Adaptive Inference-Time Scaling via Cyclic Diffusion Search

Yoshua Bengio (Mila/U. Montreal) · Minsu Kim (Mila / KAIST) · Gyubin Lee (Korea Advanced Institute of Science & Technology (KAIST)) · Bao Truong (Korea Advanced Institute of Science & Technology) · Jaesik Yoon (KAIST) · Dongwoo Lee (Korea Advanced Institute of Science & Technology) · Sungjin Ahn (KAIST)
adaptive bi-directional cyclic diffusionadaptive inferenceadaptive thinking timeautomatic exploration-exploitation balancingbi-directional diffusion cyclescomputational efficiencycomputational effortcyclic diffusion searchdenoising schedulesdiffusion modelsexploration depthgenerative capabilitiesinference-time scalingsearch-based inference frameworktermination control

Diffusion models have demonstrated strong generative capabilities across domains ranging from image synthesis to complex reasoning tasks. However, most inference-time scaling methods rely on fixed denoising schedules, limiting their ability to allocate computation based on instance difficulty or task-specific demands adaptively. We introduce the challenge of adaptive inference-time scaling—dynamically adjusting computational effort during inference—and propose Adaptive Bi-directional Cyclic Diffusion (ABCD), a flexible, search-based inference framework. ABCD refines outputs through bi-directional diffusion cycles while adaptively controlling exploration depth and termination. It comprises three components: Cyclic Diffusion Search, Automatic Exploration-Exploitation Balancing, and Adaptive Thinking Time. Experiments show that ABCD improves performance across diverse tasks while maintaining computational efficiency.