PosteriorBench accepted to NeurIPS: point reconstruction accuracy misleads on inverse solvers

AnimaAnandkumar · x · 2026-09-25

PosteriorBench, from Anima Anandkumar's group, is accepted to NeurIPS 2026 with the thesis that single point reconstruction is not enough — scientific inverse solvers should be evaluated on full posterior recovery. Many scientific problems involve indirect or partial observations where multiple physical fields explain the same measurements, making reconstruction accuracy alone misleading. The team spent substantial compute building high-fidelity reference posteriors across four tasks (Darcy flow inversion, Poisson source recovery, carbon capture and storage, light-transport material inference), enabling evaluation with five complementary metrics. Key finding: better reconstruction doesn't mean better posterior matching. Code is open-sourced as neuraloperator/PosteriorBench.

Related event: PosteriorBench Accepted at NeurIPS, Redefining Evaluation for Inverse Problems(2 posts)→

Original post →

More from Research

Research channel →