Tag
提出 VISTA 这一免训练方法，通过在少量全引导 warmup 生成中收集价值统计并用动态规划优化，在固定预算下自动为离散扩散的 SMC 奖励引导选择引导时间步，在文本、生物序列和图像生成任务上优于手工稀疏调度，实现最高 5.35× 加速。
This paper proposes a test-time alignment approach for large vision-language models using trajectory-guided structured sampling and iterative MCMC refinement, improving visual reasoning accuracy without heavy post-training.
This paper analyzes test-time alignment methods that use a small aligned model as a proxy to guide a larger unaligned model's generation. The authors propose a novel rejection criterion based on conservative confidence betting and demonstrate improvements over existing approaches on multiple datasets.