Rollback and Repair: Reusing Trajectory Assets for Test-Time Scaling in Protein Design

Reward-guided test-time scaling for protein design aims to improve task-specific quality. However, existing derivative-free approaches typically treat reward feedback as a one-shot selection criterion: they retain only high-scoring candidates and discard low-scoring denoising trajectories. This strategy concentrates computational resources around a few high-reward modes and prematurely eliminates potentially recoverable intermediate states—thereby inducing diversity collapse during test-time optimization. To overcome these issues, we propose **T**rajectory **R**epair with **A**sset **C**onstruction and **E**valuation Reuse (**TRACER**), a training-free test-time scaling framework that repurposes evaluated denoising trajectories as reusable assets, enabling rollback and targeted repair of suboptimal paths. Specifically, evaluated intermediate states—along with their associated denoising timesteps and value estimates—are organized into time-indexed trajectory assets. Leveraging these assets, TRACER selects underperforming candidates via group-relative advantages, localizes their optimal rollback times with temporal residuals, retrieves high-value references from the time-indexed assets, and repairs the rolled-back intermediate states through reference-guided re-denoising. Experiments on five protein generative models across de novo protein design, protein-binding protein design, and motif scaffolding show that TRACER consistently improves unique success rate while maintaining generative quality.