arXiv Science⌕ Search

arXiv · 2611.19501

Latent Molecular Rollout Towards Causal World Models: Direction is Learnable, Amplitude Control is Readout-Dependent

Abstract

A molecular world model should predict how structural changes alter physical properties, not merely assign labels to static molecules. We build a latent molecular rollout system on an 8B Qwen3 backbone, trained on 9,202 aqueous experimental pKa pairs from iBonD, and assess its progress toward causal control by auditing three often-conflated capabilities: predictive fidelity through the intended molecular pathway, directional learnability versus amplitude calibration of target-conditioned displacements, and transfer of latent control to decoded structures. After removing textual shortcuts, continuous latent inference achieves a test MAE of 2.087, significantly outperforming direct molecular feature readout (MAE difference -0.276, 95% CI [-0.352, -0.203], p<0.001). A target-conditioned policy reliably learns displacement directions, with a median cosine of 0.577 and 62.8% of 43 held-out pairs exceeding the 0.3 threshold. Amplitude conclusions are strictly readout-dependent. A frozen linear readout yields near-zero amplitude correlation (-0.020), while a nonlinear gradient-based readout recovers relative amplitude ordering (correlation 0.585, sign accuracy 0.744-0.767) with compressed absolute scale. At the decoded structure level, neither input-slot editing nor direct final-state injection produces meaningful changes across 80 test sources, bounding the tested control surfaces rather than indicating a universal decoder bottleneck. Generation fine-tuning shows relative improvement under internal oracle evaluation but does not surpass a simple constant baseline in single-sample settings. These results delineate the capability frontier of fixed-scale latent molecular interfaces: directional control is robustly achievable, but calibrated amplitude control and decoded structural transfer remain open challenges on the path to causal molecular world models.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Rui Wang. 2026-11-14. Latent Molecular Rollout Towards Causal World Models: Direction is Learnable, Amplitude Control is Readout-Dependent. https://arxiv.science/abs/2611.19501

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Nuclear Spin Isomers and the Pauli Principle in Polaritonic Chemistry

The Pauli principle has far-reaching consequences in quantum physics. Here, we investigate, for the first time, its implications, together with nuclear spin isomerism, in polaritonic chemistry. The theory is developed for a single and a few molecules as well as for an ensemble of molecules. As an explicit and detailed example we first present an accurate numerical description of a realistic situation involving two $^{14}$NH$_3$ molecules, existing as ortho and para spin isomers, in an infrared plasmonic cavity. Then, we generalize the approach for molecular ensembles using analytical considerations based on the Tavis-Cummings model and simulate the transmission spectrum of a Fabry-Pérot cavity filled with $^{14}$NH$_3$ gas using quantum mechanics. These results are directly relevant for recent gas-phase experiments studying rovibrational polaritons in molecules. Our findings undoubtedly demonstrate that the Pauli principle and nuclear spin isomerism significantly reshape collective light-matter coupling involving molecules with identical nuclei.

physics.chem-ph↗

Efficient strategies for obtaining accurate kinetics from biased simulations

It has been demonstrated that one can recover the unbiased rate constants for slow conformational transitions by observing transitions driven by a bias potential added along one or several collective variables (CVs). However, the choice of CV, biasing scheme, biasing parameters and the rate estimator can all affect the results and it is not obvious which to choose for a given problem. In this study, we compare accuracy versus computational cost for a previously established benchmark system, a coarse-grained protein with computationally accessible unfolding kinetics. We apply bias to four CVs of decreasing quality in both a time-dependent and quasi-static manner, and compare different estimators for the unfolding rate constant in terms of absolute accuracy to the estimated reference unbiased rate. We show that in the case of gentle biasing, the exponential average time-dependent rate (EATR) framework, which simultaneously fits the unbiased rate constant and a coordinate quality metric $γ$, recovers accurate kinetics in nearly all cases. For poor CVs, estimators equivalent to assuming a good coordinate ($γ=1$) systematically fail regardless of biasing method or parameters. We introduce heuristics to show when we expect the rates to be inaccurate, and identify which combinations of biasing scheme, estimator and number of runs reach a given accuracy at the lowest cost.

physics.chem-ph↗

Neural Network Path Optimization for Finding Transition States on a Machine Learned Potential

Transition states govern chemical reaction rates and outcomes, yet locating them remains a central challenge in computational chemistry. Double-ended search methods find transition states given reactant and product structures, but generally rely on discretized path representations developed for density functional theory, where potential energy surface evaluations are expensive. Machine learned interatomic potentials (MLIPs) dramatically reduce this cost and leverage automatic differentiation to access higher-order derivatives efficiently. We introduce Popcornn, a framework to find transition states by optimizing continuous neural network paths atop an MLIP with a parallelized adaptive numerical integrator. This combination eliminates path discretization artifacts, reduces hyperparameter sensitivity, and enables non-local optimization to overcome local features using second-order potential-energy-surface information. Popcornn consistently recovers more correct transition states than standard methods such as nudged elastic band and growing string, across MLIPs on thousands of organic reactions, establishing it as a robust, automation-ready method for high-throughput transition state search. Popcornn's infrastructure is flexible in path objective, integration strategy, and optimization algorithm, making it a practical tool and platform for investigating open questions in MLIP-based path optimization.

physics.chem-ph↗