Shanghai AI Lab's SciReasoner Tops 67 of 86 Science Benchmarks

The interesting move in Shanghai AI Lab's SciReasoner paper on Hugging Face is not the leaderboard sweep, it is what the model is being asked to reason over. Proteins, small molecules and inorganic crystals normally live in separate model zoos with separate tokenizers and separate opaque prediction heads. The paper puts all three into a single structure-aware vocabulary and treats structural tokens as evidence units the model can point at while it explains itself.

The reported numbers are the eye-catching part. SciReasoner claims state of the art on 67 out of 86 benchmarks, a 78% coverage rate, with Gene Ontology F_max lifting from 0.42 to 0.55 on low-homology proteins and single-step retrosynthesis accuracy going from 0.63 to 0.72. Expert reviewers reportedly rated the model's reasoning traces preferred or comparable to a frontier LLM in 98% of cases. Take those exact figures as reported, not as settled, until independent runs land.

Why it matters if you are not in a wet lab: scientific ML has mostly meant one narrow model per task, each with its own vocabulary and its own opaque head. A single foundation model that generalizes across biology, chemistry and materials, and that emits a fragment-level disconnection trace or a precursor-verification step a chemist can actually inspect, changes the ergonomics for drug discovery groups, materials teams and anyone who has to defend a prediction to a reviewer.

The honest caveat is that this is a fresh single-lab result and the 78% headline is an aggregate over benchmarks of very different difficulty. Expert-preference numbers as high as 98% deserve a raised eyebrow until the review protocol is picked apart. What the reporting does not give you is the training compute, the license the weights will ship under, or how the model holds up out of distribution.

If the results survive replication, the quieter shift is the interesting one: it becomes harder to argue that scientific reasoning needs a bespoke model per domain, and easier for smaller groups to bootstrap on top of one interpretable base.