• 강사 : 안성진 교수 (KAIST 전산학부)
• 일시 : 2025. 10. 13. (월) 16:00~17:30

Diffusion models have emerged as a powerful framework for long-horizon planning, but unlike Monte Carlo Tree Search (MCTS), their performance does not naturally improve with additional test-time computation. This talk introduces Monte Carlo Tree Diffusion (MCTD), a framework that integrates the generative strengths of diffusion models with the structured reasoning and adaptive search of MCTS to realize System 2–style planning. By redefining denoising as a tree-structured rollout and introducing meta-actions to balance exploration and exploitation, MCTD enables scalable refinement of trajectories as computation increases. Experimental results demonstrate substantial gains over existing diffusion-based planners, highlighting MCTD as a step toward diffusion-driven reasoning and AGI-level deliberative planning.