Pith. sign in

REVIEW 4 major objections 5 minor 32 references

S1-Omni claims a single multimodal model can unify scientific understanding, prediction, and generation, reporting wins over leading closed general-purpose systems on most of 60-plus benchmarks and parity or better against several domain-sp

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 22:34 UTC pith:5F6PHQUG

load-bearing objection A serious system paper with one broken headline: the closed-model baselines are implausible on their face, but the architecture, ablations, and several specialist comparisons deserve referee time. the 4 major comments →

arxiv 2607.15686 v1 pith:5F6PHQUG submitted 2026-07-17 cs.AI

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
show 14 more authors
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
This is my paper
classification cs.AI
keywords unified multimodal reasoningscientific AIproperty predictionspectrum-to-molecule generationprotein functional-site predictionprotein structure predictionscientific image editingreasoning supervision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

S1-Omni is the paper's claim that scientific AI can be consolidated rather than fragmented: one model takes text instructions together with molecules (SMILES), crystals (CIF), protein sequences, spectra, and scientific images, maps them into a shared representation, reasons from scientific laws and expert-structured evidence, and then routes to small task-specific decoders that output numbers, residue indices, molecular graphs, three-dimensional coordinates, or images. On more than 60 benchmarks the authors report that this unified model beats two leading closed general-purpose multimodal systems on most tasks—95.5% and 83.3% win rates in aggregate comparisons—and matches or surpasses several domain-specific specialists on drug-property, protein-site, spectrum-to-molecule, and scientific image-editing benchmarks. The sympathetic reading: a shared backbone can transfer knowledge across scientific disciplines and output spaces, with specialization pushed to the final decoder rather than to separate models. If this holds, it is a practical path toward one scientific model instead of a zoo of special-purpose systems.

Core claim

On its own terms, the paper establishes a modeling architecture for unified scientific intelligence: a vision-language backbone forms task-conditioned hidden states from an instruction and a native scientific object; structured reasoning supervision ('natural-world knowledge alignment') shapes those states so they encode evidence about composition, symmetry, contacts, peaks, or spatial relations; and a task token selects a decoder that produces an output verifiable in the domain's native form. The paper's empirical claim is that this single model outperforms leading closed general-purpose models on most of 66 evaluated tasks and matches or exceeds specialist models on a meaningful subset—not

What carries the argument

The central mechanism is the pairing of a shared VLM backbone with modular task decoders connected through prefill hidden states. A task token selects the decoder; the decoder reads a pooled or token-level condition extracted from the prefill representation, sometimes fused with auxiliary native input (pretrained protein-residue features, spectral features, input image), and emits a continuous value, residue probabilities, a molecular graph/conformation, 3D coordinates, or an image. Structured scientific reasoning is the training-side machinery that shapes these prefill states; native targets are kept out of the visible reasoning text and supervised through a separate path.

Load-bearing premise

The load-bearing premise is the cleanliness of S1-Omni-Corpus: the paper states that deduplication, benchmark test-set overlap filtering, and train/validation/test partitioning were done before data augmentation or reasoning distillation, but only a 10K subset is released, so if near-duplicates of test instances or reasoning traces that implicitly encode native targets slipped in, every reported benchmark advantage could be inflated.

What would settle it

An independent audit comparing the released corpus subset and, ideally, the full corpus against all benchmark test sets for near-duplicate molecular structures, sequences, CIFs, and images, followed by retraining on a cleaned corpus; if benchmark scores drop materially after removing overlaps, the unified-model claim as measured is not supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single frozen backbone plus trainable decoders can cover property prediction, molecule generation, structure prediction, and image editing without external tools, so scientific-task pipelines can shrink from many specialist models to one shared model with plug-in output heads.
  • Structured, property-constrained reasoning supervision is a representation-learning device: it improves downstream predictions even when the decoder never consumes the reasoning text, whereas unconstrained free-form reasoning can be worse than no reasoning.
  • Prefill hidden states (formed before answer generation) carry the task information needed by decoders; pooling answer-side or reasoning states dilutes it, which makes the prefill interface a reusable design pattern for future unified models.
  • Uniform benchmark averages can hide failure modes: on out-of-distribution QM9 properties the model shows low RMSE yet negative tail correlation on electronic-structure targets, so distribution-shift evaluation needs tail statistics, not just mean error.
  • Specialist gaps remain where three-dimensional structure and biophysical priors matter—quantum-chemical precision, materials properties, and DNA/RNA/metal-site recognition—so the unified model replaces general-purpose workflows but not all domain experts yet.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the corpus cleanliness holds, the most direct extension is adding new decoders (reaction prediction, experimental materials synthesis, instrument-to-report generation) to the same frozen backbone, since the paper shows decoders can be trained independently on native spaces.
  • Editorial inference: the negative tail-correlation results imply that for safety-critical scientific uses (drug toxicity, band-gap screening), publishing OOD/ID ratios and binned tail R2 should become a reporting norm, because aggregate win rates overstate extrapolation ability.
  • Editorial inference: the ablation showing explicit range prediction in reasoning does not help suggests future reasoning-supervision research should target evidence decomposition rather than intermediate output quantities, a testable design choice for other scientific LLMs.
  • Editorial inference: the architecture's success on segmentation-as-editing and spectrum-to-molecule reconstruction hints that a unified interface between language reasoning and generative decoders could generalize to other inverse problems (e.g., experimental spectra to crystal structures) where evidence is distributed along the input.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents S1-Omni, a unified multimodal reasoning model for scientific understanding, prediction, and generation. The architecture couples a shared vision-language backbone (S1-VL-32B) with task-specific decoders for property prediction, spectrum-to-molecule generation, protein-site prediction, protein structure prediction, and scientific image generation/editing. The model is trained on S1-Omni-Corpus, claimed to cover 200 tasks and millions of reasoning samples, and is evaluated on over 60 benchmarks. The central claim is that this single model outperforms GPT-5.5 and Gemini-3.1-Pro on most benchmarks and matches or surpasses domain-specific models on several tasks. The paper also presents controlled ablations on reasoning supervision, hidden-state pooling, predictor sharing, and frozen-decoder reuse.

Significance. If the results hold, the architecture demonstrates a useful design pattern: a shared reasoning backbone with native output decoders can handle heterogeneous scientific tasks without per-task fine-tuning of the entire stack. The controlled ablations are genuinely informative, especially the pooling-scope and reasoning-supervision comparisons, and the reuse of DiffSpectra and SimpleFold decoders isolates the backbone contribution. The release of weights, code, and a 10K corpus subset supports reproducibility. However, the headline comparison to GPT-5.5 and Gemini-3.1-Pro is severely undermined by implausible baseline numbers, and the leakage-prevention claim for the full corpus is not verifiable. The specialist comparisons and ablations are the most credible part of the paper and could stand after the closed-model claims are corrected or removed.

major comments (4)
  1. [§4.1.2, Tables 3–5] The closed-model baseline scores are not credible. GPT-5.5 achieves ESOL RMSE 68.25 on a logS scale whose typical range is roughly [-12, 2], and Gemini-3.1-Pro achieves LIPO RMSE 980.16 on a logP range of approximately [-2, 6]; GPT-5.5 QM9 GAP MAE is 4.678 eV for a property spanning roughly 0–10 eV. These magnitudes indicate that free-text outputs were not parsed and scored under a consistent protocol. The paper provides no prompts, sampling parameters, or parse-and-score code for these baselines, so the abstract's 'outperforms GPT-5.5 and Gemini-3.1-Pro' claim is unsupported as stated. This issue is load-bearing for the headline result; it must be fixed by supplying a verified baseline protocol or by removing the closed-model comparison from the main claims and keeping only the internally consistent specialist comparisons.
  2. [§3.2.2 and §1] The claim that S1-Omni-Corpus is leakage-free is load-bearing for every benchmark score in Section 4, but it is unverifiable from the manuscript. Section 3.2.2 states that 'deduplication, benchmark test-set overlap filtering, and train/validation/test partitioning are completed before data augmentation or reasoning distillation,' but only a 10K subset is released and no algorithm, hash/embedding threshold, or code for overlap filtering is provided. If near-duplicate molecules, crystals, or protein sequences from test sets entered training, all property and protein-site results could be inflated. The paper should release the overlap-detection protocol and split metadata, or clearly caveat the evaluation as conditional on an external audit.
  3. [§4.6] The aggregate win rates mix non-comparable comparators. The text says image-output tasks use GPT-Image-2 and Nano Banana 2, not GPT-5.5 or Gemini-3.1-Pro, yet Figure 6 and the surrounding text report '95.5% against the GPT model family' and '83.3% against the Gemini model family.' A single model per family is not used consistently across all 66 tasks, so the aggregated win rates conflate different systems. Provide a per-task comparator breakdown and compute the win rates separately for text/output tasks and image tasks, or the aggregate claim should be withdrawn.
  4. [§4.4.2, Table 27; §5.3] The protein-structure result is over-stated. Relative to unconditioned SimpleFold-700M, S1-Omni improves mean TM-score by 0.000366 and median RMSD by 0.126, while mean GDT-TS, lDDT, and lDDT-Ca change by less than 0.002 and mean RMSD worsens from 4.557 to 4.591. The paper acknowledges this is modest, but Section 5.3 concludes that 'a meaningful fraction of S1-Omni's benefit comes from backbone representations.' At these magnitudes, the effect is within likely noise, and the controlled comparison does not establish a meaningful folding benefit. Either provide per-target error bars or statistical significance, or soften the claim to say the conditioning signal is compatible with the geometric decoder without claiming substantial improvement.
minor comments (5)
  1. [Table 3] The dataset label 'FREESOLV' should be 'FreeSolv' for consistency with the text and standard naming.
  2. [Table 16] The entries '1.48E-11.21E-2' and '1.05E+0' are ambiguously formatted; add a space or use consistent scientific notation to avoid confusion.
  3. [Abstract vs §4.6] The abstract says 'over 60 scientific benchmarks,' while Section 4.6 says '66 tasks across six categories.' Clarify whether the count is benchmarks or tasks, and use consistent terminology.
  4. [Appendix A.1.10] The reasoning trace lists predicted DNA-binding residues such as R6, R8, R9, R13, R14, K16, R19, T32, H33, Q34, R49, Y53, Y54, H55, K59, but the native decoder output lists a different set (V3, P6, R7, S39, T40, H41, R50, A52, S53, Y55, Y56, Q117). This mismatch between the visible reasoning and the native output should be explained or aligned, otherwise it weakens the claim of evidence-grounded prediction.
  5. [§4.6] When reporting the aggregate comparison, state explicitly that image benchmarks are evaluated with GPT-Image-2 and Nano Banana 2, not GPT-5.5/Gemini-3.1-Pro, so readers are not misled by the family-level labels.

Circularity Check

0 steps flagged

No significant circularity: the central claims are benchmark evaluations against external datasets; self-citations are component dependencies, not load-bearing justifications.

full rationale

I walked the paper's claimed derivation chain. The main results are reported scores on external benchmarks (MoleculeNet, TDC ADMET, QM9, JARVIS-DFT, MPBind, CAMEO22, MSD, etc.), obtained by training S1-Omni and then evaluating on held-out test sets. These scores are not derived from fitted constants or from the training labels by construction. The paper explicitly separates native targets from visible reasoning traces: 'The native target y_task is stored separately from the visible context to prevent result information from entering the reasoning input' and 'Domain-native targets are never written into visible reasoning.' Whether the claimed deduplication and overlap filtering were actually effective is a data-hygiene and evaluation-integrity risk, not a circular reduction. The self-citations to S1-VL and S1-Omni-Image are used to identify the backbone and image decoder components; they do not by themselves justify the claim that S1-Omni outperforms GPT-5.5 or Gemini-3.1-Pro, which rests on the paper's own evaluations. The ablations (e.g., Tables 6-11, 26-28) systematically vary one factor while holding others fixed and do not rename a training target as a prediction. The implausibly extreme GPT-5.5/Gemini baseline values (e.g., LIPO RMSE 980.16, QM9 GAP MAE 4.678 eV) are a serious evaluation-protocol concern that undermines the 'outperforms' claim, but they do not make the paper circular. The Limitations section also explicitly concedes 'S1-Omni has not undergone large-scale scientific pretraining' and that molecular structures, protein coordinates, and images 'rely on diffusion or specialist decoders,' further framing the paper as a feasibility study rather than a self-justifying framework. No step in the derivation reduces by definition or by self-citation to its own input.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central results are empirical and rest on the training corpus, benchmark labels, and evaluation protocols; no new physical entities or hand-fit scalar constants are introduced. The main unverified premises are data integrity, fairness of the closed-model baselines, and the validity of simulated spectra as a proxy.

axioms (5)
  • domain assumption Benchmark labels and splits (QM9S, MPBind, RoBep, CAMEO22, JARVIS, TDC ADMET, etc.) are reliable ground truth and are used as reported.
    All conclusions in Section 4 depend on the correctness and comparability of external benchmark labels and published specialist scores; the paper does not independently validate these labels.
  • ad hoc to paper S1-Omni-Corpus overlap filtering prevents test-set leakage from benchmarks into training.
    Section 3.2.2 asserts deduplication and test-set overlap filtering but provides no verifiable details or tooling; only a 10K corpus subset is released, so this critical premise cannot be checked.
  • ad hoc to paper The Qwen3.7-distilled reasoning traces and expert-constructed rationales are scientifically valid and do not encode the native target y_task.
    Section 3.2.2 states native targets are stored separately and never introduced into rationales, but the generation and validation of millions of reasoning traces is not auditable from the paper.
  • domain assumption QM9S simulated IR/Raman/UV-Vis spectra are a valid proxy for real spectrum-to-molecule generation.
    Section 4.2 evaluates only on simulated QM9S spectra; transfer to experimental spectra is not tested.
  • domain assumption The GPT-5.5 and Gemini-3.1-Pro scores were produced under a fair prompting and output-parsing protocol.
    Implausible reported values such as GPT-5.5 ESOL RMSE=68.25 and Gemini-3.1-Pro LIPO RMSE=980.16 imply the protocol may not have captured valid numeric outputs, which would invalidate the headline comparisons.

pith-pipeline@v1.3.0-alltime-deepseek · 53656 in / 12377 out tokens · 111404 ms · 2026-08-01T22:34:04.212895+00:00 · methodology

0 comments
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.15686 by Guanming Liu, Hequn Zhou, Jiahao Zhao, Jingyi Wang, Jingyuan Shu, Junyi Liu, Keqi Wang, Lifeng Xu, Li He, Nan Xu, QingLi Wang, Qingxiao Li, Songyang Diao, Tianle Chen, Wenhui Xu, Xiaoyu Wu, Xinyu Ren, Yaqin Fan, Yawen Zheng, Yujin Zhou, Zhanao Yao, Zikai Wang.

Figure 1
Figure 1. Figure 1: Unified architecture of S1-Omni. Text, CIF, SMILES, proteins, spectra, and scientific images are first processed by the shared vision-language model. Its hidden representations are then converted by text, scalar, position￾index, coordinate, and image decoders into natural-language responses, property predictions, site predictions, three-dimensional structures, and scientific image generation or editing res… view at source ↗
Figure 2
Figure 2. Figure 2: Discipline, task, language, and dialogue-turn distributions in S1-Omni-Corpus. These statistics characterize the coverage of the training data rather than define model capability directly by task count. reasoning, the answer, and the task token, whereas y task i contains a continuous value, class label, residue annotation, molecular structure, three-dimensional coordinates, or target image. Domain-native t… view at source ↗
Figure 3
Figure 3. Figure 3: Data construction for natural-world knowledge alignment. Scientific laws and expert knowledge jointly constrain evidence dimensions, reasoning relations, task protocols, and sample validation. partitioning are completed before data augmentation or reasoning distillation to prevent data leakage, ensuring that benchmark test instances or their near-duplicates do not enter the training corpus. Each raw record… view at source ↗
Figure 4
Figure 4. Figure 4: OOD-to-ID RMSE ratios for GotenNet, Regression Transformer (RT), and S1-Omni on eight QM9 properties under BOOM label shift. Values closer to 1 indicate less relative error growth; dashes denote catastrophic RT failures excluded from the ratio scale. Tail correlation. Binned R 2 directly tests whether a model preserves variation within the extreme OOD regions ( [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicted versus reference values for S1-Omni on ID and OOD splits of eight QM9 properties. Panel titles report the OOD/ID RMSE ratio; dashed lines show ideal agreement. 4.2 Spectrum-to-Molecular Generation 4.2.1 Experimental Setup The Spectrum-to-molecular generation infers a molecular structure from IR, Raman, and UV-Vis spectra. The output includes both a two-dimensional molecular graph and a 3D conform… view at source ↗
Figure 6
Figure 6. Figure 6: Aggregate task-level comparison with general-purpose model families under the shared text-and-image input protocol. The outcome strips summarize wins and losses for S1-Omni, and the lower strip shows the composition of the six scientific task categories. requirements. 4.6 Aggregate Comparison with General-Purpose Models To consolidate the general-purpose model comparisons reported throughout Section 4, we … view at source ↗
Figure 7
Figure 7. Figure 7: Native three-dimensional molecular output reconstructed from the supplied infrared spectrum. 49 [PITH_FULL_IMAGE:figures/full_fig_p049_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Native site-prediction output for ArcA. The twelve site-head positions are highlighted in red on chain A of PDB 5DY0. 51 [PITH_FULL_IMAGE:figures/full_fig_p051_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Native SimpleFold-style structure output for GPR168, containing 3,008 atoms. 53 [PITH_FULL_IMAGE:figures/full_fig_p053_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Native generated output for the chemical-synapse illustration request. 54 [PITH_FULL_IMAGE:figures/full_fig_p054_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Native segmentation-as-editing output. The complete input image appears in the user request above. 55 [PITH_FULL_IMAGE:figures/full_fig_p055_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison of S1-Omni, Qwen-Image, GLM-Image, Nano Banana 2, and GPT-Image-2 on neuron classification, maternal care in wildlife, river-network mapping, and lignan biosynthesis. The cases expose differences in scientific organization, visual completeness, and text rendering. 56 [PITH_FULL_IMAGE:figures/full_fig_p056_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Effect of think-before-generate planning on a scientific workflow diagram. The intermediate reasoning state decomposes the requested modules, spatial relations, labels, and information flow before conditioning the image decoder. 57 [PITH_FULL_IMAGE:figures/full_fig_p057_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Representative scientific image editing cases using the shared <image_edit> route: spleen segmentation with a spatial overlay, CBCT-to-CT medical image translation, and MRI super-resolution. Each column shows the instruction, reasoning and answer, and generated result. 58 [PITH_FULL_IMAGE:figures/full_fig_p058_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Raman spectrum used by the failed reconstruction example. Model Response <think> The user needs me to predict the possible molecular structure based on the given Raman spectrum; I will first analyse the Raman spectrum. I observe the provided Raman spectrum, focusing first on the high wavenumber region where the most intense signals are located. There is a prominent cluster of sharp peaks between 2800 and … view at source ↗
Figure 16
Figure 16. Figure 16: Representative text-rendering failures in generated scientific infographics about Pascal’s law and a 3J1 alloy composition chart. Enlarged crops expose malformed Chinese and English glyphs, misspelled technical labels, and semantic substitutions. Observed Failure The global composition and scientific visual language are plausible, but locally rendered text is unreliable. Character substitutions, missing s… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 16 linked inside Pith

  1. [4]

    RPRRSAPRRAGKNP

    For stable intermetallics with heavy elements but closed f-shells and high structural order, values typically cluster in the lower quartile. Given the specific combination of moderate SOC active species (Pd) and structural rigidity, the Spillage is expected to be small but distinct from zero. A reasonable prediction interval would be between 0.05 and 0.25...

    2026

  2. [7]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811,

    Pith/arXiv arXiv

  3. [8]

    Bioreason-pro: Advancing protein function prediction with multimodal biological reasoning.bioRxiv, pp

    Adibvafa Fallahpour, Arman Seyed-Ahmadi, Parsa Idehpour, Omar Ibrahim, Purav Gupta, Jack Naimer, Kevin Zhu, Arnav Shah, Shihao Ma, Abhinav Adduri, et al. Bioreason-pro: Advancing protein function prediction with multimodal biological reasoning.bioRxiv, pp. 2026–03,

    2026

  4. [9]

    Seed-x: Multimodal models with unified multi-granularity comprehension and generation.arXiv preprint arXiv:2404.14396,

    Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation.arXiv preprint arXiv:2404.14396,

    Pith/arXiv arXiv

  5. [11]

    Team Glm, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al

    URLhttps://arxiv.org/abs/2403.05530. Team Glm, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793,

    Pith/arXiv arXiv

  6. [12]

    Learning joint 2d & 3d diffusion models for complete molecule generation.arXiv preprint arXiv:2305.12347,

    32 Han Huang, Leilei Sun, Bowen Du, and Weifeng Lv. Learning joint 2d & 3d diffusion models for complete molecule generation.arXiv preprint arXiv:2305.12347,

    Pith/arXiv arXiv

  7. [13]

    Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development.arXiv preprint arXiv:2102.09548,

    Kexin Huang, Tianfan Fu, Wenhao Gao, Yue Zhao, Yusuf Roohani, Jure Leskovec, Connor W Coley, Cao Xiao, Jimeng Sun, and Marinka Zitnik. Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development.arXiv preprint arXiv:2102.09548,

    Pith/arXiv arXiv

  8. [15]

    Bowen Jing, Ezra Erives, Peter Pao-Huang, Gabriele Corso, Bonnie Berger, and Tommi Jaakkola

    URL https://arxiv.org/abs/2406.14969. Bowen Jing, Ezra Erives, Peter Pao-Huang, Gabriele Corso, Bonnie Berger, and Tommi Jaakkola. Eigenfold: Generative protein structure prediction with diffusion models.arXiv preprint arXiv:2304.02198,

    Pith/arXiv arXiv

  9. [16]

    Alphafold meets flow matching for generating protein ensembles.arXiv preprint arXiv:2402.04845,

    Bowen Jing, Bonnie Berger, and Tommi Jaakkola. Alphafold meets flow matching for generating protein ensembles.arXiv preprint arXiv:2402.04845,

    Pith/arXiv arXiv

  10. [18]

    Drugtrail: Interpretable drug discovery via structured reasoning and druggability-tailored preference optimization

    Yurou Liu, Mingyang Li, Xinyuan Zhu, Rui Jiao, Yiming Dong, Xinyu Tang, Yang Liu, Jieping Ye, Bing Su, and Zheng Wang. Drugtrail: Interpretable drug discovery via structured reasoning and druggability-tailored preference optimization. InThe Fourteenth International Conference on Learning Representations. 33 Jiarui Lu, Xiaoyin Chen, Stephen Lu, Chence Shi,...

    2025

  11. [19]

    Techimage-bench: Rubric-based evaluation for professional image generation.arXiv preprint arXiv:2512.12220,

    Minheng Ni, Zhengyuan Yang, Yaowen Zhang, Linjie Li, Chung-Ching Lin, Kevin Lin, Zhendong Wang, Xiaofei Wang, Shujie Liu, Lei Zhang, Wangmeng Zuo, and Lijuan Wang. Techimage-bench: Rubric-based evaluation for professional image generation.arXiv preprint arXiv:2512.12220,

    arXiv

  12. [20]

    Odesign: A world model for biomolecular interaction design.arXiv preprint arXiv:2510.22304,

    ODesign Team. Odesign: A world model for biomolecular interaction design.arXiv preprint arXiv:2510.22304,

    arXiv

  13. [21]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    OpenAI. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Pith/arXiv arXiv

  14. [22]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    OpenAI. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    Pith/arXiv arXiv

  15. [23]

    Biomatrix: Towards a comprehensive biological foundation model spanning the modality matrix of sequences, structures, and language.arXiv preprint arXiv:2606.22138,

    Qizhi Pei, Zhimeng Zhou, Yi Duan, Yiyang Zhao, Wei Li, Han Guo, Liang He, Chengping Li, Chang-Yu Hsieh, Conghui He, et al. Biomatrix: Towards a comprehensive biological foundation model spanning the modality matrix of sequences, structures, and language.arXiv preprint arXiv:2606.22138,

    Pith/arXiv arXiv

  16. [24]

    Piflow: Principle-aware scientific discovery with multi-agent collaboration.arXiv preprint arXiv:2505.15047,

    Yingming Pu, Tao Lin, and Hongyu Chen. Piflow: Principle-aware scientific discovery with multi-agent collaboration.arXiv preprint arXiv:2505.15047,

    arXiv

  17. [27]

    Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh

    URL https://arxiv.org/abs/2503.23461. Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 20730–20740,

    arXiv

  18. [28]

    Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Pith/arXiv arXiv

  19. [29]

    Protein-small molecule binding site prediction based on a pre-trained protein language model with contrastive learning.Journal of cheminformatics, 16(1):125, 2024a

    Jue Wang, Yufan Liu, and Boxue Tian. Protein-small molecule binding site prediction based on a pre-trained protein language model with contrastive learning.Journal of cheminformatics, 16(1):125, 2024a. Liang Wang, Yu Rong, Tingyang Xu, Zhenyi Zhong, Zhiyuan Liu, Pengju Wang, Deli Zhao, Qiang Liu, Shu Wu, and Yang Zhang. Diffspectra: Molecular structure el...

    arXiv

  20. [30]

    Improving paratope and epitope prediction by multi- modal contrastive learning and interaction informativeness estimation.arXiv preprint arXiv:2405.20668, 2024b

    Zhiwei Wang, Yongkang Wang, and Wen Zhang. Improving paratope and epitope prediction by multi- modal contrastive learning and interaction informativeness estimation.arXiv preprint arXiv:2405.20668, 2024b. Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324,

    Pith/arXiv arXiv

  21. [31]

    Proteo-r1: Reasoning foundation models for de novo protein design.arXiv preprint arXiv:2605.02937, 2026a

    Fang Wu, Weihao Xuan, Heli Qi, Hanqun Cao, Heng-Jui Chang, Zeqi Zhou, Haokai Zhao, Ma Jian, Carl Ma, Yu-Chi Cheng, et al. Proteo-r1: Reasoning foundation models for de novo protein design.arXiv preprint arXiv:2605.02937, 2026a. Ruidong Wu, Fan Ding, Rui Wang, Rui Shen, Xiwen Zhang, Shitong Luo, Chenpeng Su, Zuofan Wu, Qi Xie, Bonnie Berger, et al. High-re...

    Pith/arXiv arXiv 2022

  22. [32]

    Next-gpt: Any-to-any multimodal llm

    Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519,

    Pith/arXiv arXiv

  23. [33]

    Qwen2.5-omni technical report.arXiv preprint arXiv:2503.20215,

    Jin Xu et al. Qwen2.5-omni technical report.arXiv preprint arXiv:2503.20215,

    Pith/arXiv arXiv

  24. [35]

    Reasoning-enhanced large language models for molecular property prediction

    Jiaxi Zhuang, Yaorui Shi, Jue Hou, Yunong He, Mingwei Ye, Mingjun Xu, Yuming Su, Linfeng Zhang, Ying Qian, Guolin Ke, et al. Reasoning-enhanced large language models for molecular property prediction. arXiv preprint arXiv:2510.10248,

    arXiv

  25. [2014]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267,

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaugh- lin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267,

    Pith/arXiv arXiv

  26. [2018]

    Speaking the language of science: Toward a general-purpose generative foundation model for the natural sciences.arXiv preprint arXiv:2606.16905, 2026a

    Mingyang Li, Yurou Liu, Jieping Ye, Bing Su, Ji-Rong Wen, and Zheng Wang. Speaking the language of science: Toward a general-purpose generative foundation model for the natural sciences.arXiv preprint arXiv:2606.16905, 2026a. Qingxiao Li, Zikai Wang, Qingli Wang, and Nan Xu. S1-omni-image: A unified model for scientific image understanding, generation, an...

    arXiv

  27. [2019]

    Omni-weather: Unified multimodal foundation model for weather generation and understanding.arXiv preprint arXiv:2512.21643,

    Zhiwang Zhou, Yuandong Pu, Xuming He, Yidi Liu, Yixin Chen, Junchao Gong, Xiang Zhuang, Wanghan Xu, Qinglong Cao, Shixiang Tang, et al. Omni-weather: Unified multimodal foundation model for weather generation and understanding.arXiv preprint arXiv:2512.21643,

    arXiv

  28. [2021]

    Efficient and accurate prediction of protein structure using rosettafold2.BioRxiv, pp

    Minkyung Baek, Ivan Anishchenko, Ian R Humphreys, Qian Cong, David Baker, and Frank DiMaio. Efficient and accurate prediction of protein structure using rosettafold2.BioRxiv, pp. 2023–05,

    2023

  29. [2023]

    URLhttps://doi.org/10.1038/s42256-023-00639-z

    doi: 10.1038/s42256-023-00639-z. URLhttps://doi.org/10.1038/s42256-023-00639-z. Valentin Boussot, Cédric Hémon, Jean-Claude Nunes, and Jean-Louis Dillenseger. Why registration quality matters: Enhancing sct synthesis with impact-based registration.arXiv preprint arXiv:2510.21358,

    arXiv

  30. [2024]

    Anthropic

    doi: 10.1038/s41586-024-07487-w. Anthropic. The claude 3 model family: Opus, sonnet, haiku,

    doi:10.1038/s41586-024-07487-w

  31. [2025]

    Sarp Aykent and Tian Xia

    URLhttps://arxiv.org/abs/2505.01912. Sarp Aykent and Tian Xia. GotenNet: Rethinking efficient 3d equivariant graph neural networks. OpenReview preprint,

    arXiv

  32. [2026]

    Jannis Born and Matteo Manica

    Accessed 2026-06-05. Jannis Born and Matteo Manica. Regression transformer enables concurrent sequence regression and generation for molecular language modelling.Nature Machine Intelligence, 5(4):432–444,

    2026