Does Learning Protein Folding Generalize to Broader Reasoning?
Yong Liu, Zhanpeng Shi, Yizhou Dang, Zhongyue Zhang, Xiaoliang Shi, Zhijian Wei, Shuangjia Zheng
Shanghai Jiao Tong University · Fudan University · Shanghai Innovation Institute · Northeastern University
Published September 30, 2026 · on the feed October 5, 2026 · ▲ 120 · arXiv 2609.38879
The abstract
Large language models rely heavily on human text, which often conveys surface answers rather than the spatial and structural logic behind them. Protein folding is a natural testbed, because one solved structure yields thousands of exactly checkable spatial and topological statements. We ask: can learning to fold proteins teach general models reusable reasoning capabilities? To answer this, we build FoldingCorpus, a protein-derived question-answer dataset, and Fold2Reason, a recipe that post-trains on it through two complementary signals: discrete structural answers predicted via 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 pp), with positive gains on all 10 benchmarks, while matched controls built from random, synthetic, and shuffled structure yield substantially smaller or negative gains. Our work shows that non-linguistic, structure-dense scientific data can systematically improve broad reasoning in language models, making a solved scientific problem a practical source of post-training supervision.
The Brief
2:22A short podcast summary of the paper. Written and voiced by AI.
Transcript
Does Learning Protein Folding Generalize to Broader Reasoning?. By Yong Liu and six others, from Shanghai Jiao Tong University, Fudan University, and two other institutions. Published on the thirtieth of September. Its code repository is worked out: about forty source files, with a readme.
Large language models learn mostly from human text, and human text tends to carry surface answers rather than the spatial and structural logic behind them. The authors ask whether learning to fold proteins can teach a general model reasoning it can reuse elsewhere. Protein folding suits the question because one solved structure produces thousands of spatial and topological statements, each exactly checkable. A single protein, in other words, is a dense source of questions with answers that can be verified, which is rarely true of ordinary text.
Their answer is yes. Non-linguistic scientific data that is dense in structure can systematically improve broad reasoning in language models.
To test it, they build FoldingCorpus, a question-answer dataset derived from proteins, and Fold2Reason, a recipe that post-trains on it with two complementary signals. One is discrete structural answers, predicted through the model's native language head. The other is continuous three-dimensional geometry, decoded from the same shared representations. The model is asked to say the answer in words and to hold the shape in its internal representation at the same time.
On FoldBench, Fold2Reason scores between two point seven and three point five times what Qwen3.5-9B scores on structure prediction. Beyond proteins, it improves all ten benchmarks tested, spanning spatial, graph, scientific and general reasoning. Macro-average accuracy rises from forty-five point zero nine percent to forty-eight point three three percent, a gain of three point two three percentage points.
The controls are what tie the gain to real structure. Matched training sets built from random, synthetic, or shuffled structure produce substantially smaller gains, or negative ones. So the improvement is not explained by extra training of any kind; it depends on the structure being genuine. The breadth claim rests on those ten benchmarks, and the gains there are positive on every one of them, though they are measured in a few points of macro-average accuracy rather than the multiples seen on structure prediction itself.
The suggestion is a practical one. A scientific problem that has already been solved stops being only a result and becomes a source of post-training supervision, offering a kind of reasoning that ordinary text rarely spells out.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Text is running out, and it hides its reasoning. Where else could structured supervision come from?
You are improving a language model's reasoning. Human text mostly gives surface answers, not the spatial and structural logic behind them, and good text is running out.
The road ahead
- 1 · The problem ~7 min Language models learn reasoning mostly from human text, which shows answers but rarely the structure behind them, and good text is running out.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Turn solved protein structures into checkable questions, train a language model on them with two signals, then test it on reasoning with no proteins.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Build checkable questions from solved proteins, post-train with answer and geometry losses, then test the adapted model alone on general reasoning.
- 6 · The experiment ~6 min You’ll guess: After post-training on protein-folding questions, by how many percentage points did macro-average accuracy rise across general reasoning benchmarks with no proteins?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.