Navigating the Exploration-Exploitation Tradeoff in Inference-Time Scaling of Diffusion Models

Inference-time scaling has achieved remarkable success in language models, yet its adaptation to diffusion models remains underexplored. We observe that the efficacy of recent Sequential Monte Carlo (SMC)-based methods largely stems from globally fitting the The reward-tilted distribution, which inherently preserves diversity during multi-modal search. However, current applications of SMC to diffusion models face a fundamental dilemma: early-stage noise samples offer high potential for improvement but are difficult to evaluate accurately, whereas late-stage samples can be reliably assessed but are largely irreversible. To address this exploration-exploitation trade-off, we approach the problem from the perspective of the search algorithm and propose two strategies: Funnel Schedule and Adaptive Temperature. These simple yet effective methods are tailored to the unique generation dynamics and phase-transition behavior of diffusion models. By progressively reducing the number of maintained particles and down-weighting the influence of early-stage rewards, our methods significantly enhance sample quality without increasing the total number of Noise Function Evaluations. Experimental results on multiple benchmarks and state-of-the-art text-to-image diffusion models demonstrate that our approach outperforms previous baselines.

Anchor Space Optimal Transport: Accelerating Batch Processing of Multiple OT Problems

The optimal transport (OT) theory provides an effective way to compare probability distributions on a defined metric space, but it suffers from cubic computational complexity. Although the Sinkhorn's algorithm greatly reduces the computational complexity of OT solutions, the solutions of multiple OT problems are still time-consuming and memory-comsuming in practice. However, many works on the computational acceleration of OT are usually based on the premise of a single OT problem, ignoring the potential common characteristics of the distributions in a mini-batch. Therefore, we propose a translated OT problem designated as the anchor space optimal transport (ASOT) problem, which is specially designed for batch processing of multiple OT problem solutions. For the proposed ASOT problem, the distributions will be mapped into a shared anchor point space, which learns the potential common characteristics and thus help accelerate OT batch processing. Based on the proposed ASOT, the Wasserstein distance error to the original OT problem is proven to be bounded by ground cost errors. Building upon this, we propose three methods to learn an anchor space minimizing the distance error, each of which has its application background. Numerical experiments on real-world datasets show that our proposed methods can greatly reduce computational time while maintaining reasonable approximation performance.

Safe Screening for Unbalanced Optimal Transport

This paper introduces a framework that utilizes the Safe Screening technique to accelerate the optimization process of the Unbalanced Optimal Transport (UOT) problem by proactively identifying and eliminating zero elements in the sparse solutions. We demonstrate the feasibility of applying Safe Screening to the UOT problem with $\ell_2$-penalty and KL-penalty by conducting an analysis of the solution's bounds and considering the local strong convexity of the dual problem. Considering the specific structural characteristics of the UOT in comparison to general Lasso problems on the index matrix, we specifically propose a novel approximate projection, an elliptical safe region construction, and a two-hyperplane relaxation method. These enhancements significantly improve the screening efficiency for the UOT's without altering the algorithm's complexity.

Wasserstein Graph Distance based on $L_1$-Approximated Tree Edit Distance between Weisfeiler-Lehman Subtrees

The Weisfeiler-Lehman (WL) test has been widely applied to graph kernels, metrics, and neural networks. However, it considers only the graph consistency, resulting in the weak descriptive power of structural information. Thus, it limits the performance improvement of applied methods. In addition, the similarity and distance between graphs defined by the WL test are in coarse measurements. To the best of our knowledge, this paper clarifies these facts for the first time and defines a metric we call the Wasserstein WL subtree (WWLS) distance. We introduce the WL subtree as the structural information in the neighborhood of nodes and assign it to each node. Then we define a new graph embedding space based on $L_1$-approximated tree edit distance ($L_1$-TED): the $L_1$ norm of the difference between node feature vectors on the space is the $L_1$-TED between these nodes. We further propose a fast algorithm for graph embedding. Finally, we use the Wasserstein distance to reflect the $L_1$-TED to the graph level. The WWLS can capture small changes in structure that are difficult with traditional metrics. We demonstrate its performance in several graph classification and metric validation experiments.