Does Learning Protein Folding Generalize to Broader Reasoning? Fold2Reason Transfers 3D Structural Logic to Boost 10 General Reasoning Benchmarks

核心背景与行业痛点

Large language models (LLMs) continue to suffer from logical brittleness and hallucinations when addressing spatial, physical, and geometric reasoning tasks. This stems from their fundamental training substrate: human-authored natural language texts convey compressed semantic conclusions rather than the rigorous continuous geometries and topological constraints of the physical universe. Concurrently, in structural biology and AI for Science, protein structure determination represents an objectively verified triumph: every experimentally resolved macromolecular structure yields thousands of mathematically verifiable 3D spatial distances, coordinate assertions, and contact topologies. This raises an intriguing foundational hypothesis: can learning the spatial and topological principles of protein folding provide language models with transferable, reusable logical reasoning capabilities across general domains?

架构亮点与底层机制

Researchers from Zhejiang University, Shanghai Jiao Tong University, and BioMap demonstrate this transferability through Fold2Reason (arXiv:2609.38879):

1. FoldingCorpus Dataset Creation: Translates thousands of verified crystallographic protein structures into question-answering pairs, probing residue-level spatial proximities, contact surfaces, and topological packings.
2. Complementary Dual-Signal Supervision: Post-trains LLMs using two synergistic training objectives simultaneously: discrete structural propositions are answered via the model's native autoregressive language head, while continuous 3D atomic coordinates are decoded concurrently from the exact same shared internal representations.
3. Implicit Topological Internalization: Propagating geometric loss gradients through shared representations forces the model to internalize continuous SE(3) spatial invariance and spatial connectivity within its attention layers, enriching abstract semantic embeddings with physical topological intuition.

权威 Benchmark 与实测跑分对比

Evaluated on the domain-specific FoldBench suite and ten external multi-domain reasoning benchmarks:

1. 2.7x to 3.5x Surge on FoldBench: Fold2Reason scores 2.7x to 3.5x higher than the baseline Qwen3.5-9B foundation model across structural biology benchmarks.
2. Consistent Gains Across All 10 External Reasoning Benchmarks: Spanning spatial reasoning, graph topology tasks, scientific inquiry, and general logic, Fold2Reason achieved uniform positive improvements on all 10 evaluated benchmarks without domain degradation.
3. Macro-Average Accuracy Climbs from 45.09% to 48.33% (+3.23 pp): Delivers a +3.23 percentage point leap in macro-average accuracy; rigorous control groups trained on synthetic, randomized, or scrambled structures yielded negligible or negative transfers, validating that physical, non-linguistic scientific structure drives generalized cognitive gains.

开发者实战落地与开箱指南

The Fold2Reason code repository and FoldingCorpus datasets have been open-sourced on GitHub (GENTEL-lab/Fold2Reason). This discovery demonstrates that post-training foundation models does not solely rely on collecting synthetic language tokens or scaling human feedback; instead, scientifically validated 3D physical structures can serve as abundant ground-truth reasoning supervision. Applied AI engineers can incorporate this dual-signal geometric pre-training pipeline into general-purpose coding agents, embodied spatial reasoning engines, and scientific problem-solving architectures.