Large language models are trained predominantly on human text, which often captures surface-level answers rather than the spatial and structural logic underlying them. A new research paper asks whether learning to fold proteins can teach general models reusable reasoning capabilities.
To investigate, the authors built FoldingCorpus, a protein-derived question-answer dataset, and Fold2Reason, a post-training recipe that uses two complementary signals: discrete structural answers predicted through the model's native language head, and continuous 3D geometry decoded from the same shared representations.
On FoldBench, Fold2Reason achieves structure prediction scores 2.7 to 3.5 times those of Qwen3.5-9B. Beyond protein structure prediction, it improves performance on all 10 benchmarks spanning spatial, graph, scientific, and general reasoning, raising macro-average accuracy from 45.09% to 48.33% (+3.23 percentage points).
Control experiments using random, synthetic, and shuffled structure data yield substantially smaller or negative gains, suggesting the improvements stem from the real structural information in protein geometry rather than simply adding more training data.
The work indicates that non-linguistic, structure-dense scientific data can systematically improve broad reasoning in language models, positioning a solved scientific problem as a practical source of post-training supervision.