ReRoute: causal what-if predictions for scientific emulators without controlled experiments

FrancescoLocat8 · x · 2026-10-16

Researchers including Dingling Yao, Anima Anandkumar and Francesco Locatello introduce ReRoute, a framework for counterfactual "what-if" prediction in scientific emulators without needing controlled intervention data.

Problem: models accurate on observed data fail when correlated inputs vary independently; the usual fix (controlled simulation data) is expensive and inherits simulator assumptions.
Method: fix the queried backbone input to a reference value, reintroduce its variation via a known mechanistic pathway, and fine-tune on factual data, leaving downstream effects to learned dynamics; the core causal identification argument is machine-checked in Lean.
Results: highly accurate counterfactuals in an advection-diffusion system with exact solutions, plus state-of-the-art climate emulation on held-out coupled-climate interventions.

Positioned as a step toward scientific AI that explores what-ifs beyond reproducing observations.

Original post →

ReRoute: causal what-if predictions for scientific emulators without controlled experiments

More from AGI Musings