TL;DR: A UC Berkeley and LBNL team trained an unmodified Transformer on Cartesian atomic coordinates to predict molecular energies and forces, and the model learned electrostatic-like interactions and physics-consistent cutoffs with predictable scaling.

Answer first: what the September 24 revision reports

A research team from UC Berkeley and Lawrence Berkeley National Laboratory reports that a Transformer trained without domain-specific priors can recover physical structure directly from data when trained to predict molecular energies and forces. According to the arXiv record for [2510.02259] Transformers Discover Molecular Structure Without Graph Priors, the revised version labeled v3 was posted on September 24, 2026, after an initial submission on October 2, 2025 and a second version on September 17, 2026.

The authors listed on the arXiv page are Tobias Kreiman, Yutong Bai, Fadi Atieh, Elizabeth Weaver, Eric Qu, and Aditi S. Krishnapriyan, with affiliations identified as UC Berkeley and LBNL. The paper is categorized under Machine Learning, Materials Science, Chemical Physics, and Biomolecules. Its central question is framed in the abstract as the extent to which physical inductive biases — described in the abstract as prior assumptions about the structure of the physical world — can emerge by learning from data alone.

The abstract states the main results in three parts. First, the model autonomously recovers key physical structure, including learned interatomic interaction strengths that mirror classical electrostatics and interaction cutoffs consistent with traditional physical models. Second, the work demonstrates predictable neural scaling law behavior with increased data and compute. Third, it finds accuracy on certain metrics competitive with physics-informed architectures.

The proposed interpretation, as stated in the abstract, is that the results clarify the boundary between engineered inductive biases and learnable physical structure. The authors suggest that general-purpose architectures can serve as principled baselines for scientific modeling by learning fundamental physical structure directly from data.

What was built: coordinates in, energies and forces out

The modeling choice is deliberately minimal. In contrast to approaches that prescribe locality through an interaction graph and encode geometric features directly into model architectures, the team represents a molecular system simply as a set of atomic coordinates in Cartesian space. That description comes directly from the paper text indexed before the cutoff.

The study does not construct a graph and does not embed chemical priors. The abstract emphasizes that the training removes domain-specific priors, including any manually defined atomistic pairwise interactions. The test is whether an unmodified Transformer can recover fundamental physical structure when trained to predict molecular energies and forces.

That framing matters because atomistic modeling is described in the paper as a representative field in the computational sciences where machine-learning architectures have historically embedded strong physical inductive biases, such as geometric locality and graph structure. The paper develops what it calls a systematic understanding of how physical patterns can alternatively be discovered directly from data.

In practical terms, the input is atomic positions. The supervision is energy and force labels. The internal behavior is then probed for physical structure, rather than imposing that structure in advance through architecture.

Why atomistic modeling is the test case

Computational simulations are described in the paper as playing a central role in scientific discovery. Machine learning has emerged, in the authors' description, as a promising alternative to traditional physics-based modeling. The complication is that scientific modeling requires physically meaningful predictions.

The paper contrasts two traditions. One tradition builds ab initio methods on known physical laws to describe matter accurately, but notes that high computational cost often makes those methods prohibitively expensive for systems of practical interest. Machine-learning methods, by contrast, are presented as leveraging data to model physical systems at a fraction of the cost of traditional approaches.

Because physical meaning matters, many machine-learning approaches for the physical sciences incorporate physical priors, such as specific symmetries, directly into domain-specific model architectures. The broader machine-learning landscape, the paper notes, has shifted toward a domain-agnostic paradigm centered on the Transformer architecture.

The tension the authors identify is direct: to what extent must physical laws be explicitly encoded into machine-learning models, and to what extent can they be learned directly from data? Atomistic modeling provides a concrete setting for that question because it contains rich physical structure and a history of architecture-level physical assumptions.

What the model learned without being told

The most closely watched claim concerns emergence. Probing the model's internal representations, the authors report that the Transformer autonomously recovers functional forms consistent with classical physics, such as a learned interatomic interaction strength that mirrors electrostatic interactions.

The abstract repeats the same finding in compact form: learned interatomic interaction strengths that mirror classical electrostatics and interaction cutoffs consistent with traditional physical models. The paper text adds that Transformers also learn to flexibly adapt their effective receptive field, focusing on long-range interactions in the settings examined.

The significance, as presented by the authors, is not that electrostatics was programmed into the network. It was not. The model was trained without manually defined atomistic pairwise interactions, without constructing a graph, and without embedding chemical priors. The electrostatic-like dependence and the cutoff-like behavior are described as discovered patterns, not imposed rules.

The paper positions this as evidence for a boundary worth mapping. Some structure that has historically been engineered into models as inductive bias appears, in this experimental setup, to be learnable from data. The authors do not state that all physical priors are unnecessary; they state that the results clarify the boundary between engineered inductive biases and learnable physical structure.

Scaling: predictable gains with data and compute

A second pillar of the paper is scaling behavior. The abstract reports predictable neural scaling law behavior with increased data and compute.

That language connects to the broader Transformer literature discussed in the paper. The success of the Transformer, the authors write, has been guided by empirical scaling laws that allow for precise forecasting of performance gains as a function of dataset size, computational budget, and number of model parameters.

The paper notes that this common architectural framework, paired with established scaling laws, has accelerated research by providing a standardized recipe for addressing a broad class of machine-learning problems. It links that recipe to powerful multi-modal models and to specialized hardware and software for Transformers.

For scientific modeling, predictable scaling would be operationally useful if confirmed across settings. It would mean that additional data and compute translate into expected accuracy changes, rather than unpredictable jumps. The paper reports such predictable behavior in its atomistic experiments, alongside accuracy on certain metrics competitive with physics-informed architectures.

The qualifier matters. The claim is accuracy on certain metrics, not superiority on all benchmarks. The suggested role is also carefully worded: general-purpose architectures can serve as principled baselines for scientific modeling.

Why Open Molecules 2025 makes this test possible now

The authors point to data scale as the enabling change. Despite Transformer success in language, vision, and robotics without explicit domain-specific inductive biases, adoption in the physical sciences has been limited. The paper attributes that limitation partly to historically smaller availability of data and to the resulting perspective that models without physical inductive biases cannot recover complex physical interactions without explicit geometric guidance.

The recent emergence of massive scientific datasets, in the authors' account, provides an opportunity to test that assumption. The dataset named in the paper is Open Molecules 2025, abbreviated as OMol25.

Independent work published in the same week illustrates why OMol25 is drawing attention. According to the PNNL journal record for Prediction and Experimental Verification of Electrolyte Solvation Structure from an OMol25-Trained Interatomic Potential, published September 22, 2026, OMol25 provides large-scale molecular density functional theory training data for machine-learning interatomic potentials with broad elemental coverage and specifically samples tens of millions of electrolyte configurations.

That PNNL-linked study describes machine-learning interatomic potentials trained on large, chemically diverse datasets as enabling molecular dynamics simulations of battery electrolytes with near-density functional theory accuracy over 10,000 times faster than density functional theory. It reports that an OMol25-trained Universal Model of Atoms predicted experimentally measured densities and X-ray structure factors in substantially better agreement than state-of-the-art models trained only on inorganic materials data.

The Berkeley and LBNL Transformer paper does not make the same electrolyte claim. Its OMol25 relevance is different: a large molecular dataset makes it possible to ask whether a prior-free architecture can discover atomistic structure from examples alone.

How this differs from physics-informed atomistic models

The contrast is architectural. The prevailing paradigm described in the paper prescribes locality through an interaction graph and encodes geometric features directly into model architectures. Those choices embed assumptions about geometric locality and graph structure.

The alternative tested here removes those supports. The molecular system is represented as coordinates in Cartesian space. No graph is constructed. No chemical priors are embedded. No pairwise interactions are manually defined.

Training supervision remains physical: molecular energies and forces. The question is what the architecture contributes versus what the data teaches. The reported answer is that interaction strengths resembling classical electrostatics, cutoffs resembling traditional physical models, and an adaptable receptive field can arise during training.

For readers tracking machine-learning interatomic potentials, the distinction is between encoding symmetry, locality, and interaction form in advance versus probing whether a general sequence model discovers analogous behavior. The paper's competitive-accuracy statement is limited to certain metrics, so it does not displace physics-informed architectures. It proposes a baseline against which specialized models can be compared.

What remains bounded by the September 24 record

The public record available by the archive cutoff is a preprint history, not a journal publication event. The arXiv page shows submission on October 2, 2025, revision on September 17, 2026, and the current v3 revision on September 24, 2026. The subjects are listed as Machine Learning, Materials Science, Chemical Physics, and Biomolecules.

Within that record, the supported statements are specific. Computational simulations play a central role in scientific discovery. Ab initio methods build on known physical laws but are often prohibitively expensive for practical systems. Machine-learning methods offer modeling at a fraction of the cost. Many physical-science approaches embed symmetries and other priors in architecture. Transformers have unified tasks across language, vision, and robotics without explicit domain-specific inductive biases.

The new experimental statements are equally specific. A model without domain-specific priors, including manually defined pairwise interactions, was trained on Cartesian coordinates to predict energies and forces. It recovered interaction strengths mirroring classical electrostatics. It recovered cutoffs consistent with traditional models. It showed predictable scaling with data and compute. It reached accuracy competitive with physics-informed architectures on certain metrics.

The broader suggestion follows from those points: general-purpose architectures can serve as principled baselines for scientific modeling by learning fundamental physical structure directly from data. As phrased in the abstract, that is a proposal for how to use such models, grounded in the reported emergence of physical patterns without engineered atomistic priors.

Sources referenced