PosteriorBench’s authors describe a benchmark for scientific solvers that work backward from indirect or partial observations. It evaluates posterior distributions—the probabilities assigned to possible solutions—rather than reconstruction accuracy alone.

The team says it built reference distributions across four tasks: Darcy flow inversion, Poisson source recovery, carbon capture and storage, and light transport material inference. Researchers can evaluate their solvers using five complementary metrics. The authors report that even strong generative samplers struggle to capture both the mean and variance, often underestimating uncertainty.

A September 25 announcement says PosteriorBench was accepted to NeurIPS 2026. Paper and code links are available.