AI Insight
This study introduces Perception-RFT, a training framework for multimodal document question answering that uses reinforcement learning to directly align visual features with answer locations, bypassing intermediate reasoning steps. Testing on a 4 billion parameter model, researchers found that models trained to reason actually suppressed their reasoning traces during optimization, converging to direct perception-based responses while reducing inference token length by over 60% compared to reasoning-enabled approaches. The work reveals that reasoning chains provide no performance benefit at this scale and identifies a trade-off between semantic accuracy and geometric precision when models are optimized for both simultaneously.
Why it matters
This research could significantly reduce computational costs for document analysis AI systems by eliminating unnecessary reasoning steps, making such tools more efficient and accessible. The findings challenge assumptions about when chain-of-thought reasoning benefits multimodal AI, suggesting simpler direct perception approaches may be preferable for certain document understanding tasks.
Understand the Science
arXiv:2607.14682v1 Announce Type: cross
Abstract: Efficient multimodal document question answering with explicit visual grounding, locating the precise document region that supports each answer remains an open challenge. Current approaches bifurcate into Supervised Fine-Tuning (SFT), which requires large annotated datasets and reaches optimization plateaus, and reasoning-centric Reinforcement Learning (RL), which depends on verbose intermediate traces that inflate inference token cost without clear benefit. We introduce Perception-RFT, a training framework that applies Group Relative Policy Optimization (GRPO) to multimodal document QA, bypassing intermediate reasoning tokens to directly align visual features with structured grounding outputs. To rigorously evaluate the necessity of reasoning, we construct a reasoning variant under identical reward settings. We find that reasoning-enabled models suppress their reasoning traces during training, converging to direct perception-based policies at the 4B parameter scale, reducing per-query inference token length by more than 60%, while reasoning-enabled RL underperforms perception-only training. Through a fine-grained analysis of Qwen3-VL-4B optimization dynamics, we confirm that SFT saturation and cold-start RL instability established in text-domain post-training extend to multimodal, and identify a previously uncharacterized Grounding Divergence: a selective trade-off between semantic robustness and geometric precision on two out of distribution (OOD) benchmarks (4,828 samples) under joint RL optimization. We further show that an early SFT$rightarrow$RL transition achieves comparable precision with 65% less training data.