News · 2026-07-09

SciReasoner, a science AI whose reasoning experts prefer 98% of the time

SciReasoner, a new multimodal scientific foundation model, reached state-of-the-art results on 67 of 86 benchmarks spanning biology, chemistry, and materials science, and in blind evaluation domain experts preferred its reasoning traces over those of frontier large language models in 98% of cases. That second number is the more important one: it means scientists trusted not just the answers but the way the model got there.

Key facts

• What: A multimodal scientific foundation model that reasons over molecular and material structure.
• Result: State-of-the-art on 67 of 86 benchmarks; expert reviewers preferred its reasoning in 98% of blind comparisons.
• How: Encodes coordinates, topologies, and connectivities into a unified 'structure-aware vocabulary' treated as citable evidence.
• Primary source: arXiv:2607.07708, with a project page and code.

The core idea is a translation problem. A general language model reads a protein or a crystal as text -- a string of symbols with no built-in sense of three-dimensional structure. SciReasoner instead discretizes the structural elements themselves -- atomic coordinates, molecular topologies, the periodic connectivities of a crystal lattice -- into a single "structure-aware vocabulary." Each structural token becomes an addressable unit the model can point to while reasoning, the way you might cite a specific line in a document rather than paraphrasing the whole thing from memory.

That design pays off across very different sciences. In biology, it improves the annotation of protein cellular components for low-homology, orphan-like proteins -- the hard cases where a protein has few known relatives to compare against, raising the relevant score from 0.42 to 0.55. In chemistry, it improves single-step retrosynthesis (working backward from a target molecule to the reactions that could make it) from 0.63 to 0.72, and crucially it generates fragment-level "disconnection traces" -- showing which bonds it proposes breaking and why. In materials science, it resolves both high- and low-band-gap regimes and cleanly separates elemental from compound phases.

By analogy, most scientific AI today is like a brilliant but silent oracle: it gives an answer with no showing of work, so an expert cannot tell a lucky guess from real understanding. SciReasoner is built to argue its case, exposing structural evidence at each step. That is why the 98% expert-preference figure is the headline: in a field where a wrong answer delivered confidently can waste months of lab time, an explanation a chemist or biologist can inspect and challenge is worth more than a marginally higher score. This connects to broader work on making models' reasoning legible, discussed in our lesson on chain-of-thought reasoning.

Why it matters: AI-for-science has been dominated by narrow, single-domain tools -- one model for protein folding, another for materials, another for chemistry. A single model that is state-of-the-art across all three, and that experts trust to explain itself, points toward general-purpose scientific assistants that a working researcher could actually adopt. The honest caveat: "expert preferred the reasoning" is a judgment about persuasiveness and legibility, not a guarantee of correctness, and 67 of 86 benchmarks means it still loses on nearly a quarter of them. Benchmark wins also do not automatically translate into novel discoveries -- a model that annotates known proteins well has not yet designed a new drug. But as an argument that scientific AI should show its structural work, SciReasoner is a strong one.

Primary source, verified: read the paper → (arXiv 2607.07708)

Ground Truth. (2026, July 9). SciReasoner, a science AI whose reasoning experts prefer 98% of the time. Ground Truth. https://groundtruth.day/news/scireasoner-science-model-shows-its-work.html

@misc{groundtruth:scireasoner-science-model-shows-its-work,
  title  = {SciReasoner, a science AI whose reasoning experts prefer 98% of the time},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/scireasoner-science-model-shows-its-work.html}
}

Topics: AI-for-science · foundation-models · chemistry · biology · materials