Does Learning Protein Folding Generalize to Broader Reasoning?

The authors introduce FoldingCorpus, a protein-derived question-answer dataset, and Fold2Reason, a recipe that post-trains language models on structural data using both discrete structural answers and continuous 3D geometry. This approach investigates whether dense scientific structures can systematically improve general reasoning capabilities. On FoldBench, Fold2Reason achieves structure prediction scores 2.7 to 3.5 times those of Qwen3.5-9B while raising macro-average accuracy across 10 benchmarks spanning spatial, graph, scientific, and general reasoning.

On FoldBench, Fold2Reason achieves structure prediction scores 2.7 to 3.5 times those of Qwen3.5-9B.

It raises macro-average accuracy from 45.09% to 48.33% (+3.23 pp) across 10 benchmarks spanning spatial, graph, scientific, and general reasoning.