relative Policy Optimization (GRPO) to multimodal document QA. This approach directly aligns visual features with structured grounding outputs, effectively bypassing the need for intermediate reasoning tokens. A rigorous evaluation, including a reasoning variant under identical reward settings, revealed that reasoning-enabled models often suppress their reasoning traces during training, converging to direct perception-based policies at the 4B parameter scale. This results in a more than 60% reduction in per-query inference token length, with reasoning-enabled RL even underperforming perception-only training. The study also identified a "Grounding Divergence" where a trade-off exists between semantic robustness and geometric precision under joint RL optimization, further supporting the efficiency of the perception-only approach.