PaperlayerView exact revision on arXiv

Source-linked AI summary

Does Learning Protein Folding Generalize to Broader Reasoning?

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

arXiv:2609.38879v1cs.LGcs.AI

TL;DR

The paper asks whether training on protein structures can support reasoning beyond protein tasks, where supervision typically relies on human text. It builds FoldingCorpus and post-trains models with answer and geometry signals; performance improves across all 10 reasoning benchmarks, with a higher General-10 macro-average.

  • Problem

    Which properties of supervision support transfer beyond a model’s source domain remains open, motivating tests of structured scientific data as training supervision.

    Source: §1 · p. 2

  • Method

    The paper builds FoldingCorpus from protein structures and post-trains models with verified structural answers and continuous geometry supervision on shared adapted representations.

    Source: §1 · p. 2

  • Results

    General-10 macro-average accuracy rises from 45.09% to 48.33% (+3.23 pp), with positive dataset means across all 10 benchmarks.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The results support behavioral transfer from protein-derived supervision, while leaving open whether it reflects reusable reasoning, more reliable use of pretrained capabilities, or answer-format adaptation.

    Source: §6 · p. 9

    The evidence is limited to 1,000 training proteins, rank-16 adapters, three seeds, and General-10; remote-homology and novel-fold generalization remain unverified.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

The paper asks which kinds of supervision support transfer beyond a model’s training domain, using protein structures as a source of checkable spatial targets. It introduces FoldingCorpus and Fold2Reason, and reports gains across general reasoning benchmarks.

1 INTRODUCTION

  • Prior work shows specialized training can improve behavior beyond its source domain, but broader behavior can also remain unchanged or worsen, leaving transfer-supporting supervision properties open..Source: §1 · p. 2
  • Solved protein coordinates provide deterministic contact, distance, orientation, and coordinate targets without additional annotation.This makes protein structures a setting for testing transfer from scientific structures to non-protein tasks.Source: §1 · p. 2
  • Fold2Reason post-trains Qwen3.5-9B with verified FoldingCorpus answers through its native language head and continuous geometry supervision on shared adapted representations.Downstream evaluation uses the adapted base model without protein inputs or the training-time workspace and geometry decoder.Source: §1 · p. 2
  • Across three adapters, Fold2Reason raises General-10 from 45.09% to 48.33% (+3.23 pp), with positive dataset means on all 10 benchmarks.The reported gains span text, image, and video tasks.Source: §1 · pp. 2–3
  • FoldingCorpus contains 1,200 proteins and 14,400 question–answer records across 12 structural operators, with cluster-disjoint partitions and independently recomputed answers.The core split uses 1,000 training, 100 development, and 100 frozen-test proteins.Source: §1 · p. 2

2 RELATED WORK

Prior work connects protein structure and language models to molecular capabilities, and separately studies which training data improves general capability. This paper instead uses protein structure as training supervision and evaluates reasoning after removing the protein interface.

2 RELATED WORK

  • Protein–language alignment efforts evaluate molecular or protein capabilities, whereas this work uses protein structure as training signal and evaluates downstream reasoning after detaching the protein view, workspace, and geometry decoder.Related work also includes protein language models and alignment of sequence models with structural graph encoders.Source: §2 · p. 3
  • Data-centric training work includes task-targeted selection, synthesized corpora, and verifiable-answer supervision, while FoldingCorpus draws answer targets from three-dimensional protein coordinates.Prior studies report mixed results on how far such training transfers.Source: §2 · p. 3

3 METHOD

The method converts known protein structures into discrete question answers and continuous geometry targets while keeping coordinates out of the model input. A shared workspace connects both training signals to the same adapted representations.

3.1 TRAINING TARGETS: KNOWN ANSWERS, HIDDEN EVIDENCE

  • Known coordinates generate targets and losses but remain outside the model input; the frozen backbone has trainable LoRA parameters.The model processes a protein prompt with one marker per residue.Source: §3.1 · p. 3
  • Each protein yields 12 packed questions whose labels are computed from coordinates by deterministic structural operators, with numerical evidence retained for audit rather than shown in the prompt.The operators cover contacts, distances, orientations, proximity, local direction, chirality, multi-constraint queries, and global summaries.Source: §3.1 · p. 3

3.2 ONE WORKSPACE, TWO READOUTS

  • A shared workspace reduces residue states, exchanges local and sampled long-range messages, and returns residue-aligned states E plus pooled evidence tokens M.Sixteen learned queries pool the residue set into M while E preserves residue-level correspondence.Source: §3.2 · p. 4
  • The FoldingCorpus path prepends M to packed questions and supervises only the answer tokens and EOS, while the geometry path sends E to a frozen coordinate and distogram decoder.The geometry loss combines coordinate, pair-distance, contact, distogram, local-frame, torsion, and radius-of-gyration terms.Source: §3.2 · p. 4
  • Freezing the geometry decoder forces its objective to update the shared workspace and LoRA, and gradients from both answer and geometry losses reach adapted representations.At transfer evaluation, the workspace and decoder are discarded and only the learned LoRA adapter is applied to the native benchmark interface.Source: §3.2 · pp. 4–5

4 EXPERIMENTAL DESIGN

The experiments train Qwen3.5-9B on 1,000 proteins and evaluate the adapted model on a ten-benchmark reasoning suite and a separate protein-folding benchmark. Ablations and matched controls distinguish the contributions of training components and supervision sources.

4 EXPERIMENTAL DESIGN

  • Training uses 1,000 proteins, 12,000 FoldingCorpus records, three epochs, 375 optimizer steps, and three training seeds.The protein views comprise sequence-only, sequence-plus-MSA, and sequence-plus-MSA-plus-template inputs.Source: §4 · p. 5
  • General-10 is the unweighted macro across FTB-Core, spatial, graph, broad reasoning, chemistry, laboratory, and scientific benchmarks; FoldBench334 evaluates protein structures with four metrics.The evaluation also reports External9, which excludes the internally constructed FTB-Core benchmark.Source: §4 · p. 5
  • Results are averaged across three independently trained adapters, and component ablations share the base model, endpoint, seeds, and General-10 protocol.The ablations are designed to localize which loss path drives external transfer and which improves structural readout.Source: §4 · pp. 5–6

5 RESULTS

Across the results, FoldingCorpus supervision improves General-10 reasoning across benchmarks and model families, while transfer varies with training scale and architecture. Ablations separate broad behavioral transfer from Geometry’s more localized structural and 3D-oriented effects.

5.1 GENERAL-10 TRANSFER AND MATCHED FOLDINGCORPUS-FORMAT CONTROLS

  • General-10 macro accuracy rises from 45.09% to 48.33% (+3.23 pp), with positive mean changes on all 10 datasets.The aggregate gain remains positive when GraphQA Easy and Hard are removed (+2.62 pp) or FTB-Core is excluded (+3.14 pp).Source: §5.1 · p. 6
  • Hidden Geometry improves General-10 by 0.92 pp, while Format Copy changes it by −0.09 pp and Fixed Shuffle by −0.27 pp; protein-derived Fold2Reason reaches +3.23 pp.The controls distinguish coherent random-geometry targets from answer copying and fixed input–label mismatch.Source: Table 1 · p. 6, §5.1 · p. 6

5.2 FOLDINGCORPUS SUPERVISION TRANSFERS ACROSS MODEL SCALES AND FAMILIES

  • FoldingCorpus raises General-10 by 5.41, 4.84, and 3.22 pp for Qwen3.5-2B, 4B, and 9B, respectively, and by 1.53 pp for InternVL3.5-8B.The InternVL result is positive for every seed.Source: Figure 3 · p. 7, §5.2 · p. 7
  • Gemma-4-12B-IT remains near its base level (+0.07 pp), while Figure 3 shows contributions from all four reasoning categories and especially large SpatialViz gains for smaller Qwen models.The Gemma result has two positive and one negative seed.Source: Figure 3 · p. 7, §5.2 · p. 7

5.3 GENERAL TRANSFER UNDER JOINT SCALING OF PROTEIN COVERAGE AND TRAINING STEP

  • General-10 gains rise from +0.49 pp at 50 proteins to a peak of +3.70 pp at 2,000, then remain +3.02 pp at 4,000 proteins under the three-epoch schedule.Protein coverage and optimization steps increase together in this scaling study.Source: Figure 4 · p. 7, §5.3 · p. 8
  • At 4,000 proteins, held-out FoldingCorpus accuracy becomes unstable at 0.418, yet General-10 remains positive; transfer is not determined by endpoint accuracy alone.FoldBench lDDT-Cα rises from 0.244 to 0.252 and then saturates, while Contact F1 continues rising from 0.0666 to 0.0768.Source: §5.3 · p. 8

5.4 ABLATIONS: FOLDINGCORPUS DRIVES BROAD TRANSFER WHILE GEOMETRY CONCENTRATES ON 3D

  • Table 2 separates component roles: w/o Geometry yields +2.93 pp General-10, while w/o FoldingCorpus preserves structural readouts near full but gives smaller FTB-Core and SpatialViz gains.With the workspace retained, Geometry adds 1.26 pp on FTB-Core, 0.90 pp on SpatialViz, and 1.08 pp on VSI over w/o Geometry.Source: Table 2 · p. 8, §5.4 · p. 9
  • Across three example proteins, Fold2Reason produces sharper contact frequencies and more pronounced distance-map block structure than FoldingCorpus-only, but both remain far from ground-truth global topology.The maps compare ground truth, FoldingCorpus-only, and Fold2Reason for proteins of lengths 134, 250, and 286 residues.Source: Figure 5 · p. 8, §5.4 · p. 9

6 DISCUSSION AND CONCLUSION

The discussion presents protein-derived answer supervision as the main source of broad behavioral transfer, with Geometry adding more localized structural and 3D-oriented effects. It also leaves the mechanism and several generalization boundaries unresolved.

6 DISCUSSION AND CONCLUSION

  • Fold2Reason combines verified FoldingCorpus answers and continuous geometry supervision, with separable effects: most of the 3.23 pp General-10 gain comes from answers, while Geometry adds 3D-oriented gains and local structural decodability.The geometry signal raises lDDT-Cα and Contact F1, while TM-score dips marginally and global folding quality stays low.Source: §6 · p. 9
  • Real protein structure is strongest among the tested sources, but whether transfer reflects reusable reasoning, more reliable use of pretrained capabilities, or answer-format adaptation remains unresolved.Hidden random-geometry targets improve the aggregate, while format copying and fixed shuffled labels remain near zero on average.Source: §6 · p. 9
  • Transfer peaks at 2,000 proteins when coverage and compute scale together; increasing labels per protein at fixed coverage and optimizer steps gives no monotonic gain.Gains extend to InternVL3.5-8B and Qwen3.5-2B, 4B, and 9B, but not Gemma-4-12B-IT.Source: §6 · p. 9
  • The evidence is bounded by 1,000 training proteins, rank-16 adapters, three seeds, and General-10; FoldBench remote-homology and novel-fold generalization remain unverified.The frozen decoder measures structural decodability rather than accurate global reconstruction; task heterogeneity and answer-format ambiguity also qualify behavioral gains.Source: §6 · pp. 9–10

A TRAINING AND DATA CONTRACT

The section specifies the canonical training contract, protein-data preparation, FoldingCorpus construction, and the FTB-Core benchmark and its audit. It also distinguishes cluster-disjoint partitions from broader forms of isolation that remain unverified.

A TRAINING AND DATA CONTRACT

  • Table 3 identifies the canonical Fold2Reason-full training contract, while Table 4 distinguishes an operator–protein record from a packed sample containing all 12 records.Three instrumentation reruns use canonical seeds, data order, rank-16 LoRA, a shared workspace, and a frozen decoder; logged losses and gradient norms are diagnostics, not checkpoint-selection criteria.Source: Table 3 · p. 15, Table 4 · p. 16, §A.1 · p. 16

A.2 PROTEIN SOURCE, PARTITIONS, AND STRUCTURAL PREPROCESSING

  • Protein structures are filtered to selected relaxed monomers with standard amino acids and finite backbone atoms, then centered, framed, and quantized before coordinate caching.Distances and contacts are recomputed from cached coordinates and are invariant to translation and proper rotation; the right-handed frame preserves chirality.Source: §A.2 · p. 16
  • Partitions have disjoint upstream MMseqs2 clusters at 30% identity and 80% coverage, but this does not establish exhaustive pairwise or template isolation.Full MMseqs2, deposition-date, CATH, and SCOP isolation remain unverified; the separate scaling-pool audit reports nonzero template-source and alignment overlaps.Source: §A.2 · p. 16

B FOLDINGCORPUS CONSTRUCTION

  • FoldingCorpus applies 12 structural operators to each protein, with deterministic class and candidate ordering, valid-tuple sampling, and a fallback when a requested class has no valid tuple.Every answer is independently recomputed from coordinates, and all 14,400 exported records pass schema, candidate, split, and answer-recomputation checks.Source: B FOLDINGCORPUS CONSTRUCTION · p. 17, Table 5 · p. 17

C FTB-CORE CONSTRUCTION AND AUDIT

  • FTB-Core v1 contains train, validation, test, and size/range-shifted test-OOD splits; the paper evaluates 500 examples per split for each of 12 tasks.The paper uses all 12,000 test and test-OOD examples.Source: C FTB-CORE CONSTRUCTION AND AUDIT · p. 18, Table 6 · p. 18
  • FTB-Core audits hashes, schemas, IDs, numeric targets, answer formats, and task counts; exact prompt overlap is zero, though semantic near-duplicates remain possible.Answer correctness is exhaustively program-audited, but prompt clarity lacks separately logged human validation.Source: C FTB-CORE CONSTRUCTION AND AUDIT · p. 18

D GENERAL-10 EVALUATION PROTOCOL

The section defines how benchmark inputs are prepared and scored, then describes the frozen geometry decoder and training path. It specifies how cached protein examples supervise the model and how three evaluation interfaces measure different outcomes.

D GENERAL-10 EVALUATION PROTOCOL

  • General-10 uses benchmark-specific prompts and deterministic decoding; malformed, missing, non-finite, or out-of-tolerance answers receive zero credit.Gemma uses fixed answer-prefix wrappers for SpatialViz and VSI, but equivalent audits are unavailable for other model-family settings, and the diagnostics do not fully separate reasoning from answer-format or answer-selection changes.Source: Table 7 · p. 18, D GENERAL-10 EVALUATION PROTOCOL · p. 19

E GEOMETRY DECODER PRETRAINING PROTOCOL

  • The pretrained geometry decoder is held fixed during training, receives residue representations as input, and is removed for external evaluation.Its structural readout prior comes from 1,000 training proteins; General-10 evaluates behavioral transfer after the protein and decoder interfaces are removed.Source: E GEOMETRY DECODER PRETRAINING PROTOCOL · pp. 19–20, Table 8 · p. 20

F ALGORITHMIC SPECIFICATION AND WORKED DATA EXAMPLES

  • The training cache is built from protein sequences and assigned evidence views, with coordinates used offline to construct and independently check structural answers.Cached prompts, answer tokens, masked labels, geometry targets, and audit metadata are saved for reuse across epochs.Source: §F.1 · p. 20

F.1 CONSTRUCTING THE FROZEN TRAINING CACHE

  • Each protein pack contains 12 structural question–answer pairs; the cache masks prompt positions and supervises answer and EOS tokens.The generation procedure checks candidate membership and independently recomputes answers from the structural targets.Source: §F.1 · p. 20, §F.3 · p. 21

F.2 POST-TRAINING WITH SHARED AND LANGUAGE-ONLY PATHS

  • The canonical Full run combines causal answer loss with geometry loss through a shared workspace and frozen decoder, while updating only LoRA and workspace parameters.The Pure-LoRA specialization omits the workspace and geometry forward pass and optimizes only LoRA parameters.Source: §F.2 · p. 21

F.3 PACKED DATA AND SUPERVISION MASKS

  • Full prepends 16 workspace embeddings to the packed answer sequence, while Pure LoRA uses the same cached answers without those embeddings or the protein-stream forward pass.Prompt tokens are masked from the answer loss.Source: §F.3 · p. 21

F.4 EVALUATION INTERFACES AND AGGREGATION

  • FoldingCorpus accuracy evaluates one-token decisions, FoldBench evaluates the coordinate readout, and General-10 evaluates the adapted language model on downstream inputs.General-10 pairs adapted and base scores by sample ID under identical prompts and scoring; each dataset has equal weight in the macro-average.Source: §F.4 · p. 22

G MODEL-SCALE AND MODEL-FAMILY RESULTS

The section compares FoldingCorpus transfer across model families and presents complete benchmark profiles. Transfer is positive across seeds for several families, while the reported Gemma mean is near neutral; the profiles are within-model comparisons, not a common-budget ranking.

G MODEL-SCALE AND MODEL-FAMILY RESULTS

  • All three Qwen sizes and InternVL improve in each of three seeds, whereas Gemma’s mean change is +0.072 pp with a t interval spanning zero.The reported mean General-10 gains for the three Qwen sizes and InternVL are 5.414, 4.844, 3.223, and 1.529 pp, respectively.Source: G MODEL-SCALE AND MODEL-FAMILY RESULTS · p. 23

G.1 COMPLETE DATASET PROFILES

  • The complete profile retains every benchmark and model, including negative means, but its rows are within-model comparisons rather than a common-budget ranking.The nine-billion-parameter Qwen reference uses Pure LoRA and the corrected answer-tail GraphQA scorer; other model rows retain their archived scoring contracts.Source: §G.1 · p. 23, Table 11 · p. 23

H TASK-LEVEL DECOMPOSITION OF SPATIAL TRANSFER

The task-level audit finds gains across spatial benchmarks, but their distribution is uneven and some comparisons have interpretation limits. Component comparisons attribute most broad transfer to FoldingCorpus supervision, while the additional Geometry increment is small and seed-sensitive.

H TASK-LEVEL DECOMPOSITION OF SPATIAL TRANSFER

H.1 OUTPUT VALIDITY, EXTRACTION FAILURES, AND ANSWER-SELECTION BIAS

  • On sparse-constraint selection, all FTB-Core outputs are valid, but Base selects A/B on 946/1,000 questions despite approximately balanced reference labels, leaving answer-selection bias a possible contributor.The comparison does not establish uniform spatial improvement.Source: §H.1 · p. 24
  • On the same 1,160 Base-valid SpatialViz questions, Full scores 32.53% versus 27.07% for Base, a +5.46 pp difference.Base has 20 parse failures, while every Full seed has valid outputs; the common-valid comparison is diagnostic, not a counterfactual estimate of reasoning with output format held fixed.Source: §H.1 · pp. 24–25, Table 14 · p. 25

I COMPLETE COMPONENT ABLATION RESULTS

  • The component arms use shared-contract comparisons to assess the two loss paths, but a full factorial interaction analysis remains a separate experiment.The historical w/o FoldingCorpus arm is geometry-dominant and includes a legacy retrieval auxiliary.Source: I COMPLETE COMPONENT ABLATION RESULTS · p. 26, Table 15 · p. 26

I.1 SEED-LEVEL COMPONENT COMPARISONS

  • FoldingCorpus supervision accounts for most broad transfer; the Full-minus-w/o-Geometry increment averages 0.303 pp, with seed differences of +0.575, +0.486, and −0.152 pp.Its t interval spans zero, and the direct Pure-LoRA comparison shows no resolved aggregate advantage for the additional modules.Source: §I.1 · p. 26

J NATIVE LANGUAGE-HEAD FOLDING EVALUATION

The native language-head evaluation compares direct coordinate generation from the base model with Fold2Reason checkpoints on the complete 334-protein monomer cohort. The reported native coverage averages 97.22% of reference-valid residues.

J NATIVE LANGUAGE-HEAD FOLDING EVALUATION

  • The unadapted Qwen3.5-9B baseline generates one Cα coordinate row per residue through its native language head; Fold2Reason is evaluated using epoch-3 checkpoints with the workspace and frozen decoder.The scorer maps finite coordinate rows by unique residue indices and retains partial predictions without coordinate imputation.Source: J NATIVE LANGUAGE-HEAD FOLDING… · pp. 26–27
  • Native predictions cover 97.22% of reference-valid residues on average across the 334-protein evaluation cohort.All proteins receive equal weight, including three without usable coordinates, which score zero.Source: J NATIVE LANGUAGE-HEAD FOLDING… · p. 27, Table 17 · p. 27

K STRUCTURAL READOUT AUDITS

  • With amino-acid character matching additionally required, native scores are 2.05/2.22/1.05 for TM-score/lDDT-Cα/Contact F1, respectively, on the same 334-protein denominator.The full structural-readout comparison includes a frozen-base decoder and three adapted rows, with lower Cα MAE better.Source: J NATIVE LANGUAGE-HEAD FOLDING… · p. 27, Table 17 · p. 27

K.1 PER-PROTEIN STRUCTURAL EFFECTS

Protein-level audits show that Geometry changes local structural readouts more consistently than global topology, with heterogeneous per-protein outcomes. The 334-target cohort is complete for the stated monomer list, but homology and deposition-date isolation remain incomplete.

K.1 PER-PROTEIN STRUCTURAL EFFECTS

  • Full improves per-protein lDDT-Cα and Contact F1 for 75.4% and 78.7% of proteins, respectively, while TM-score and Cα MAE contain influential tails.Figure 6 compares Full with the workspace-equipped w/o-Geometry arm, not Pure LoRA.Source: §K.1 · p. 27, Figure 6 · p. 28
  • The selected lower, median, and higher Contact-F1-change cases show local distance MAE changes of 6.311→6.178 Å, 2.620→2.821 Å, and 5.982→5.533 Å, respectively.The middle case worsens while the lower and higher cases improve; selection used ground truth before visual inspection.Source: Figure 7 · pp. 28–29

K.2 WHY THE STRUCTURAL COHORT CONTAINS 334 PROTEINS

  • The cohort comprises all 334 distinct targets in the local FoldBench monomer-protein manifest, rather than a performance-based subsample; other FoldBench task categories are outside this readout evaluation.The table reports manifest sequence lengths separately from residue counts in the structural-evaluation cache.Source: §K.2 · pp. 29–30, Table 19 · p. 30

K.3 STRUCTURAL READOUT SCOPE

  • Full improves local readouts relative to w/o Geometry but has lower TM-score and worse distance MAE; FoldBench homology, deposition-date, and fold-level isolation remain incomplete.The independent scaling audit also found template-source and alignment overlaps, so these measurements are diagnostics under the stated conditions.Source: §K.3 · p. 30

L DATA SCALING AND CHECKPOINT DYNAMICS

The scaling and checkpoint analyses vary protein coverage, label density, and training progress, while retaining fixed evaluation contracts. Transfer grows over optimization, but dataset responses differ and the label-density experiment does not establish a positive dose–response relationship.

L DATA SCALING AND CHECKPOINT DYNAMICS

  • Across seven training-set sizes from 50 to 4,000 proteins, every scale uses three epochs and the complete held-out FoldingCorpus, FoldBench334, and General-10 endpoints.The 4,000-protein pool also changes its confidence-quality gate relative to the 2,000-protein pool, so that comparison changes both sample count and quality support.Source: L DATA SCALING AND CHECKPOINT DYNAMICS · p. 30, Table 20 · p. 30

L.4 NUMERICAL CHECKPOINT TRAJECTORIES

  • Full-RG General-10 changes rise from −0.02/+0.03/+0.04 pp at step 5 to +3.15/+2.94/+3.85 pp at step 375; all seeds in both Full-RG and Pure-LoRA are classified as continuous growth.The timing pattern is not specific to the Geometry loss, and gradual format adaptation remains an alternative explanation.Source: L DATA SCALING AND CHECKPOINT DYNAMICS · p. 30, Figure 8 · p. 31

L.1 EVALUATION CONTRACTS AND COMPLETE SCALE-BY-SEED RESULTS

L.2 WHICH DATASETS ACCOUNT FOR DATA-SCALING GAINS?

  • Dataset responses differ with protein coverage: GraphQA Hard and SpatialViz contribute substantial gains toward 2K proteins, ChemBench4K peaks earlier, and ChemBench and Lab-Bench have small negative mean changes at 4K.The macro combines strengthening, saturating, and declining task responses.Source: §L.2 · p. 32, Table 23 · p. 32, Figure 9 · p. 33

L.3 FOLDINGCORPUS-LABEL DENSITY AT FIXED PROTEIN COVERAGE

  • With 1,000 proteins and 375 optimizer steps, using 3, 6, or 12 labels per protein yields General-10 changes of 3.391, 2.951, and 3.311 pp, respectively, showing no monotonic increase.The 95% t interval for the mean within-seed slope is [−0.814, 0.734], permitting positive or negative density effects.Source: §L.3 · p. 33, Table 24 · p. 33

M BROADER IMPACT

  • The paper identifies benchmark contamination and misuse of structural readouts as folding predictions as main risks, and reserves biological interpretation for validated structure-prediction systems.The released models and documentation will identify geometry output as a training diagnostic.Source: M BROADER IMPACT · p. 34

N MACHINE-READABLE EVIDENCE FILES

  • The evidence package preserves run-specific evaluation contracts, splits, and checkpoint distinctions; inspecting exported scores, table cells, and figures requires only the supplementary source package.Rebuilding tables from scratch requires the referenced experiment files.Source: N MACHINE-READABLE EVIDENCE FILES · pp. 34–35, Table 25 · p. 34
Loading 2609.38879v1…