Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Two Phase Rapid Simulation Based Inference with Differentiable Simulators
Uros Zivanovic, Rahul Srinivasan, Roberto Trotta, Andre Scaffidi
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.
Introduces Rapid Simulation Based Inference (RSBI), a diffusion-based variational method for likelihood-free Bayesian inference using Schrödinger Bridge sampling, with likelihood from a neural ratio estimator or a differentiable simulator.
Problem
Likelihood-free Bayesian inference (simulation-based inference) struggles with posterior sample efficiency when prior-to-posterior volumes are large and multi-dimensional, and with multi-modal posteriors.
Approach
A diffusion-based variational approach that builds on Schrödinger Bridge diffusion sampling to handle multi-modal posteriors, with the likelihood supplied either by a neural ratio estimator or by a differentiable simulator (two-phase design).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RSBI (surrogate-guided zoom + NRE-ASBS) instead of SNPE | SBI on gw_ringdown · gw_ringdown default prior, 1e5 sims | C2ST: 0.935 -> 0.891 | Improved |
| use RSBI instead of SNPE on wide prior | SBI · gw_ringdown wide prior | C2ST: 0.985 -> 0.903 | Improved |
| use RSBI instead of SNPE/SNVI on Lotka-Volterra | SBI · Lotka-Volterra, 1e5 sims | C2ST: 0.94 -> 0.55 | Improved |
| use RSBI at very large prior volume ratio | SBI · Two Moons log10 Rp ~4.3 | C2ST: ~1 (baselines) -> <=0.7 | Improved |
| increase prior volume on Bernoulli GLM (to 12.2) | RSBI and baselines · Bernoulli GLM | C2ST: 0.66 -> 0.89 (RSBI) | Hurt |
| increase simulation budget from 1e5 to 1e6 | RSBI · Two Moons default prior | C2ST: 0.543 -> 0.536 | No clear effect |
| seed SNPE/SNVI with Phase-1 proposal (ablation) | SNPE/SNVI · sbibm tasks | C2ST | Improved |
| add Phase-2 NRE-ASBS transport on top of Phase-1 proposal (ablation) | RSBI · sbibm tasks | C2ST | Improved |
| use RSBI on SLCP at larger prior | all methods · SLCP | C2ST: baselines degrade 0.1-0.2 | Mixed |
Experimental setup
Builds on: Schrödinger Bridge diffusion sampling
Closest papers by meaning
- Closing the Approximation Gap in Simulation-free Latent SDEs (NeurIPS 2026 · similarity 0.87)
- Inflationary Flows: Calibrated Bayesian Inference with Diffusion-Based Models (NeurIPS 2024 · similarity 0.87)
- Dynamic Diffusion Schrödinger Bridge in Astrophysical Observational Inversions (NeurIPS 2025 · similarity 0.86)
- Learning Likelihood-Free Reference Priors (ICML 2025 · similarity 0.85)
- BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling (NeurIPS 2026 · similarity 0.85)
- Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models (NeurIPS 2026 · 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.
