What it is
Post-trains an LLM on protein-structure QA plus 3D geometry supervision to test whether folding signals transfer to broader reasoning. It improves protein folding and all 10 downstream reasoning benchmarks. Instead of training only on text or synthetic logic, it uses solved protein structures as exactly checkable supervision: LM-head answers to structural QA plus shared-state 3D geometry decoding, then discards protein-specific modules at inference.
Key results
- General-10 macro 48.33% vs 45.09% base (+3.23 pp)
- Positive gains on all 10 benchmarks; External9 improves by +3.14 pp
- FoldBench structure scores are 2.7-3.5x Qwen3.5-9B
Structure-dense scientific supervision can transfer beyond its domain and improve broad reasoning without keeping task-specific modules at inference.