conditional sampling with sir

Slice sampling suffers from low efficiency and heavy reliance on manual tuning when applied to complex target distributions—such as highly skewed or constrained spaces. To address this, we propose Quantile Slice Sampling (QSS), a novel framework that (1) integrates probability integral transformation with quantile mapping to enable automatic initialization and unit-interval standardization; (2) introduces an evaluable pseudo-target importance reweighting mechanism, coupled with dual-metric quality assessment and adaptive parameter optimization; and (3) extends slice sampling to multivariate and constrained state spaces by incorporating elliptical slicing, Neal’s shrinkage, and Gibbs-like coordinate updates. Experiments on benchmark distributions and Bayesian modeling tasks demonstrate that QSS significantly outperforms conventional slice sampling and Metropolis–Hastings: in highly skewed and constrained settings, it reduces rejection rates by over 30%, while delivering enhanced robustness, full automation, and practical usability.

For expensive black-box target density sampling, this paper proposes an active sampling method framed within the multi-armed bandit (MAB) paradigm. Unlike conventional approaches that optimize a proposal distribution, our method models sample location selection as a sequential decision-making process, jointly leveraging a Gaussian process surrogate model and space-filling criteria to adaptively select evaluation points with maximal information gain. To our knowledge, this is the first work to directly apply MAB for sampling point scheduling—bypassing explicit distribution modeling and substantially reducing the number of target function evaluations. Experiments demonstrate superior performance over state-of-the-art importance sampling methods on multimodal and heavy-tailed distributions. In Bayesian inference tasks, our approach achieves higher approximation accuracy with significantly fewer evaluations.

This work addresses the inefficiency of retraining models for each importance-weighting function in multi-task biased sampling. We propose a **training-free importance sampling framework**, leveraging a pre-trained score-based generative model (SGM). By combining its score function with an arbitrary importance weight function, we formulate importance sampling as a **controllable reverse diffusion process**, requiring neither fine-tuning nor retraining. Our key contribution is the first fully training-free importance sampling method, featuring an explicitly derived weight-adaptive reverse stochastic differential equation (SDE) grounded in the score function. Extensive experiments on industrial and natural image datasets demonstrate high scalability and effectiveness: the approach significantly reduces computational and training overhead in multi-task settings while enabling flexible adaptation of a single base distribution to diverse bias objectives.

This paper addresses efficient conditional sampling under partially observed conditional distributions. We propose the CGMMD framework, which—uniquely—integrates Maximum Mean Discrepancy (MMD) with nearest-neighbor regression to formulate a non-adversarial, directly optimizable training objective, enabling theoretically guaranteed one-shot conditional sampling via a single forward pass. Theoretical contributions include deriving a unified generalization bound for nearest-neighbor functions and proving that the generated distribution converges in probability to the true conditional distribution. Empirically, CGMMD achieves state-of-the-art performance on synthetic data and real-world inverse problems—including image denoising and super-resolution—while requiring only O(1) forward computations at test time, drastically reducing inference complexity. The method thus bridges rigorous theoretical foundations with practical deployability.

We propose a training-free conditional sampling method for flow matching models based on importance sampling. Because a na\"ive application of importance sampling suffers from weight degeneracy in high-dimensional settings, we modify and incorporate a resampling technique in sequential Monte Carlo (SMC) during intermediate stages of the generation process. To encourage generated samples to diverge along distinct trajectories, we derive a stochastic flow with adjustable noise strength to replace the deterministic flow at the intermediate stage. Our framework requires no additional training, while providing theoretical guarantees of asymptotic accuracy. Experimentally, our method significantly outperforms existing approaches on conditional sampling tasks for MNIST and CIFAR-10. We further demonstrate the applicability of our approach in higher-dimensional, multimodal settings through text-to-image generation experiments on CelebA-HQ.

This work addresses the challenge of implementing resampling steps in conditional sequential Monte Carlo (CSMC) algorithms, particularly when dealing with complex dependency structures. The authors propose a unified framework that accommodates a wide range of standard and nonstandard resampling schemes—including systematic, adaptive, and chopthin resampling—without requiring strong assumptions such as unbiasedness or exchangeability. By preserving the order of ancestor indices and avoiding random permutations, the framework not only simplifies algorithmic implementation but also significantly broadens the theoretical applicability of CSMC. This approach guarantees the validity of CSMC under substantially weaker conditions than previously required, thereby enhancing its robustness and flexibility in practical applications.

Traditional inverse design methods are limited to point-wise target outputs and struggle to accommodate design requirements expressed as target distributions. This work formalizes, for the first time, the distribution-level inverse design problem and introduces a new paradigm termed Conditional Distribution Matching (CDM), defining two task variants: CDMS and CDMO. The authors propose MLGD-F, a plug-and-play inference algorithm that efficiently solves these tasks without additional training. MLGD-F leverages a pre-trained score-based diffusion model combined with a single-step conditional sampler, using a matching loss to guide gradient updates. The method successfully recovers inputs whose outputs align with complex target distributions—including discrete mixtures and continuous low-rank supports—demonstrating effectiveness across synthetic data, structured image transformation, and generative editing tasks.

This work addresses the challenge of evaluating conditional generation quality in compositional extrapolation settings, where the true target distribution is unavailable. The authors propose a post-hoc, instance-wise confidence scoring mechanism that requires no access to the target distribution. By constructing estimable metrics based on data manifold compatibility and attribute contrastive distance, the method holistically assesses both global realism and attribute fidelity. Notably, it incurs no additional training and is directly applicable to off-the-shelf pre-trained generative models. To the best of our knowledge, this is the first approach enabling effective evaluation of compositional extrapolation samples, facilitating sample filtering, ranking, and pre-generation abstention. Experiments on biological imaging and visual benchmarks demonstrate substantial improvements in morphological fidelity and downstream predictive performance, along with the capability for early abstention during generation.

This study addresses the challenge of specifying concentration parameter priors in Dirichlet process mixture models, where default hyperpriors often impose overly strong or unintended assumptions. To resolve this, the authors propose the Design Conditional prior Elicitation (DCE) framework, which translates practitioners’ prior beliefs about clustering structure into a Gamma hyperprior tailored to a fixed sample size, thereby jointly regulating the number of clusters and the distribution of mixture weights. The approach employs a two-stage moment-matching procedure to enhance computational efficiency and introduces a dual-anchor protocol to diagnose and mitigate risks of weight dominance. Experiments demonstrate that DCE-calibrated priors substantially reduce posterior collapse rates—by over 60% compared to the default Gamma(1,1)—and consistently improve clustering accuracy and robustness across varying data informativeness. An open-source R package, DPprior, and reproducible diagnostic workflows are provided.