Does Learning Protein Folding Generalize to Broader Reasoning?

Yong Liu; Zhanpeng Shi; Yizhou Dang; Zhongyue Zhang; Xiaoliang Shi; Zhijian Wei; Shuangjia Zheng

cs.LGSep 30, 2026

Does Learning Protein Folding Generalize to Broader Reasoning?

Authors: Yong Liu, Zhanpeng Shi, Yizhou Dang, Zhongyue Zhang, Xiaoliang Shi, Zhijian Wei, Shuangjia Zheng

Organizations: Shanghai Jiao Tong University · Fudan University · Shanghai Innovation Institute · Northeastern University

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.

Figures & tables

Figure 1: Folding performance and transfer to general reasoning. (a) Full-target folding scores on all 334 FoldBench proteins ( Xu et al., 2025 ) , for Fold2Reason and for general-purpose language models prompted to emit C

\alpha

coordinates directly. (b) Gains of Fold2Reason over its base model on the ten General-10 benchmarks, in percentage points, grouped into the four reasoning categories used throughout the paper.

Figure 2: The Fold2Reason pipeline from protein structures to general reasoning transfer. (a) Known protein structures generate deterministic FoldingCorpus labels and continuous geometry targets; ground-truth coordinates are used only for supervision. (b) The adapted model transfers to general reasoning benchmarks without protein inputs at inference. (c) A frozen language-model backbone with trainable LoRA and a shared workspace feeds FoldingCorpus and geometry readouts during post-training.

Category	Benchmark	Base	Post-training
Category	Benchmark	Base	Hidden Geom.	Format Copy	Fixed Shuffle	Fold2Reason
Spatial reasoning	FTB-Core	36.18	37.25 (+1.07)	36.04 (-0.14)	35.41 (-0.77)	40.26 (+4.08)
	SpatialViz	26.61	23.16 (-3.45)	28.56 (+1.95)	25.76 (-0.85)	32.71 (+6.10)
	VSI Bench	58.53	59.22 (+0.69)	56.32 (-2.21)	58.42 (-0.11)	60.16 (+1.63)
Graph reasoning	GraphQA Easy	65.14	66.90 (+1.76)	63.95 (-1.19)	64.91 (-0.23)	69.70 (+4.56)
Graph reasoning	GraphQA Hard	32.65	34.83 (+2.18)	33.95 (+1.30)	35.88 (+3.23)	39.48 (+6.83)

Table 1: Dataset-level General-10 results for matched FoldingCorpus-format controls and the full Fold2Reason recipe. Values are mean accuracies in percent over three seeds; parentheses report absolute change from Base in percentage points. Categories follow Figure 1 (b). Green/red cells indicate gains/declines; stronger shading indicates larger absolute changes on a shared scale. The three controls share the same LoRA recipe, supervised-token budget, seeds, and full General-10 evaluation.

Figure 3: FoldingCorpus supervision across model scales and families. (a) General-10 change relative to each model’s own base checkpoint. Bars show the three-seed mean and open circles show the individual seeds. (b) The same change resolved per benchmark, with one marker shape and colour per model and the categories of Figure 1 (b). The horizontal axis is broken to accommodate the large SpatialViz gains of the two smaller Qwen3.5 models.

Figure 4: Data scaling of structural readout and general transfer. The three panels report FoldBench lDDT-C

\alpha

, FoldBench Contact F1, and General-10 change for an independent three-seed Qwen3.5-9B Full-RG scaling run over seven nested subsets spanning 50–4,000 proteins. Every subset is trained for three epochs. Thin curves show individual seeds, center curves show means, and transparent bands with boundary lines show three-seed 95%

t

-intervals. The upper axis gives the corresponding number of FoldingCorpus records at 12 targets per protein.

Arm	G10	3D macro	FTB	SpatialViz	VSI	Text G7
Base	45.09	40.44	36.18	26.61	58.53	47.09
w/o FoldingCorpus	45.77 (+0.68)	41.47 (+1.03)	38.44 (+2.26)	27.40 (+0.79)	58.58 (+0.05)	47.61 (+0.52)
w/o Geometry	48.02 (+2.93)	43.30 (+2.86)	39.00 (+2.81)	31.81 (+5.20)	59.08 (+0.54)	50.05 (+2.96)
Fold2Reason-full	48.33 (+3.23)	44.38 (+3.94)	40.26 (+4.08)	32.71 (+6.10)	60.16 (+1.63)	50.02 (+2.93)

Table 2: Component results on external benchmarks over three seeds. Values are mean accuracies in percent; parentheses report absolute change from Base in percentage points. The 3D macro averages FTB-Core, SpatialViz, and VSI; Text G7 averages the remaining seven General-10 datasets.

Figure 5: Contact and distance maps for three example proteins. Rows show three FoldBench334 proteins of increasing length (8wt3_A, $L=134$ ; 8qjp_A, $L=250$ ; 7xg9_A, $L=286$ ), comparing ground truth, FoldingCorpus-only, and Fold2Reason. Contacts use C $\alpha$ distance $<8$ Å and sequence separation $\geq 6$ ; predicted maps show frequency across all three seeds, with gray diagonal bands marking excluded near-sequence pairs. Distance maps show the three-seed mean, clipped at 20 Å. Panels are unsmoothed and cover the complete protein.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Item	Value
Base model	Qwen3.5-9B
Protein views	336 sequence; 332 +MSA; 332 +MSA+template
Length range / mean	29–199 / 144.42 residues
FoldingCorpus records	12 per protein; 12,000 total
Packed supervision	12 labels + EOS per protein
Workspace	width 256; 16 pooled tokens; at most 2,048 pairs

Appendix

Table 3: Canonical Fold2Reason-full training contract.

Partition	Proteins	Questions/protein	FoldingCorpus records	Packed samples
Train	1,000	12	12,000	1,000
Development	100	12	1,200	100
Frozen test	100	12	1,200	100
Total	1,200	12	14,400	1,200

Appendix

Table 4: Protein data partitions and supervision counts. A FoldingCorpus record is one operator applied to one protein; a packed sample contains all 12 records.

Operator / train labels	Definition and threshold	Sampling rule
CONTACT_SHORT A470/B530	$A$ iff $d(i,j)<8$ Å	$6\leq\mathrm{sep}\leq 11$
CONTACT_MEDIUM A458/B542	$A$ iff $d(i,j)<8$ Å	$12\leq\mathrm{sep}\leq 23$
CONTACT_LONG A447/B553	$A$ iff $d(i,j)<8$ Å	$\mathrm{sep}\geq 24$
DISTANCE_ORDER_1 A481/B519	$A$ iff $d(i,j)<d(k,l)$	low/high distance deciles; local/medium pairs
DISTANCE_ORDER_2 A512/B488	$A$ iff $d(i,j)<d(k,l)$	low/high distance deciles; long-range pairs
SEGMENT_ORIENTATION A331/B297/C372	For endpoint vectors $u,v$ : A if $\cos(u,v)\geq 0.5$ ; B if $\cos(u,v)\leq-0.5$ ; C otherwise	segment starts separated by at least eight residues

Appendix

Table 5: The 12 FoldingCorpus operators. Counts are observed labels in the 1,000-protein training partition. Indices refer to valid residues; sep is sequence separation.

Task	Answer/parser	Representative generated query
nearest_neighbor	point ID / exact	choose the nearest labeled 3D point
bond_angle	degrees / $\pm 0.5$	compute an angle from three coordinates
signed_dihedral	degrees / $\pm 0.5$	compute a signed four-point dihedral
tetrahedral_chirality	positive/negative	determine the handedness of four points
proper_rigid_equivalence	yes/no	distinguish a proper rigid transform from reflection
sparse_constraints_candidate_selection	A–D	choose the structure satisfying sparse distances

Appendix

Table 6: FTB-Core v1 task composition. The final column gives one prompt schema per task; each task contributes 500 test and 500 test-OOD examples to the headline score.

Benchmark (revision)	$n$	Modality	New tokens	Score / extraction
FTB-Core v1.0.0	12,000	text	32	numeric tolerance or categorical EM
SpatialViz ( f38482a8 )	1,180	image+text	128	option-letter accuracy
VSI ( d7cb1a39 )	5,130	video+text	16	MC accuracy/MCA; aggregation below
GraphQA Easy ( 18f55c29 )	21,600	text	128	canonical answer-tail EM
GraphQA Hard ( b7e910dd )	21,600	text	128	canonical answer-tail EM
BBH ( 982bb89f )	6,511	text	128	normalized EM

Appendix

Table 7: Per-benchmark evaluation contract. All rows use greedy decoding. EM denotes the benchmark-specific normalized exact matcher; MCA is mean relative accuracy.

Item	Protocol
Initialization	Pretrained Phase-0 head-only decoder
Trainable in Phase 0	Coordinate and distogram heads only; base LM and LoRA frozen
Phase-0 data	1,000 training proteins from the OpenFold high-confidence split
Fold2Reason training	Decoder loaded with requires_grad=False and excluded from optimization
Held-out exclusions	No FoldBench334, General-10, FTB-Core, SpatialViz, VSI, GraphQA, BBH, ChemBench, ChemBench4K, Lab-Bench, or SciBench examples
Extra structural data	No additional PDB-derived proteins beyond the 1K training split

Appendix

Table 8: Frozen geometry decoder provenance.

	Input: training proteins $(s,v,Y,m,w)$ , tokenizer $T$ , fixed data seed $s_{d}$ ; $v$ is the assigned evidence view.
	Output: frozen protein inputs $x$ , marker indices $I$ , FoldingCorpus tokens $z$ , masked labels $\ell$ , and geometry targets.
1	Compute valid C $\alpha$ distances and structural summaries from $Y,m$ .
2	Build the training-pool hard-negative index used by the 32-way summary question.
3	For each protein $p$ in the training partition:
4	Serialize $(s_{p},v_{p})$ and a residue skeleton with exactly one marker per residue; tokenize to obtain $(x_{p},I_{p})$ .
5	For each of the 12 operators in Table 5 :

Appendix

Algorithm 1 Construct a FoldingCorpus training cache

	Input: frozen cache, base $\theta_{0}$ , pretrained decoder $g_{\psi}$ , training seed, three-epoch schedule.
	Output: LoRA adapter $\Delta\theta$ and workspace $\phi$ ; unchanged decoder $\psi$ .
1	Initialize rank-16 LoRA and workspace; freeze $\theta_{0}$ , $\psi$ , and the unused retrieval projection.
2	Create AdamW groups for LoRA and active workspace parameters only.
3	For each epoch and seed-shuffled gradient-accumulation window:
4	Zero optimizer gradients; distribute protein packs across workers.
5	For each active protein pack $(x,I,z,\ell,Y,m,w)$ :

Appendix

Algorithm 2 Canonical Full post-training and its Pure-LoRA specialization

	Input: declared checkpoint(s), frozen evaluation manifests, prompts, decoding settings, and scorers.
	Output: per-example predictions, per-seed metrics, and seed-aggregated results.
1	For each training seed, load its declared checkpoint in evaluation mode.
2	Held-out FoldingCorpus questions: for each protein, compute its matched workspace memory for Full; omit memory for Pure-LoRA.
3	Ask each question independently using the sequence and that question; do not append earlier gold answers.
4	Take the first next-token argmax over the complete vocabulary and compare with the canonical label token.
5	Average correctness within each operator, then average the 12 operator accuracies. Record candidate-restricted accuracy separately as a diagnostic.

Appendix

Algorithm 3 Evaluate fixed checkpoints on FoldingCorpus, FoldBench, and General-10

Model	Epochs	Steps	Peak LR	LoRA (M)	GPUs/seed
Qwen3.5-2B	3	375	1e-04	16.82	4
Qwen3.5-4B	3	375	1e-04	32.46	4
Qwen3.5-9B	3	375	1e-04	43.28	4
InternVL3.5-8B	3	375	1e-04	43.65	4
Gemma-4-12B-IT	3	375	1e-04	65.57	4

Appendix

Table 9: Model-family training contracts from the saved run configurations. All arms use rank-16, alpha-32 LoRA with dropout 0.05; the parameter count includes only trainable adapters.

Model	S1	S2	S3	Mean $\pm$ SD	95% t interval
Qwen3.5-2B	+6.448	+5.048	+4.747	$5.414\pm 0.907$	[+3.160, +7.668]
Qwen3.5-4B	+4.636	+5.155	+4.742	$4.844\pm 0.274$	[+4.164, +5.525]
Qwen3.5-9B	+3.216	+3.204	+3.249	$3.223\pm 0.024$	[+3.164, +3.281]
InternVL3.5-8B	+1.754	+1.225	+1.607	$1.529\pm 0.273$	[+0.851, +2.206]
Gemma-4-12B-IT	-0.690	+0.509	+0.398	$0.072\pm 0.663$	[-1.574, +1.718]

Appendix

Table 10: General-10 gains by model and training seed. Each row is paired with its own base checkpoint and archived evaluation contract.

Benchmark	Qwen 2B	Qwen 4B	Qwen 9B	InternVL 8B	Gemma 12B
FTB-Core	+10.45	+7.46	+4.10	+2.64	-1.23
SpatialViz	+21.21	+22.20	+6.69	-0.45	-0.79
VSI	-0.28	+3.99	+1.50	+0.20	+2.95
GraphQA Easy	+2.80	-2.82	+1.96	+0.22	-0.43
GraphQA Hard	+6.51	+3.69	+9.95	+3.42	-1.93
BBH	+1.35	+2.91	+2.38	+1.35	+1.53

Appendix

Table 11: Complete Pure-LoRA mean changes from each model’s own base (percentage points). All benchmarks and model settings are retained.

FTB-Core task	Base	Full	$\Delta$	SpatialViz task	Base	Full	$\Delta$
Bond angle	0.60	0.23	-0.37	2DRotation	0.00	5.00	+5.00
Distance audit	9.40	7.50	-1.90	3DRotation	30.00	35.42	+5.42
Fragment assembly	28.30	30.10	+1.80	3ViewProjection	31.00	38.67	+7.67
Nearest neighbor	41.50	50.47	+8.97	ArrowMoving	30.00	33.75	+3.75
Noisy template	25.80	26.97	+1.17	BlockMoving	31.25	31.25	+0.00
Rigid equivalence	47.20	48.20	+1.00	CrossSection	14.17	11.39	-2.78

Appendix

Table 12: Complete task-level mean scores for Full. Scores are percentages; changes are percentage points. Abbreviated FTB task names follow the construction table.

Task	$n$	Base VO	Full VO	Common $\Delta$
all	12000	36.18	40.26	+4.08
Bond angle	1000	0.60	0.23	-0.37
Constraint audit	1000	9.40	7.50	-1.90
Fragment assembly	1000	28.30	30.10	+1.80
Nearest neighbor	1000	41.50	50.47	+8.97
Template selection	1000	25.80	26.97	+1.17

Appendix

Table 13: FTB-Core parser audit. PF: parse-failure rate (percent of all rows); VO: accuracy conditional on a valid output (percent). Common-valid

\Delta

uses the same IDs for Base and Full in each seed. Full values are means of three seed-wise rates, not pooled votes. Overall rates use question weighting.

Task	$n$	Base PF	Base VO	Full VO	Common $\Delta$
all	1180	1.69	27.07	32.71	+5.46
2DRotation	80	0.00	0.00	5.00	+5.00
3DRotation	80	0.00	30.00	35.42	+5.42
3ViewProjection	100	0.00	31.00	38.67	+7.67
ArrowMoving	80	0.00	30.00	33.75	+3.75
BlockMoving	80	0.00	31.25	31.25	+0.00

Appendix

Table 14: SpatialViz parser audit. PF: parse-failure rate (percent of all rows); VO: accuracy conditional on a valid output (percent). Common-valid

\Delta

uses the same IDs for Base and Full in each seed. Full values are means of three seed-wise rates, not pooled votes. Overall rates use question weighting.

Benchmark	Base	w/o FoldingCorpus	w/o Geometry	Fold2Reason-full
FTB-Core	36.18	38.44 (+2.26)	39.00 (+2.81)	40.26 (+4.08)
SpatialViz	26.61	27.40 (+0.79)	31.81 (+5.20)	32.71 (+6.10)
VSI	58.53	58.58 (+0.05)	59.08 (+0.54)	60.16 (+1.63)
GraphQA Easy	65.14	67.34 (+2.20)	68.72 (+3.58)	69.70 (+4.56)
GraphQA Hard	32.65	32.63 (-0.02)	44.44 (+11.79)	39.48 (+6.83)
BBH	54.45	54.64 (+0.19)	55.17 (+0.72)	56.81 (+2.36)

Appendix

Table 15: Dataset-level component ablation results. Values are mean accuracies in percent over three seeds; parentheses report absolute change from Base in percentage points. The w/o FoldingCorpus arm removes FoldingCorpus CE, the w/o Geometry arm removes the Geometry objective, and Fold2Reason-full combines FoldingCorpus CE and Geometry. The historical w/o FoldingCorpus arm is geometry-dominant and includes the legacy retrieval auxiliary.

Comparison	S1	S2	S3	Mean $\pm$ SD	95% t interval
w/o FoldingCorpus	+0.688	+0.783	+0.556	$0.676\pm 0.114$	[+0.392, +0.959]
w/o Geometry	+2.611	+2.629	+3.551	$2.931\pm 0.538$	[+1.595, +4.266]
Full RG	+3.186	+3.115	+3.400	$3.234\pm 0.148$	[+2.866, +3.601]
Full $-$ w/o Geometry	+0.575	+0.486	-0.152	$0.303\pm 0.396$	[-0.681, +1.287]

Appendix

Table 16: Component gains and the paired Geometry increment on General-10 (percentage points).

Readout condition	TM-score $\uparrow$	lDDT-C $\alpha$ $\uparrow$	Contact F1 $\uparrow$	C $\alpha$ MAE (Å) $\downarrow$
Frozen-base readout	0.1742	0.2400	0.0664	10.753
w/o Geometry	0.1710	0.2435	0.0627	12.096
w/o FoldingCorpus	0.1685	0.2534	0.0696	12.262
Fold2Reason-full	0.1688	0.2532	0.0690	12.244

Appendix

Table 17: Complete FoldBench334 structural readout comparison over three seeds. The frozen-base row trains a decoder of the same architecture on frozen Qwen features; the three adapted rows use the Phase-0 decoder as a fixed readout. The historical w/o FoldingCorpus arm is geometry-dominant and includes a legacy retrieval auxiliary. Lower C

\alpha

MAE is better.

Figure 6: Geometry changes local readouts more consistently than global topology. Each point is one FoldBench334 protein after averaging that protein over three training seeds; distributions show Fold2Reason-full minus FoldingCorpus-only. Thick bars span the interquartile range, white circles mark medians, and diamonds mark means. Positive values favor Full except for C

\alpha

MAE, where negative is better. Full improves per-protein lDDT-C

\alpha

and Contact F1 for 75.4% and 78.7% of proteins, respectively, while TM-score and MAE contain influential tails.

Panel A: proteins displayed in Figure 5
Selection	Protein	$L$	TM-score FC-only $\rightarrow$ Full	lDDT-C $\alpha$ FC-only $\rightarrow$ Full	Contact F1 FC-only $\rightarrow$ Full	Distance MAE (Å) FC-only $\rightarrow$ Full
Rank 1	8wt3_A	134	0.2438 $\rightarrow$ 0.2742 (+0.0304)	0.2767 $\rightarrow$ 0.3034 (+0.0267)	0.1146 $\rightarrow$ 0.1169 (+0.0023)	6.0422 $\rightarrow$ 5.3820 (-0.6602)
Rank 2	8qjp_A	250	0.2255 $\rightarrow$ 0.2586 (+0.0331)	0.2462 $\rightarrow$ 0.2629 (+0.0167)	0.0377 $\rightarrow$ 0.0552 (+0.0175)	10.8590 $\rightarrow$ 9.1586 (-1.7004)
Rank 3	7xg9_A	286	0.2310 $\rightarrow$ 0.2568 (+0.0258)	0.2607 $\rightarrow$ 0.2834 (+0.0227)	0.0435 $\rightarrow$ 0.0493 (+0.0058)	12.2372 $\rightarrow$ 10.8577 (-1.3794)
Panel B: Contact-F1-change quantiles displayed in Figure 7
Lower (10th)	7urp_A	159	0.1917 $\rightarrow$ 0.2035 (+0.0118)	0.2125 $\rightarrow$ 0.2275 (+0.0150)	0.0707 $\rightarrow$ 0.0595 (-0.0112)	9.3934 $\rightarrow$ 8.0564 (-1.3371)

Appendix

Table 18: Preselected FoldBench334 cases for qualitative structural audit. Panel A matches Figure 5 : the three proteins have the highest Full mean TM-score among all 334 targets. Panel B matches Figure 7 : proteins are selected at the 10th, 50th, and 90th percentiles of the per-protein Contact F1 change. Selection uses the three-seed mean, with protein ID breaking ties, and precedes visual inspection. Each metric cell reports FC-only

\rightarrow

Full, followed by Full minus FC-only in parentheses; lower distance MAE is better.

Figure 7: Local distance-map audit for the preselected proteins of Table 18 . For each protein, the anchor is the residue with the most ground-truth contacts; the displayed neighborhood contains that anchor and its 15 nearest ground-truth C $\alpha$ neighbors. Ground truth alone determines the selection. Full changes the local distance MAE from 6.311 to 6.178 Å in the lower case, from 2.620 to 2.821 Å in the median case, and from 5.982 to 5.533 Å in the higher case. The middle case worsens while the lower and higher cases improve, matching the heterogeneous Geometry effect in the aggregate analysis. Green squares mark the anchor residue.

Representation	N	Min	Q1	Median	Q3	Max	Mean
Manifest sequence	334	28.0	140.2	225.5	344.8	1414.0	261.43
Evaluated residues	334	26.0	140.2	225.5	344.8	1414.0	261.10

Appendix

Table 19: FoldBench334 cohort size and residue-count distribution. Manifest sequence length and the residue count in the structural evaluation cache are reported separately.

Proteins	FoldingCorpus targets	Steps	FoldingCorpus macro	lDDT-C $\alpha$	Contact F1	General-10 $\Delta$
50	600	21	0.478	0.244	0.0666	0.49
100	1,200	39	0.486	0.244	0.0671	0.61
250	3,000	96	0.485	0.246	0.0666	1.37
500	6,000	189	0.492	0.247	0.0675	2.92
1,000	12,000	375	0.501	0.252	0.0694	3.31
2,000 †	24,000	750	0.546	0.252	0.0739	3.70

Appendix

Table 20: Fixed-epochs data-scaling endpoints. Values are three-seed means from the independent scaling run. General-10 values are changes from the base model in percentage points.

Figure 8: Training and transfer dynamics at pre-registered checkpoints. Columns show training losses, held-out FoldingCorpus answer accuracy, FoldBench structural readouts, and General-10 changes. Legend entries carry the run identifiers full_rg for Fold2Reason-full (blue) and pure_lora for FoldingCorpus-only Pure-LoRA (orange). The upper axis reports cumulative protein and answer-label exposures. All benchmark evaluations were run offline after training and were not used for checkpoint selection.

Dataset	N	Canonical base	Scaling base
FTB-Core	12000	36.18	36.23
SpatialViz	1180	26.61	26.61
VSI	5130	58.53	56.93
GraphQA Easy	21600	65.14	67.71
GraphQA Hard	21600	32.65	33.39
BBH	6511	54.45	54.00

Appendix

Table 21: The two frozen base-evaluation contracts used by the canonical endpoint study and the independent scaling study. Each reported gain uses the base from its own column.

Proteins	Seed	Steps	FC acc.	lDDT	Contact F1	General-10 $\Delta$
50	20260729	21	0.4608	0.2384	0.0650	+0.406
50	20260803	21	0.4875	0.2458	0.0667	+0.686
50	20260804	21	0.4858	0.2474	0.0680	+0.392
100	20260729	39	0.4758	0.2432	0.0660	+0.232
100	20260803	39	0.4975	0.2432	0.0670	+0.593
100	20260804	39	0.4850	0.2469	0.0682	+1.013

Appendix

Table 22: All fixed-three-epoch data-scaling endpoints by seed. FoldingCorpus and structure metrics are fractions; General-10 changes are percentage points.

Dataset	50	100	250	500	1K	2K †	4K †
FTB-Core	-0.13	-0.44	+0.50	+4.26	+4.26	+5.00	+4.56
SpatialViz	+0.76	+0.65	+1.53	+5.17	+5.93	+7.12	+6.92
VSI	+0.12	-0.08	+0.33	+1.17	+1.13	+1.82	+1.58
GraphQA Easy	+1.36	+2.13	+1.77	+2.58	+2.49	+2.76	+1.47
GraphQA Hard	+0.78	+0.56	+1.94	+3.90	+6.35	+8.67	+8.46
BBH	+0.57	+0.74	+2.37	+3.30	+3.77	+3.68	+2.90

Appendix

Table 23: Dataset-level changes along the complete fixed-three-epoch curve (percentage points). The final row reports the macro mean and training-seed SD.

Figure 9: Dataset-level transfer along the fixed-three-epoch curve. Cells show the three-seed mean change from the scaling study’s frozen base in percentage points. A common diverging color scale is centered at zero. Asterisks mark the fresh 2K and 4K extension runs; the 4K pool also changes the confidence-quality support described above.

Labels/protein	S1	S2	S3	$\Delta\pm$ SD	FC acc.	lDDT	Contact F1
3	+2.971	+3.731	+3.472	$3.391\pm 0.386$	$0.4942\pm 0.0175$	$0.2544\pm 0.0020$	$0.0690\pm 0.0007$
6	+3.145	+2.441	+3.268	$2.951\pm 0.446$	$0.4994\pm 0.0047$	$0.2542\pm 0.0028$	$0.0693\pm 0.0003$
12	+3.145	+2.941	+3.848	$3.311\pm 0.476$	$0.5014\pm 0.0155$	$0.2523\pm 0.0012$	$0.0694\pm 0.0005$

Appendix

Table 24: FoldingCorpus-label density at fixed 1,000 proteins and 375 optimizer steps. S1–S3 and the first mean report General-10 changes (pp); the remaining metrics are fractions.

Figure 10: FoldingCorpus-label density with 1,000 proteins and 375 steps. Solid lines show means over three training seeds. Shaded regions extend from mean minus one sample SD to mean plus one sample SD; the thin boundary lines mark those limits. FoldingCorpus accuracy uses the frozen test split. FoldBench uses all 334 targets, and external transfer uses the scaling evaluation contract.

File	Contents
endpoint_seed_scores.json	All ten dataset scores by seed for Fold2Reason-full, the component arms, the model families, and the source-control aggregates.
model_training_contracts.json	Saved LoRA modules, parameter counts, steps, seed, and hardware world size for the fifteen model-family runs.
task_breakdown.json	Complete FTB split/family/task, SpatialViz category/task, and VSI question-group scores.
foldbench_audit.json	The 334 target IDs, sequence and evaluation length distributions, and structural metrics by arm and seed.
scaling_evidence.json	Fixed-three-epoch endpoints, label-density results, and numerical checkpoint trajectories.
scaling_evaluation_manifest.json	Archived checkpoint and split assignments and expected evaluation counts for the included original scaling runs.

Appendix

Table 25: Machine-readable evidence included with the appendix source package.

Explore similar work

Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning
Jul 8, 2026Chen Tang, Yizhou Wang, Jianyu Wu +26Scientific Discovery Reasoning Skills
Shanghai Artificial Intelligence Laboratory, China. · The Chinese University of Hong Kong, Hong Kong. · Shanghai Jiao Tong University, China. +8
How Post-Training Shapes Biological Reasoning Models
Jun 15, 2026Lukas Fesser, Hanlin Zhang, Michelle M. Li +5Large Reasoning Models Post-Training
Harvard University · Google DeepMind · Google Research
Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores
Aug 31, 2026Qiyao Yan, Chenpeng Wang, Liangming PanLLM Reasoning Strategies Reasoning Benchmark
State Key Laboratory of Multimedia Information Processing, Peking University · School of Computer Science, Peking University · YiXin-AILab, YIXIN, Beijing, China +1

Jul 8, 2026, cs.CLJ/K move · Enter open · S save

Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning

Chen Tang, Yizhou Wang, Jianyu Wu, Lintao Wang+25

Shanghai Artificial Intelligence Laboratory, China. · The Chinese University of Hong Kong, Hong Kong. · Shanghai Jiao Tong University, China.+8

Scientific Discovery Reasoning Skills Interdisciplinary Research

Open paper AskPDF arXiv

Structure-property relationships are foundational to biology, chemistry and materials science, where function, reactivity and physical response emerge from spatial, chemical and periodic organization. Mechanistically explaining these relationships requires interpreting structural evidence through scientific principles and physical constraints, from stereochemistry and bonding to symmetry, energetics and periodic order. However, applying artificial intelligence to this process presents a joint challenge of representation and reasoning: models must preserve domain-native structural information while showing how specific evidence supports predictions under these constraints. Here we introduce SciReasoner, a multimodal scientific foundation model for native structural reasoning across proteins, small molecules and inorganic crystals. SciReasoner discretizes coordinates, topologies and periodic connectivities into a unified structure-aware vocabulary, treating structural tokens as addressable evidence units during reasoning. In homology-controlled Gene Ontology prediction, SciReasoner improves Cellular Component annotation for low-homology and orphan-like proteins, increasing

F_{\max}

from 0.42 to 0.55. In chemistry, it raises single-step retrosynthesis accuracy from 0.63 to 0.72 while generating fragment-level disconnection and precursor-verification traces. In materials science, its representations separate elemental and compound phases and resolve high- and low-band-gap regimes. Across 86 benchmarks, SciReasoner achieves state-of-the-art performance on 67 tasks. Double-blind expert evaluation rates its reasoning traces as preferred or at least comparable to those of a frontier large language model in 98% of cases. By making structure an inspectable substrate for reasoning under scientific constraints, SciReasoner connects accurate prediction with interpretable scientific inference.

Abstract

Explore similar work

Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning

How Post-Training Shapes Biological Reasoning Models

Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores