Minimum Variance Path Principle for Stable Score-Based Density Ratio Estimation
Researchers have introduced the Minimum Variance Path (MVP) Principle to address a critical paradox in score-based machine learning methods, which are theoretically path-independent but practically path-dependent. The study identifies that practical training objectives diverge from ideal ground-truth objectives due to an overlooked term: the path variance of the score function. To resolve this, the authors derived a closed-form expression for this variance, enabling tractable optimization. By parameterizing the path using a flexible Kumaraswamy Mixture Model, the proposed method learns data-adaptive, low-variance paths without requiring heuristic manual selection. This principled approach optimizes the complete objective, resulting in more accurate and stable density ratio estimators. The method has established new state-of-the-art results on challenging benchmarks and provides a general framework for optimizing score-based interpolation. The research, accepted at the Fourteenth International Conference on Learning Representations (ICLR) 2026, includes open-source code to facilitate further adoption and verification within the artificial intelligence and machine learning communities.
Wire timeline
Minimum Variance Path Principle for Stable Score-Based Density Ratio Estimation
Researchers have introduced the Minimum Variance Path (MVP) Principle to address a critical paradox in score-based machine learning methods, which are theoretically path-independent but practically path-dependent. The study identifies that practical training objectives diverge from ideal ground-truth objectives due to an overlooked term: the path variance of the score function. To resolve this, the authors derived a closed-form expression for this variance, enabling tractable optimization. By parameterizing the path using a flexible Kumaraswamy Mixture Model, the proposed method learns data-adaptive, low-variance paths without requiring heuristic manual selection. This principled approach optimizes the complete objective, resulting in more accurate and stable density ratio estimators. The method has established new state-of-the-art results on challenging benchmarks and provides a general framework for optimizing score-based interpolation. The research, accepted at the Fourteenth International Conference on Learning Representations (ICLR) 2026, includes open-source code to facilitate further adoption and verification within the artificial intelligence and machine learning communities.
cs.AI updates on arXiv.org