Position Paper: Generative Molecular Dynamics Evaluations Let Models Claim Arbitrary Wins

rishabh16_ · x · 2026-10-28

A NeurIPS 2026 position paper by Simon Olsson, Frank Noé, Grant Rotskoff, and Kresten Lindorff-Larsen argues the generative molecular dynamics (GenMD) community evaluates progress improperly: models can match reported metrics while being arbitrarily wrong on the claims made, risking over-claimed advantages over classical baselines.

Key points:

Ground truth — exhaustive molecular dynamics on systems of interest — is computationally intractable, so evaluation drifts toward convenient proxies and those norms entrench as a community-level equilibrium.
The paper identifies misaligned evaluation patterns, e.g. papers touting "1000× faster than MD" inference speedups while hiding data-generation and training costs; for per-system methods a 1000× speedup on a handful of queries can leave the method net more expensive than the baseline.
It proposes a tiered evaluation checklist: Tier A (always required, nearly free) covers target distribution, system scope, ID vs OOD claims, explicit cost accounting (Cdata, Ctrain, Cinf, break-even query count), and data/protocol release; Tier B applies when specific claims are made (reference fitness, thermodynamic vs kinetic metrics, ESS after reweighting); Tier C encourages strong classical baselines at matched compute and claim-evidence alignment.

The goal is to give the field shared evaluation language and tools before current conventions calcify.

Original post →

Position Paper: Generative Molecular Dynamics Evaluations Let Models Claim Arbitrary Wins

More from Research