DiMOS: Doob-Guided Inference-Time Multi-Objective Search for Scientific Design

2026-10-05

AI Summary

This study addresses the challenge of optimization within limited inference budgets in scientific design, where stringent multi-objective constraints, gradient-free black-box rewards, and narrow feasible regions pose significant difficulties. To overcome these issues, this work proposes a training-free framework built upon a frozen masked diffusion model. The method introduces a novel local resampling mechanism guided by an approximate Doob h-transform that operates without reward gradients, coupled with a budget-aware trajectory search algorithm for efficient computational resource allocation. Experimental results demonstrate that the proposed approach achieves state-of-the-art joint success rates across six DNA, protein, and RNA design tasks, outperforming the strongest baseline by a factor of 1.98 while effectively preserving sequence uniqueness and naturalness.

Abstract

Scientific design often requires jointly satisfying multiple objectives and constraints. Pretrained masked diffusion models provide a generative foundation for this task, but fine-tuning them to meet these objectives and constraints incurs additional training costs, motivating inference-time guidance with frozen models. However, such guidance faces two challenges: pass-or-fail constraints and black-box reward models may provide no useful gradients, while jointly satisfying multiple requirements can leave a small feasible region, making feasible designs difficult to find within a limited inference budget. To address these challenges, we introduce DiMOS, a training-free framework for multi-objective scientific design. Using joint rewards from candidate completions, DiMOS performs approximate Doob-guided local resampling without requiring reward gradients. To allocate computation efficiently, it uses budget-efficient trajectory search to focus computation on promising continuations. Across six DNA, protein, and RNA tasks, DiMOS attains the highest joint success rate at comparable generation times, up to $1.98\times$ the strongest baseline on DNA and protein, while maintaining high sequence uniqueness and naturalness.

Problem

• multi-objective scientific design
• inference-time guidance
• masked diffusion models
• black-box reward
• feasible region

Innovation

• Multi-Objective Search
• Masked Diffusion Models
• Inference-Time Guidance
• Doob-Guided Resampling
• Budget-Efficient Trajectory Search