PosteriorBench: From Point Estimates to Posterior Matching in Evaluating Generative Inverse Solvers
PosteriorBench tests whether inverse solvers recover the full answer distribution, not just one plausible reconstruction.
The benchmark covers four physics-based inverse problems and compares solvers against reference posteriors built with rejection sampling and MCMC. Its metrics check mean error, uncertainty, distributional alignment, and frequency fidelity. The authors report sizable posterior-matching gaps in current solvers, even when point estimates look strong. They also find neural operators help with resolution robustness, while guidance weights and generation noise matter for variance calibration. ArXiv · AI/CL/LG's note
The benchmark covers four physics-based inverse problems and compares solvers against reference posteriors built with rejection sampling and MCMC. Its metrics check mean error, uncertainty, distributional alignment, and frequency fidelity. The authors report sizable posterior-matching gaps in current solvers, even when point estimates look strong. They also find neural operators help with resolution robustness, while guidance weights and generation noise matter for variance calibration. ArXiv · AI/CL/LG's note
score 5