Sequential Monte Carlo (SMC)-style frameworks are a common approach to inference-time search, but they depend on per-step reward signals and gradients. According to the new arXiv preprint, those dependencies introduce systematic bias into the sampling process.
To address this, the paper proposes a tree-based sampler — specifically a "Bootstrap Stochastic Interpolant Tree" — that avoids these dependencies entirely. The authors frame tree-based samplers as a compelling option for scalable inference-time search, since they do not require the same step-by-step reward or gradient feedback.
One caveat: the abstract provided is truncated after the opening sentences, so this digest can only report the motivation and framing. The full methodology, experimental setup, and results are not included in the available excerpt.