SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines

1. Introduction

• Existing scientific models remain limited in cross-domain generalization, heterogeneous-modality coverage, and explicit multi-step reasoning, while specialist pipelines often require retraining for unseen tasks.The paper also notes that multi-discipline models may cover only a limited task set, such as NatureLM’s 30 tasks.Source: §1 · p. 5
• The model unifies natural language, scientific sequences, molecular strings, and material representations through multi-representation pretraining, instruction-aligned schemas, and reasoning-oriented post-training.Its reasoning stage bootstraps long-form rationales and uses physics- and task-aware rewards.Source: §1 · p. 5
• Five capability families cover translation, text and knowledge extraction, property prediction, property classification, and constrained sequence generation and design.The paper presents these capabilities as supporting interoperability, screening, and design across scientific domains.Source: §1 · pp. 5–6
• The resulting single backbone supports prediction, design, translation, and literature-grounded extraction in a read–reason–design loop, with strong cross-domain generalization across chemistry, proteins, genomics, and materials.The framework also describes a solve–check protocol using retrieval and domain tools, with rewards that enforce unit, stoichiometry, and constraint consistency.Source: §1 · p. 6

2.1. Pretraining Data Types

• Pretraining combines scientific text, pure sequences, sequence–sequence and sequence–text pairs, and general text to support scientific and linguistic capabilities.The data include DNA/RNA, proteins, small molecules, materials, and cross-discipline examples such as molecule–protein binding and multiple encodings of the same molecule.Source: §2.1 · pp. 6–7

Data Preparation Details

• Protein data contribute 55 billion tokens (33.1%), chemistry 50 billion (30.1%), DNA 30 billion (18.1%), RNA 26 billion (15.7%), and materials 5 billion (3.0%).The composition is summarized across the five scientific domains in Figure 2.Source: Figure 2 · p. 7, Data Preparation Details · p. 8

2.2. Training Protocol

• The model is trained for one epoch with mixed scientific and general text and paired sequence data in warm-up, then SFT-style data to refine instruction following and task-aligned outputs.The base models are Qwen3 1.7B and 8B; training loss curves are reported in Figure 3.Source: §2.2 · pp. 8–9, Figure 3 · p. 9

3.1.1. Data

• The SFT corpus integrates about 100 scientific subtasks across chemistry, nucleic acids, proteins, materials, and cross-domain tasks, with data tagged and converted to a unified dialogue format.Table 2 describes a corpus of 63.35 million samples.Source: §3.1.1 · p. 9, Table 2 · p. 10

3.1.2. Training Protocol

• Joint multidisciplinary training showed performance degradation on some subtasks from data imbalance; additional pretraining, expanded materials SFT data, and task-level reweighting were used to mitigate it.The paper reports that reweighting alleviates performance decline in mixed-dataset training.Source: §3.1.2 · p. 10

3.2.1. Data

• RL examples are selected using empirical solve rates to target medium-difficulty instances, excluding nearly always-solved or nearly always-failed items when possible.The method samples stochastic rollouts, retains partial-success examples, and forms a 1,000-example training pool per subtask; the paper states this focus yields stronger and more stable improvement.Source: §3.2.1 · p. 11

3.2.2. Training Protocol

• Annealed cold-start adaptation replaces SFT targets with CoT-augmented examples for thinking tasks while preserving direct-answer data for instant tasks; DAPO then optimizes with task-specific softened rewards.The reward schemes use distance-based measures for prediction, matching for retrieval and extraction, and tool verification for scientific software tasks.Source: §3.2.2 · pp. 11–12

4.1. Task Overview

• The evaluation covers translation across scientific representations, text and knowledge extraction or QA, property prediction and classification, and unconditional or conditional generation and design.Tasks span chemistry, materials, nucleic acids, and proteins, including targeted antibody design and programmable RNA switches.Source: §4.1 · p. 13

4.2. Baseline

• Baselines include closed- and open-source general-purpose models, non-language expert systems, and the best-performing specialist language model for each task.The named general-purpose baselines are Gemini-2.5-pro and GPT-o3 (closed-source), and GPT-oss-120B and DeepSeek-r1 (open-source).Source: §4.2 · p. 13

4.3. Scientific Translation Tasks

• SciReasoner-8B achieves >93% formula-recovery accuracy, molecular-captioning MENTOR 0.61, and protein-translation ROUGE-L up to 0.98, consistently outperforming translation baselines.The evaluation also includes SMILES↔IUPAC, SMILES↔formula, and IUPAC↔formula translation, with exact-match and structure-equivalence assessment.Source: Table 3 · p. 14, §4.3 · p. 17
• In translation case studies, SciReasoner generates the ZIF-8 IUPAC name successfully where GPT-5 fails, and its SMILES-derived name preserves the reference structure and stereochemistry.For the SMILES→IUPAC example, GPT-5, LLaSMol, and Gemini-2.5-Pro produce chemically inconsistent names; the model’s molecular caption also matches the reference’s core facts without unsupported inferences.Source: Figure 6 · p. 16, §4.3 · pp. 16–17, Figure 7 · p. 17

4.4. Scientific Knowledge Extraction and Question Answering

• Scaling from SciReasoner-1.7B to 8B improves scientific extraction and QA, with the 8B model surpassing prior Specialist LLMs on every task.The largest gains occur in chemical–protein and chemical–disease relation extraction; multiple-choice QA approaches a ceiling while open-ended and true/false performance remains competitive with strong general-purpose models.Source: Table 4 · p. 18, §4.4 · pp. 18–19

4.5. Property Prediction Tasks

• The model predicts a broad range of structural, electronic, and thermodynamic properties across multiple materials databases.The unified property suite also covers molecule and multi-omics regression, using chemical formulas, SMILES, and DNA/RNA/protein sequences as inputs.Source: §4.5 · pp. 19, 21
• Scaling from SciReasoner-1.7B to 8B yields consistent property-prediction accuracy gains and matches or surpasses strong baselines on multiple benchmarks.Reported gains span materials, molecules, and biology multi-omics; the largest jumps occur on multi-output or context-dependent tasks, including enhancer activity and APA isoform usage.Source: §4.5 · p. 23

4.6. Property Classification Tasks

• Across classification tasks, SciReasoner-8B improves over 1.7B, often surpasses GPT-o3/oss, and is best or second-best on most tasks.Reported gains include about 3% AUC on MPclassification and SNUMATclassification, up to 15 F1 on gene-centric association tasks, and 2–10 points MCC/accuracy across several biology multi-omics tasks.Source: Table 6 · p. 24, Table 7 · p. 26, §4.6 · p. 30
• On GO-term classification, SciReasoner-8B has higher medians and upper quartiles than 1.7B across all three test sets, with gains most evident on IDFilterSeq and UniProtSeq.Upper whiskers reach 9–10 correctly predicted terms for 8B versus 5–8 for 1.7B, indicating stronger generalization beyond CASP-similar sequences.Source: Figure 12 · p. 29, §4.6 · p. 29

4.7. Generation and Design Tasks

• Unconditional generation evaluates plausible RNA, material, molecule, and protein candidates under minimal format constraints, with generated samples also considered as seeds for conditioned design.The RNA-generation visualization reports pLDDT scores of 74.6, 72.3, and 71.5 for three generated sequences.Source: §4.7 · pp. 31–32, Figure 14 · p. 32
• Unconditional evaluation tested 800 molecules for RDKit validity and uniqueness, 500 proteins for length, diversity, and validity, 5,000 RNAs for average MFE of valid sequences, and 5,000 materials for SMACT validity.Protein generation also probes the model’s ability to create biologically feasible proteins from scratch.Source: §4.7 · pp. 32–33
• Conditional material tasks ask the model to design compositions containing specified elements or materials with a target bulk modulus.These tasks test generation under elemental or physical-property constraints.Source: §4.7 · p. 33
• Conditional material generation is evaluated with SMACT and three complementary metrics: Precision, Novelty, and Success Rate.Precision measures overlap with ground-truth elements, Novelty measures predictions absent from the ground-truth set, and Success Rate measures exact recovery of all ground-truth elements.Source: §4.7 · p. 33
• Conditional small-molecule generation uses reactants or reagents, text descriptions, scaffolds, or target properties to test instruction following, constraint satisfaction, and novelty-correctness balance, following Mol-Instruction [81] and SMolInstruct [100].Source: §4.7 · p. 33

4.8. Reasoning Examples

• The section presents examples across materials science and molecular biology, showing input prompts, reasoning, final answers, and references.The examples are framed as illustrations of reasoning capabilities across scientific domains.Source: §4.8 · p. 36

4.8.1. Material Task: SNUMAT Classification

• The RbCuCl3 example links its ionic components to Cu’s +2 oxidation state and d9 configuration to predict its HSE bandgap type.The passage describes the resulting prediction as correct and the reasoning as a physically meaningful sequence.Source: §4.8.1 · p. 36

4.8.2. DNA Task: Promoter-300

• For the 300 bp DNA sequence, the model weighs length, GC-rich content, and CpG islands, while recommending further bioinformatics analysis for definitive promoter identification.It notes the lack of an explicit TATA box and allows for other transcription-factor binding sites.Source: §4.8.2 · p. 37

4.8.3. Protein Task: Solubility

• The protein-solubility example relates abundant hydrophilic and charged amino acids, and the absence of long hydrophobic regions or transmembrane domains, to predicted solubility.The account connects these sequence features to interactions with water and aggregation or insolubility.Source: §4.8.3 · pp. 37–38

4.9. Effectiveness of Pretraining

• Pretraining improved all ten evaluated tasks: median gain was +15.0 pp across nine higher-is-better tasks, while ESOL RMSE fell by 98.7%.The largest absolute gains occurred on sequence-interaction and regulatory tasks; promoter detection at 300 bp benefited more than core-promoter detection at 50–100 bp.Source: §4.9 · pp. 38–39, Table 9 · p. 40

5. Conclusion

• Task-aware tokenization maps natural language, biological sequences, molecular strings, and materials representations into a shared backbone with consistent input–output schemas.The framework combines multi-representation pretraining, instruction-driven alignment, and reasoning that supports fast responses and step-by-step deliberation.Source: §5 · p. 40
• The model supports property prediction and classification, constrained sequence generation and design, scientific translation, and text or knowledge extraction.The conclusion positions this design as moving beyond narrow discipline-specific solutions and limited task menus.Source: §5 · p. 40
• Retrieval and tool execution are used to ground outputs for improved fidelity and verifiability.The conclusion describes this grounding alongside the model’s shared representation and consistent schemas.Source: §5 · p. 40

Appendix

• The evaluation tables cover scientific translation, knowledge extraction and QA, property prediction, classification, and generation and design tasks.The captions identify Tables 10–13 and 15 as covering these task groups.Source: Table 10 · p. 49, Table 11 · p. 50, Table 12 · p. 51
• For some extremely large test sets, results use randomly sampled 1,000-item test subsets to avoid substantial testing costs.The sampling note appears in the captions for the translation, prediction, and classification tables.Source: Table 10 · p. 49, Table 12 · p. 51, Appendix · p. 52
• Table 14 covers biology multi-omics property-classification tasks, with dashed lines separating tasks that report multiple metrics.The supplied table material identifies multi-omics classification as including binary and multi-label groups.Source: Appendix · p. 53, Table 14 · p. 54
• Table 15’s caption uses N/A to indicate that the model cannot generate samples meeting the required specifications.Source: Table 15 · p. 55

Protein Generation

• Figure 20 is labeled as an additional visualization for protein generation.Source: Figure 20 · p. 56
• Figure 21 is labeled as an additional visualization for protein generation.Source: Figure 21 · p. 56

• Figure 22 provides an additional visualization for the SMILES-to-IUPAC translation task.Source: Figure 22 · p. 56

Task Examples

• Molecular prompts show numeric solubility and logD prediction alongside yes/no predictions for blood-brain barrier permeability, toxicity, HIV inhibition, and side effects.Examples report a log solubility of -1.41 mol/L and a logD of 1.090.Source: Task Examples · p. 57
• Materials prompts request JSON-formatted property values from crystalline-material formulas, with examples spanning volume, band gap, formation energy, final energy, density, and CO2 adsorption.One example predicts volume as 128.62405239223665 for KPrMnNbO6.Source: Task Examples · pp. 58–59
• Protein prompts cover catalytic-reaction and fluorescence prediction as well as protein stability and cellular thermostability assessment.The examples report GFP fluorescence intensity of 3.67 and a thermostability measure of 56.77.Source: Task Examples · pp. 60–61
• DNA and RNA sequence prompts illustrate enhancer activity, isoform likelihood, ribosome loading, RNA-switch states, CRISPR disruption, and siRNA efficiency predictions.The RNA-switch example reports ON state 0.02, OFF state 0.47, and ON/OFF ratio 0.06.Source: Task Examples · pp. 61–63
• Other examples ask for RNA-protein interaction detection and antibody-antigen binding detection, with responses predicting positive interaction and plausible binding, respectively.The examples also show materials classification prompts for stability and SOC.Source: Task Examples · pp. 63–65
• Protein analysis prompts include domain/motif search, protein function and localization, solubility, and enzyme EC-number prediction.One function example identifies rRNA binding and a ribosomal location; a motif example reports a DDXXXE motif.Source: Task Examples · pp. 63–64
• Examples cover epigenetic-mark prediction and non-coding RNA family classification, alongside RNA modification prediction.The listed test sets contain 28,741 epigenetic-mark, 2,600 non-coding RNA classification, and 1,200 RNA modification samples.Source: Task Examples · p. 66
• DNA examples ask the model to identify transcription-factor binding sites in both mouse and human sequences.The examples are labeled TF-Mouse and TF-Human, with 10,005 and 28,741 test samples, respectively.Source: Task Examples · pp. 66–67
• Other DNA examples predict enhancer–promoter interaction potential and identify promoter or core-promoter sequences.The associated test sets include 1,762 enhancer–promoter, 5,920 promoter, and 5,000 core-promoter samples.Source: Task Examples · pp. 66–67
• Protein examples also predict subcellular localization and fold type, with separate test sets of 2,772 and 3,243 samples.The examples return a localization label or a fold-type value.Source: Task Examples · pp. 69, 71
• Protein annotation examples return UniProtKB keywords or Gene Ontology terms from sequences, including CASP-similar, ID-filtered, and UniProt sequence sets.The cited examples include keyword sets for CASP-similar and ID-filtered sequences and GO-term sets for CASP-similar and ID-filtered sequences.Source: Task Examples · pp. 69–72
• Gene-symbol prompts request tissue-expression predictions or associated cancer names, with the two cancer tasks each listing 148 test samples.The tissue task lists 2,181 test samples.Source: Task Examples · pp. 72–73
• Molecule examples translate SMILES to IUPAC names and IUPAC names to SMILES, and ask for molecular formula prediction from either representation.The translation tasks each list 2,993 test samples; the examples show a name, SMILES string, and formula response.Source: Task Examples · p. 73

Molecular and protein descriptions

• Protein and molecule examples translate input sequences or structures into functional or descriptive text, including enzyme activity, chromatin roles, and drug properties.The molecule-captioning example describes etomidate’s sedative and anesthetic properties; protein examples describe functions including pyrophosphatase activity and chromatin regulation.Source: Task Examples · pp. 74–75

Unconditional generation

• Unconditional generation examples produce molecular, protein, RNA, and material outputs in their respective scientific formats.The prompts request a molecule, protein sequence, novel RNA sequence, or material with a composition.Source: Task Examples · pp. 76–77

Conditional design

• Conditional design examples generate molecules from descriptions, proteins for a specified function, and materials from compositions or target bulk modulus.One protein prompt specifies modulation of glycine decarboxylation; material prompts specify constituent elements or a bulk modulus of 75.321 GPa.Source: Task Examples · pp. 77–78

Chemical reaction tasks

• Chemical reaction examples cover forward product prediction, reagent prediction, retrosynthesis, and forward synthesis from reactants.The examples include prompts to propose products from reactants and to infer possible reactants from products.Source: Task Examples · pp. 78–79

Biomedical information extraction

• Biomedical text examples extract chemical–disease relations, chemical entities, and subject–relation–object triples from clinical or research passages.The examples return a chemical–disease pair, a chemical entity, or triples describing a drug’s antagonist relation.Source: Task Examples · pp. 79–80

Medical question answering

• Medical question-answering examples include a multiple-choice response, a direct open-ended answer, and a qualified true/false response.The true/false example notes conditions under which routine intraoperative frozen-section examination may be useful or not indicated.Source: Task Examples · pp. 80–81

• Figure 23 reports TMScore for the last 1,000 samples.Source: Figure 23 · p. 82