Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document Question Answering via Reasoning-Free Alignment
Merged summary
TL;DR - Perception-RFT is a post-training framework that applies GRPO reinforcement learning to multimodal document QA while skipping intermediate reasoning tokens, directly aligning visual features with structured grounding outputs. It matters because it shows explicit chain-of-thought reasoning is often unnecessary—and even counterproductive—for visual grounding tasks, cutting inference cost sharply.
- Reasoning-free wins: At the 4B scale (Qwen3-VL-4B), reasoning-enabled models suppress their own traces during training and converge to direct perception-based policies; reasoning-enabled RL underperforms perception-only training while inflating token cost.
- Efficiency gains: Bypassing intermediate reasoning reduces per-query inference token length by more than 60%.
- Post-training dynamics carry over: SFT optimization saturation and cold-start RL instability known from text-domain post-training extend to the multimodal setting.
- New failure mode + data savings: The authors identify a "Grounding Divergence"—a trade-off between semantic robustness and geometric precision on two OOD benchmarks (4,828 samples)—and show an early SFT→RL transition matches precision with 65% less training data.
Sources (1)
Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document Question Answering via Reasoning-Free Alignment
TL;DR - Perception-RFT is a post-training framework that applies GRPO reinforcement learning to multimodal document QA while skipping intermediate reasoning tokens, directly aligning visual features with structured grounding outputs. It matters because it shows explicit chain-of-thought reasoning is often unnecessary—and even counterproductive—for visual grounding tasks, cutting inference cost sharply.
- Reasoning-free wins: At the 4B scale (Qwen3-VL-4B), reasoning-enabled models suppress their own traces during training and converge to direct perception-based policies; reasoning-enabled RL underperforms perception-only training while inflating token cost.
- Efficiency gains: Bypassing intermediate reasoning reduces per-query inference token length by more than 60%.
- Post-training dynamics carry over: SFT optimization saturation and cold-start RL instability known from text-domain post-training extend to the multimodal setting.
- New failure mode + data savings: The authors identify a "Grounding Divergence"—a trade-off between semantic robustness and geometric precision on two OOD benchmarks (4,828 samples)—and show an early SFT→RL transition matches precision with 65% less training data.