The score says the simulation worked. The physics disagrees

Confident and wrong

A generative sampler's free energy estimate comes with a confidence score: effective sample size, or ESS. In a preprint posted this week, researchers show that ESS diagnostics are structurally invariant to the Boltzmann mass of unvisited basins in generative samplers — in plain terms, the score cannot see the parts of the physics the sampler never visited, so it stays high while the answer is wrong. m:…5651

The paper makes the failure concrete with Trp-cage, a small protein used as a test system. Omitting the Trp-cage native basin, which accounts for 50.3% of the mass, results in a free energy shift of 0.417 kcal/mol — an error the standard diagnostics never flag, because the missing basin is invisible to them by construction. The sampler never knew the basin was there, and nothing in its own output could tell it. m:…5651

Close is not correct

A preprint on machine learning for transition-state geometry finds the same pattern in a different metric. RMSD and D-MAE — measures of how close a predicted geometry is to a reference — are unreliable proxies for predicting whether a predicted geometry optimizes to a genuine transition state. Geometry-based quality measures saturate in a narrow AUC band of 0.77-0.84 for predicting optimization success, meaning they top out well short of certainty no matter how good the geometry looks. m:…5744

The numbers are stark: 18% of geometry guesses within 0.28 Å of a reference transition state fail to optimize to a genuine transition state. A guess can be nearly on top of the right answer and still be chemically wrong — and the score that measures closeness cannot tell the difference. The metric rewards geometric closeness, not chemical validity, and closeness is the wrong target when the question is whether the geometry works. m:…5744

The same blindness, out of distribution

Another preprint looks at HIP, a machine-learned interatomic potential, away from the data it was trained on. HIP_v3.ckpt exhibits a mean Eckart-projected eigenvalue error of 1.8234 eV/(Ų·amu) on PubChem geometries, exceeding the in-distribution HORM baseline by over 300x. The authors trace part of the failure to a practical detail, not a mystery of deep learning: the 32-neighbor cap in torch_cluster.radius_graph without distance sorting causes significant edge loss in high-degree molecules. m:…9747

The fix is also practical. Setting cutoff_hessian to 7.5 Å and using distance-based neighbor selection reduces the 90th-percentile force MAE from 0.1410 to 0.0074 eV/Å. The improvement came from construction, not from a new model: the graph was throwing away the connections it needed. m:…9747

Check against something independent

The papers converge on the same remedy: check against something the model did not produce. For generative samplers, reverse checks using independent reference samplers are more effective than forward Jensen gaps at flagging missing basins. The check works precisely because the reference is independent: it can see what the sampler never visited. m:…5651

None of the papers argues that these tools are useless — only that their headline scores are not evidence of correctness. A sampler that reports high confidence, a geometry that sits 0.28 Å from the reference, a potential that performs well in distribution: each can be wrong in ways its own score cannot detect. Until the check comes from outside the model, the number to trust least is the score the model gives itself. m:…5651 m:…5744 m:…9747