Research map / Probabilistic and scientific modeling / Variational and generative sampling
Geometry-Aware Score-Repellent Monte Carlo
Jie Hu, Lingyun Chen, Do-Young Eun
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a geometry-aware extension of Score-Repellent Monte Carlo (SRMC) that replaces the single scalar repellence strength with a direction-dependent one to handle steep and flat directions of the target.
Problem
Score-Repellent Monte Carlo tilts the target density with a single scalar repellence strength, which cannot simultaneously handle steep and flat directions of the target distribution in MCMC sampling.
Approach
Extends SRMC's score-based repellence factor exp(-alpha theta_n^T s(x)) to be geometry-aware so that repellence strength can adapt across steep and flat directions instead of using one scalar alpha.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Euclidean alignment (SRMC) with Fisher-weighted alignment (GA-SRMC) on MALA | MALA + SRMC · Gauss-2D kappa=100 | worst-coordinate ESS: 104 -> 452 | Improved |
| replace SRMC with GA-SRMC on MALA | MALA + SRMC · MNIST BLR | worst-coordinate ESS: 74 -> 501 | Improved |
| replace SRMC with GA-SRMC on MALA | MALA + SRMC · Australian, Heart, Pima BLR | worst-coordinate ESS: 16-32% higher ESS | Improved |
| replace SRMC with GA-SRMC on FisherMALA (proposal already preconditioned) | FisherMALA + SRMC · Gauss-2D | worst-coordinate ESS: 12459 -> 18425 (ratio 1.48) | Improved |
| replace SRMC with GA-SRMC on FisherMALA | FisherMALA + SRMC · Caravan BLR | worst-coordinate ESS: 6381 -> 6022 | Hurt |
| replace SRMC with GA-SRMC on MALA | MALA + SRMC · German BLR | worst-coordinate ESS: 4663 -> 4173 | Hurt |
| replace SRMC with GA-SRMC | MALA/FisherMALA + SRMC · all 8 targets | posterior-mean MSE: lower in all 16 full-Fisher comparisons | Improved |
| replace SRMC with GA-SRMC on near-isotropic target | MALA + SRMC · Ripley (kappa=1.3) | posterior-mean MSE: 0.205 -> 0.198 | No clear effect |
| add GA-SRMC to MALA versus swapping MALA for FisherMALA | MALA · all 8 targets | posterior-mean MSE: 2.45x on MNIST to 32.9x on Ripley lower than FisherMALA | Improved |
| use smaller Fisher regularizer delta (0.01 instead of 10) (ablation) | FisherMALA + GA-SRMC · 7 targets, delta sweep | worst-coordinate ESS: Gauss-2D 10 -> 12695; Ripley 12027 -> 13958 | Improved |
| use diagonal instead of full Fisher estimate in GA-SRMC under MALA (ablation) | MALA + GA-SRMC · coupled BLR posteriors | worst-coordinate ESS | Hurt |
| use diagonal Fisher estimate in history feedback only (MALA + diagonal GA-SRMC) (ablation) | MALA + SRMC · all 8 targets | posterior-mean MSE | Improved |
Experimental setup
Baselines: Score-Repellent Monte Carlo
Builds on: Score-Repellent Monte Carlo
Closest papers by meaning
- Score-Repellent Monte Carlo: Toward Efficient Non-Markovian Sampler with Constant Memory in General State Spaces (ICML 2026 · similarity 0.89)
- Accelerating Constrained Sampling: A Large Deviations Approach (JMLR 2026 · similarity 0.86)
- Stochastic Quasi-Newton Langevin Monte Carlo (ICML 2016 · similarity 0.86)
- Reflective Hamiltonian Monte Carlo: Mixing Analysis and Application to Sampling on Stiefel Manifold (ICML 2026 · similarity 0.86)
- Sampling by averaging: A multiscale approach to score estimation (NeurIPS 2025 · similarity 0.86)
- AutoStep: Locally adaptive involutive MCMC (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
