Diffusion vision-language models improve answers by adapting reasoning length
Routing by Reasoning Need: Trajectory-Aware Decoding Control for Diffusion Vision-Language Models
Artificial Intelligence
Summary
Some AI models answer questions by slowly refining their guesses step by step. But using the same number of steps for all questions can cause problems: simple questions may get overthought, while complex questions may be cut short. The authors studied this problem with a model called LLaDA-V and created a way to watch how the answer changes during thinking. Their system decides when to stop early, keep going normally, or focus on detailed reasoning, based only on signals seen without knowing the true answer. Tests showed this approach gave better, more reliable answers across different types of questions.
What this means in practice
- •For natural language processing teams: Improve question-answering systems by adapting reasoning steps per question complexity during model inference.
- •For chatbot developers: Create chatbots with more reliable and context-appropriate answers by dynamically controlling model reasoning length without extra training.$Commercial implications: Enables building more efficient and accurate conversational AI products that adjust reasoning per question complexity, improving user experience.
Authors
Yixiang Liu, Zhongxing Xu, Zhonghua Wang, Xiaoying Tang
Abstract
Diffusion vision-language models generate answers through iterative refinement, exposing intermediate answer trajectories that can be inspected and controlled at inference time. However, this controllability creates a reasoning-need mismatch, where a universal generation length is applied to questions with different reasoning demands. Visually closed questions may be harmed by continued refinement after a stable answer has formed, whereas reasoning-sensitive questions may be harmed by premature commitment. We formulate this problem as reasoning-budget mismatch and study it in LLaDA-V. Rather than choosing a universal generation length, our training-free controller routes each example to early commitment, baseline preservation, or reasoning-supportive decoding using trajectory signals from answer closure, commitment evidence, and representation revision pressure, without using ground-truth answers. Across answer-focused, mixed-reasoning, and CoT-sensitive benchmarks, routed control improves robustness over fixed long decoding, pure short decoding, and single-rule interventions. The gains are not explained by shorter outputs alone. Answer-closed examples often benefit from commitment, whereas CoT-sensitive examples require preserving or supporting intermediate reasoning. Taken together, these results suggest diffusion VLM decoding should route inference-time control by the state suggested by the observed trajectory instead of relying on a universal decoding length.