AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AI Finds Its Third Dimension: SciReasoner Explains the Physical World From the Inside Out

Artificial intelligence has gotten remarkably good at mimicking human language, but when it comes to the complex, three-dimensional physical sciences, today’s models often struggle. Standard large language models (LLMs) view a protein or a crystal simply as a string of text characters. When asked to predict how a drug might bind or whether a material will conduct electricity, they rely on superficial statistical patterns rather than genuine physical intuition.

To bridge this gap, an international team of researchers has unveiled SciReasoner, a multimodal scientific foundation model that doesn’t just guess properties—it reasons through them using the native, three-dimensional geometry of matter.

Learning the “Language of Shape”

The core innovation behind SciReasoner is its ability to translate spatial coordinates, molecular bonds, and crystal symmetries into a unified “structure-aware” vocabulary. While a traditional LLM might look at the chemical formula for carbon (C) and fail to distinguish soft, flaky graphite from an ultra-hard diamond, SciReasoner translates the physical architecture of the atoms into distinct, addressable digital “tokens.”

This allows the model to perform what the researchers call “native structural reasoning.” Instead of serving as a black box that spits out a final prediction, SciReasoner generates an inspectable, step-by-step chain of thought, interleaving scientific explanations with actual molecular coordinates, protein residues, and crystal descriptors.

Grounded in Real Chemistry and Biology

To see how this works in practice, consider the challenge of retrosynthesis—working backward from a target medicine to find the chemical ingredients needed to build it. A traditional AI might suggest a chemical route based on recipes it memorized during training. SciReasoner, however, acts like a master organic chemist. Faced with a complex target molecule, the model physically locates a strategic carbon-oxygen (C–O) ester bond, argues in plain English why this specific bond should be severed, identifies the resulting fragments, and verifies that the starting precursors are chemically stable. This approach boosted single-step retrosynthesis accuracy from 63% to 72% over previous state-of-the-art models.

Similarly, in biology, many proteins share very little genetic sequence similarity (low homology) but fold into nearly identical 3D shapes. Standard sequence-matching tools fail to predict the functions of these “orphan” proteins. SciReasoner, however, analyzes their 3D physical interfaces. For instance, when looking at DNA-binding proteins, the model focuses its attention directly on the residues that physically touch the DNA, raising functional prediction accuracy ($F_{max}$) from 0.42 to 0.55 in tricky, low-homology scenarios.

In materials science, the model can successfully evaluate the thermodynamic stability of crystal structures like $TiGaCo_2$ by mapping its input tokens to specific symmetry groups and evaluating the density of its periodic bonding networks.

Expert Approved

Across 86 benchmarks spanning biology, chemistry, and materials science, SciReasoner claimed state-of-the-art performance on 67 tasks. More importantly, when human scientists reviewed the model’s step-by-step reasoning in a double-blind test, they preferred or tied SciReasoner’s explanations over those of frontier generalist LLMs in 98% of cases.

By transforming raw structural data into an inspectable ledger of physical evidence, SciReasoner represents a major step toward scientific AI that does not just predict what a molecule will do, but finally explains why.