PaperlayerView exact revision on arXiv

Source-linked AI summary

S1-Omni: A Unified Multimodal Reasoning Model for Scientific Understanding, Prediction, and Generation

Jiahao Zhao, Junyi Liu, Lifeng Xu, Nan Xu, Qingli Wang, Qingxiao Li, Tianle Chen, Xiaoyu Wu, Yawen Zheng, Zikai Wang, Guanming Liu, Hequn Zhou, Jingyi Wang, Jingyuan Shu, Keqi Wang, Li He, Songyang Diao, Wenhui Xu, Xinyu Ren, Yaqin Fan, Yujin Zhou, Zhanao Yao

arXiv:2607.15686v1cs.AI

TL;DR

Existing scientific intelligence remains fragmented across modalities, disciplines, and task types. S1-Omni unifies scientific data representation, knowledge alignment, and task-specific decoding in one multimodal reasoning model. Across over 60 benchmarks, it outperforms major closed models on most metrics and matches or surpasses specialists on several benchmarks.

  • Problem

    Existing scientific models remain fragmented across domains and have not unified heterogeneous data, scientific principles, and constraints within one model.

  • Method

    S1-Omni combines unified scientific-data representations, natural-world knowledge alignment, and task-specific decoders for native scientific outputs.

  • Results

    Across over 60 benchmarks, S1-Omni outperforms GPT-5.5 and Gemini-3.1-Pro on most metrics and matches or surpasses specialists on several benchmarks.

  • Takeaways & Limitations

    S1-Omni demonstrates unified scientific understanding and reasoning across modalities and tasks within a single process, with outputs verifiable in native scientific forms.

  • Takeaways & Limitations

    A clear specialist gap remains on tasks governed by three-dimensional structure and biophysical priors, because sequence and semantic representations do not fully encode spatial constraints.

Abstract

from arXiv · show

We present S1-Omni, a unified multimodal reasoning model for scientific understanding, prediction, and generation. AI for Science (AI4S) has advanced significantly through domain-specific models, tool-augmented LLMs, and scientific language models. However, model capabilities remain highly fragmented, limiting the joint modeling of heterogeneous data, scientific laws, and expert knowledge. S1-Omni addresses this gap by consolidating these capabilities into a single, coherent scientific reasoning model. The architecture of S1-Omni is built upon three core components: unified representation of scientific data, natural-world knowledge alignment, and decoding for domain-specific tasks. First, S1-Omni maps natural-language instructions and scientific objects, including CIF, SMILES, protein sequences, spectra, and scientific images, into a shared representation space. Second, it incorporates scientific laws and expert knowledge into data construction and training, enabling the model to reason from scientific evidence. Third, it performs task-specific decoding to support a broad range of applications, including property prediction, spectrum-to-molecular generation, protein site and structure prediction, and scientific image generation and editing. S1-Omni is trained on S1-Omni-Corpus, which covers 200 scientific tasks and contains millions of reasoning samples, and is evaluated on over 60 scientific benchmarks. It outperforms GPT-5.5 and Gemini-3.1-Pro on most benchmarks and matches or surpasses domain-specific models on several benchmarks. Overall, S1-Omni provides a practical path toward unified scientific modeling.

1 Introduction

S1-Omni addresses fragmentation in scientific intelligence with a unified multimodal reasoning model for scientific understanding, prediction, and generation. It combines shared representations, evidence-informed reasoning, and task-specific decoding across scientific modalities and tasks.

  • Motivation: Existing scientific intelligence remains fragmented across domain-specific models, tool-augmented LLMs, and scientific language models.These approaches remain limited by specialization, external-tool dependence, or incomplete cross-disciplinary integration.
  • Unified model: S1-Omni unifies natural-language instructions with scientific objects including CIFs, SMILES, protein sequences, spectra, and scientific images.The model organizes these modalities into a unified task representation.
  • Training data: S1-Omni-Corpus covers 200 scientific tasks and contains millions of high-quality scientific reasoning samples.Its construction uses scientific laws and expert knowledge to design task-specific reasoning chains involving evidence sources and output constraints.
  • Evaluation: Over 60 benchmarks evaluate S1-Omni across molecules, materials, proteins, spectra, and scientific images, including prediction, generation, and editing tasks.The evaluation includes property prediction, spectrum-to-molecular generation, protein functional-site and structure prediction, and scientific image generation and editing.
  • Results: S1-Omni outperforms GPT-5.5 and Gemini-3.1-Pro on most evaluation metrics and matches or surpasses task-specific models on several benchmarks.The reported results support unified scientific understanding and reasoning across modalities and tasks.

2 Related Work

Related work spans domain-specific scientific models, tool-augmented general and multimodal systems, scientific language models, and scientific reasoning models. S1-Omni differs by unifying heterogeneous scientific data representation, natural-world knowledge alignment, and domain-specific decoding in one multimodal reasoning architecture.

  • Domain-specific scientific models: Domain-specific scientific models encode scientific objects, domain inductive biases, and evaluation protocols to achieve strong performance on targeted tasks.Examples include AlphaFold 3 for biomolecular-complex structure prediction and ESM3 for protein sequence, structure, and function generation.
  • Tool-augmented general models: Tool-augmented general models use natural language as a unified interface for perception, generation, and tool use across multimodal systems.Scientific systems such as GPT-Rosalind and Claude Science combine agents, scientific tools, and workflows for literature analysis, data processing, and experimental collaboration.
  • Scientific language models: Scientific language models map diverse scientific objects into shared representation spaces to improve interoperability across heterogeneous scientific data.NatureLM represents molecules, proteins, RNA, and materials as sequences, while LOGOS uses a scientific grammar for objects, spatial constraints, and interactions.
  • Scientific reasoning and knowledge alignment: Scientific reasoning models explicitly incorporate scientific laws and domain knowledge into reasoning for tasks such as protein design and function prediction.Examples include Proteo-R1, BioReason-Pro, and DrugTrail, which introduce reasoning mechanisms, multimodal biological context, or structured reasoning traces.
  • S1-Omni: S1-Omni unifies scientific understanding, prediction, and generation through data representation, natural-world knowledge alignment, and domain-specific decoding.Its framework encodes text, molecules, materials, proteins, spectra, and scientific images within a unified modeling framework.

3 Method

S1-Omni unifies multimodal scientific inputs, evidence-aligned reasoning, and domain-specific decoding within one model process. Its shared representation supports textual responses alongside native outputs such as predictions, structures, and scientific images.

  • Unified architecture: S1-Omni combines unified scientific representations, natural-world knowledge alignment, and task-specific decoding in a single model process.A shared vision-language model produces task-conditioned hidden representations, optional reasoning, textual answers, and task tokens before domain decoding.
  • Unified architecture: Text, CIF, SMILES, proteins, spectra, and scientific images enter a shared vision-language model and feed text, scalar, position-index, coordinate, and image decoders.These decoders produce natural-language responses, property and site predictions, three-dimensional structures, and scientific image generation or editing results.
  • Unified architecture: The shared backbone performs cross-modal task understanding and scientific reasoning, while result decoders preserve numerical, discrete, structural, geometric, and visual output constraints.S1-VL-32B serves as the shared backbone, communicating with task decoders through VLM prefill hidden representations.
  • Knowledge alignment: Natural-world knowledge alignment organizes user intent, scientific evidence, and their relations in the shared representation through evidence selection, scientific relations, and reasoning supervision.The resulting representation supports both textual responses and domain-native result decoding.
  • Domain decoding: Task-specific extraction adapts the shared condition to each domain: property tasks use pooled features, protein-site prediction retains token-level features, and spectrum generation projects into molecular-generation space.Protein-site prediction additionally fuses ESM2 residue features, while spectrum-to-molecular generation reads SpecFormer spectral features.
  • Training data: More than 8 million filtered reasoning and supervision records span over 200 research tasks across chemistry, biology, materials, physics, and medical imaging.The corpus includes general text and multimodal tasks together with domain-native targets such as values, labels, residue annotations, structures, coordinates, and images.

4 Evaluations

S1-Omni achieves broad cross-domain performance, surpassing stronger closed models on most molecular, quantum-chemical, and materials properties while specialist models retain advantages on several structure-sensitive tasks. Ablations show that structured reasoning, question-only pooling, data scaling, and end-to-end adaptation materially affect performance, whereas explicit range prediction is ineffective.

  • Cross-domain evaluation: S1-Omni exceeds the stronger closed model on all seven MoleculeNet tasks, 16 of 18 ADMET endpoints, and 40 of 48 materials and related properties.It exceeds ChemFM-3B on BBBP at 0.8578 ROC-AUC and FreeSolv at 0.7092 RMSE, while specialists remain stronger on several endpoints.
  • Cross-domain evaluation: Specialist models retain substantially lower error on quantum-chemical properties and at least one specialist outperforms S1-Omni on each of the 12 material properties.S1-Omni’s COMPAS-1D MAEs range from 0.0508 to 0.0699 versus 0.0067 to 0.0170 for ChemFM-3B, while its EPS_DFPT MAD/MAE is 1.5559 versus 1.5780 for LLM-Prop-35M.
  • Reasoning and representations: Structured reasoning lowers material MAE to 17.255, compared with 34.818 for free-form reasoning and 20.457 without reasoning text.The benefit depends on organizing scientific evidence consistently around the target property; arbitrary reasoning text does not improve prediction by itself.
  • Reasoning and representations: Question-only pooling improves every metric, raising chemical F1 from 0.773 to 0.822, lowering chemical RMSE from 1.663 to 0.908, and raising material MAD/MAE from 0.899 to 1.454.This indicates that property information can be extracted before answer generation, with reasoning supervision shaping the prefill representation.
  • Ablations and limitations: Explicit range prediction leaves RMSE essentially unchanged at 0.2812 with range prediction versus 0.2814 without it.The model does not reliably predict the target range, so explicit range prediction is not an effective intermediate reasoning signal.
  • Scaling and adaptation: Increasing Stage-2 data from 27k to 140k raises chemical F1 from 0.822 to 0.874 and material MAD/MAE from 1.454 to 1.555, while Stage 3 further improves the metrics.At 140k records, Stage 3 lowers RMSE from 0.855 to 0.790 and raises material MAD/MAE from 1.525 to 1.865.

5 Discussion

Reasoning supervision improves representations for property prediction, protein structure prediction, and scientific image generation, but its effectiveness depends decisively on how reasoning data are constructed. On six material regression properties, structured, property-constrained supervision substantially improves regression performance over unconstrained free-form reasoning.

  • Reasoning supervision: Reasoning supervision improves representations supplied to downstream modules across property prediction, protein structure prediction, and scientific image generation.The discussion identifies reasoning supervision as beneficial across all three application areas.
  • Reasoning-data construction: 17.255 MAE is achieved with structured, property-constrained supervision, versus 34.818 with unconstrained free-form reasoning on six material regression properties.This comparison is reported in Table 6.
  • Reasoning-data construction: 1.260 MAD/MAE is achieved with structured, property-constrained supervision, compared with 0.888 under unconstrained free-form reasoning.The reported metric increase accompanies the reduction in MAE.

Appendix Sections A.1.4 and A.1.5 illustrate the distinction on the same Lu2ZnPd crystal and Spillage

The appendix illustrates that property-constrained reasoning improves training signals by stabilizing evidence coverage and causal ordering, while reasoning-conditioned shared representations support heterogeneous scientific decoders. Controlled analyses further attribute meaningful benefits to the shared backbone rather than task experts alone.

  • Reasoning traces: Property-constrained traces reduce nuisance variation in training compared with fluent unconstrained rationales, which can underperform omitting reasoning text.The structured trace standardizes evidence selection and explicitly states competing property-relevant directions.
  • Reasoning traces: Reasoning supervision teaches shared backbone states that remain useful to numerical predictors, molecular and protein generators, and image decoders without requiring generated reasoning text downstream.Image generation benefits when explicit planning decomposes instructions into visual units and specifies their layout and connections.
  • Feature extraction: Question-only pooling improves property classification and regression over pooling reasoning and answer tokens, while task information is largely present before answer generation.Long reasoning and final-answer states can mix scientific evidence with linguistic organization, autoregressive history, and formatting.
  • Backbone versus experts: Controlled reuse of strong expert decoders shows that shared-backbone conditioning improves spectrum-to-molecular generation and modestly improves protein-folding metrics.These experiments change only the conditioning representation supplied by the shared backbone while holding task-specific decoders fixed.
  • Backbone versus experts: The controlled results support attributing a meaningful fraction of S1-Omni’s benefit to backbone representations that enhance downstream experts rather than to task decoders alone.The architecture nevertheless passes conditions predominantly one way, leaving deeper bidirectional fusion as an open direction.

6 Limitations

S1-Omni is an initial step toward scalable unified multimodal scientific reasoning, but its current design and evaluation leave important limitations in pretraining scale, native generation, task coverage, and long-horizon capabilities.

  • Overall scope: The present work tests feasibility of linking heterogeneous scientific data, shared task representations, and domain-native results, but limitations become more important in larger and more open scientific settings.The authors characterize S1-Omni as an initial step toward scalable unified multimodal reasoning.
  • Large-scale scientific pretraining: S1-Omni has not undergone large-scale scientific pretraining, limiting conclusions about scaling with data, parameters, or compute.It starts from an existing vision-language model and trains on curated scientific tasks.
  • Modular generation across scientific output spaces: Molecular structures, protein coordinates, and scientific images rely on diffusion or specialist decoders rather than native generation by the shared backbone.This preserves geometric accuracy and generation quality but leaves the model between unified task representation and unified native generation.
  • Task coverage: Current evaluation covers initial scientific test beds but lacks sufficient disciplinary depth, complex multistage problems, out-of-distribution generalization, and systematic cross-task transfer.Broader tasks and transfer studies are needed to test composition of existing capabilities.
  • General capabilities and long-horizon tasks: Scientific task training reduces performance on some general tasks relative to the base model, while evaluation focuses on bounded single-turn problems.Multi-turn planning, cross-modal interaction, long-horizon state tracking, tool use, and correction across steps remain underexplored.

7 Conclusion

S1-Omni is a unified multimodal reasoning model for scientific understanding, prediction, and generation, trained on S1-Omni-Corpus. It combines unified scientific-data representation, natural-world knowledge alignment, and task-specific decoding to connect heterogeneous inputs, evidence-grounded reasoning, and scientific outputs.

  • S1-Omni unifies scientific understanding, prediction, and generation within a single multimodal reasoning model trained on S1-Omni-Corpus.
  • Its architecture combines unified scientific-data representation, natural-world knowledge alignment, and decoding for domain-specific tasks.
  • These components link heterogeneous scientific inputs, evidence-grounded reasoning, and native scientific outputs, offering a scalable path toward unified multimodal reasoning for science.

A Appendix

The appendix presents thirteen complete reasoning examples, targeted visual analyses, representative failure modes, and a stability evaluation after extended general-purpose conversation. It retains the original language of every interaction.

  • Reasoning examples: The appendix presents thirteen complete reasoning examples ordered by task family, matching Section 4.The examples use the same task-family order as Section 4.
  • Visual analyses: It analyzes scientific image generation and editing through targeted visual analyses.
  • Robustness analysis: It examines representative failure modes and tests whether scientific prediction remains stable after a long, general-purpose multi-turn conversation.
  • Interaction language: The appendix retains the original language of every interaction.

A.1 Reasoning Examples

The reasoning examples present complete inputs, model responses, and task tokens, including native numerical, structural, or image outputs when available. Two controlled reasoning-supervision examples lack native scalar outputs, while Table 31 catalogs cases by evaluation task family and route token.

  • Example format: Each example includes the full user input, complete model response, and emitted task token.Examples begin on separate pages.
  • Example format: Native numerical, structural, or image outputs appear when available in the source record.The examples do not infer unavailable native outputs.
  • Output availability: Two controlled reasoning-supervision examples report no native scalar output.Accordingly, no scalar output is inferred for them.
  • Case catalogue: Table 31 catalogs cases by evaluation task family, with route tokens selecting the corresponding native result interface.The catalogue is ordered according to the task families used in the Evaluation section.

A.1.1 Crystal Property Prediction: DFPT Dielectric Constant

The section predicts the DFPT total dielectric constant of triclinic Pr2AlNO3, linking its dielectric response to Pr-driven polarizability, low symmetry, and dense crystal packing. The material is characterized as a potential dielectric with moderately high dielectric behavior.

  • Crystal structure: Pr2AlNO3 crystallizes in the triclinic P 1 space group, with a 3.752 Å, b 3.752 Å, c 6.875 Å, and unit-cell volume 89.298 Å3.The unit cell contains one formula unit and seven atoms, with minimal symmetry constraints on atomic positions.
  • Dielectric-response drivers: Pr’s high polarizability and f-electrons are expected to enhance electronic and ionic contributions to the dielectric response.The presence of Pr is identified as a key structural feature influencing the total dielectric constant.
  • Prediction target: The target property is the dimensionless total dielectric constant ε calculated using density functional perturbation theory (DFPT).Its response includes joint contributions from low-frequency polarization and electronic-cloud distortion.
  • Dielectric-response drivers: Low triclinic symmetry permits more lattice-relaxation modes, while dense atomic packing limits the likelihood of an extremely low dielectric response.Together, these factors support a moderately high dielectric response and motivate consideration of Pr2AlNO3 as a potential dielectric material.

Native Output · A.1.2 Crystal Property Prediction: Exfoliation Energy

This section estimates the exfoliation energy of Ta2CS2 from its CIF crystal structure. The analysis connects the material’s geometry and bonding environment to its suitability for isolating single layers.

  • A.1.2 Crystal Property Prediction: Exfoliation Energy: The task takes a crystal structure supplied as a CIF file and requests exfoliation energy in meV/atom.The target metric is explicitly labeled Exf. En.
  • A.1.2 Crystal Property Prediction: Exfoliation Energy: The CIF identifies the material as Ta2CS2 with triclinic P 1 symmetry and lattice parameters a=b=c=8.833 Å.The unit-cell metadata gives the structural formula and equal cell lengths.
  • A.1.2 Crystal Property Prediction: Exfoliation Energy: All three unit-cell angles are acute at 21.505°, producing a highly symmetric geometric configuration despite the low-symmetry P 1 label.The structure is described as potentially suggesting rhombohedral stacking associated with layered transition-metal compounds.
  • A.1.2 Crystal Property Prediction: Exfoliation Energy: The unit cell contains five atoms: two tantalum, one carbon, and two sulfur atoms.This composition corresponds to the Ta2CS2 structural formula in the CIF metadata.
  • A.1.2 Crystal Property Prediction: Exfoliation Energy: Exfoliation energy measures the feasibility of isolating single layers from a bulk crystal for nanotechnology applications.It is presented as a key metric for layered materials.
  • A.1.2 Crystal Property Prediction: Exfoliation Energy: The property reflects weak van der Waals interactions between sulfur-terminated layers balanced against strong covalent Ta-C bonding within the slabs.This bonding contrast determines the interlayer-versus-intralayer energy balance relevant to exfoliation.

Native Output

The model predicts an exfoliation energy of 49.97 meV/atom.

  • 49.97 meV/atom is the predicted exfoliation energy.

A.1.3 Crystal Property Prediction: Topological Spillage User Input

The section predicts VSe2’s dimensionless Spillage from its layered crystal structure and electronic factors. It concludes that Spillage should be low but nonzero, reflecting weak or nontrivial topology rather than a strong topological-insulator state.

  • Crystal structure: VSe2 is a layered transition-metal dichalcogenide with a=b=3.372 Å, c=6.201 Å, γ=120°, and unit-cell volume 61.070 Å3.The CIF identifies composition V1 Se2 and space group P1.
  • Physical determinants: Spillage is primarily controlled by Se spin–orbit coupling and hybridization between V 3d and Se 4p orbitals.Se contributes moderate SOC, while orbital hybridization influences band dispersion and possible band inversion.
  • Physical determinants: The layered coordination environment may induce slight band inversion, but metallic or semimetallic behavior limits Spillage and P1 symmetry may lift degeneracies.The absence of a bulk gap distinguishes VSe2 from a typical topological insulator and constrains the upper bound of Spillage.
  • Prediction: Spillage is expected to be much less than 1 but nonzero, indicating weak or nontrivial topological character rather than a strong topological insulator.The qualitative prediction combines Se SOC, VSe2’s electronic structure, elemental composition, and structural symmetry.

Decoder Output … Decoder Output

The decoder outputs span scientific property prediction, molecular generation, biochemical classification, DNA-binding-site prediction, and failure analyses. Results combine mechanistic reasoning with explicit outputs, while highlighted failures show that misread evidence and unsupported assumptions can distort downstream predictions.

  • A.1.4 Reasoning-Supervision Comparison: Unconstrained Free-Form Reasoning: 0.05 to 0.25 is the estimated Spillage interval for Lu2ZnPd, reflecting limited wavefunction mixing from closed-shell Lu and a highly ordered lattice.The effective Spillage prediction range is [0, 4].
  • A.1.5 Reasoning-Supervision Comparison: Structured, Property-Constrained Reasoning: Structured reasoning constrains the teacher to quantify lattice, composition, coordination, topology, distortion, and symmetry before selecting three to five property-relevant factors.The Stage-2 predictor, training scale, optimization, and question-only pooling remain fixed in the comparison.
  • A.1.6 Molecular Property Prediction: ESOL Aqueous Solubility: CCCCCC is identified as n-pentane, whose nonpolar hydrocarbon structure lacks heteroatoms, hydrogen-bonding capacity, and ionizable groups, implying poor aqueous solubility.The decoder output is introduced as a predicted aqueous-solubility result, but no numerical log S value is supplied.
  • A.1.7 CYP2C9 Substrate Classification: The CYP2C9 assessment balances hydrophobic and potentially oxidizable features against rigidity, electron-poor quinazoline, fluorine-induced deactivation, and steric protection.Potential metabolic sites include methoxy O-demethylation, benzylic oxidation, and aromatic epoxidation.
  • A.1.8 CYP2D6 Inhibition Classification: The molecule is predicted unlikely to inhibit CYP2D6 because it lacks a protonatable basic amine and instead has an anionic carboxylate and high polarity.These features are inconsistent with the ionic and hydrophobic interactions needed for CYP2D6 active-site engagement.
  • A.1.9 Infrared Spectrum-to-Molecular Generation: The infrared interpretation supports a saturated cyclopropane-containing nitrile structure from sp3 C–H absorptions, strained-ring features above 3000 cm−1, and a diagnostic nitrile stretch.The absence of a C=C band favors cyclopropane over alkene or aromatic assignments.
  • A.1.10 DNA-Binding-Site Prediction: ArcA: ArcA DNA-binding-site prediction identifies a basic N-terminal patch and assigns helix-recognition and wing-region residues to phosphate-backbone, major-groove, and minor-groove contacts.The listed basic patch is R6, R8, R9, R13, R14, K16, R19.
  • A.3 Failure Mode Analysis: Failure analyses show that incorrect peak assignment can produce a hydrocarbon cyclopropane prediction instead of an oxygen-containing ground truth, while unsupported sequence claims can bias structure prediction.One protein response reported 255 residues and a specific HAD-IIB classification, whereas the input contains 254 residues and lacked sufficient evidence for those claims.
Loading 2607.15686v1…