Overview
| Name | SciReasoner |
|---|---|
| Full name | The paper gives no expansion; the name is the code name |
| Ambiguous names (not auto-linked) | SciReasoner |
| Family | Two papers share the name but are different work; this record means the 2026 structure-reasoning version, and the site must state which is meant
|
| Organization | Shanghai AI Laboratory, The Chinese University of Hong Kong, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Stanford University, University of Oxford |
| Organization type | Nonprofit institute |
| Country | China, United States, United Kingdom |
| Model class | Multimodal science LLM (structure reasoning) |
| Domain | Text, LLMs & agents (also: Proteins, Small molecules) |
| First release | † unverified · paper revision Basis: Preprint v1, with the repository README the same day Note: Two papers share the name and the intended entity must be confirmed first source |
| Other milestones | |
| Latest update | † unverified · no later event Basis: The release event itself The repository holds two commits and no releases; weights and demo are marked as coming soon and the evaluation scripts are said to arrive with the code. No later substantive event was found as of 2026-11-16. Note: Commit dates were not visible when fetched, so the recorded date may differ from the repository's by a day or two source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Released within 12 months with no event found since |
| Openness | Code: Unclear Weights: Unclear The project site lists a repository but no weights or licence were visible as of 2026-11-16; evaluation code for the earlier same-named version sits in a separate public repository under a permissive licence |
| Benchmarks reported | 1 counted (self 1, unconfirmed 7) |
| Wet-lab validation | None recorded |
| Unverified fields | † first_release, release_milestones.paper_v1, latest_update, open_status.code, open_status.weights, training_data.b_private_licensed, benchmarks_reported |
| As of | 2026-11-16 |
What it does
Built on an open 14-billion-parameter general model, it encodes the three-dimensional structure of proteins, small molecules and inorganic crystals as shared structural tokens so that reasoning chains can cite structural evidence, covering proteins, small molecules, crystals and nucleic acids; training runs through alignment and continued pretraining, a supervised cold start and then reinforcement learning. It belongs to the same institute's science-model line as Intern-S but is a mid-size specialist rather than a general science base model, with ether0 and TxGemma as chemistry-specialist comparisons.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 8 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 07 USPTO family (+PaRoutes) | Developer report, official protocol | — | Self-reported | USPTO-50K single-step retrosynthesis, matrix note only, unverified †unverified · no source link |
| 14 PubMedQA | Unconfirmed (not counted) | — | Unconfirmed source | Not in the paper text †disputed · no source link |
| 17 CAFA | Unconfirmed (not counted) | — | Unconfirmed source | Self-reports on DeepFRI-GO; CAFA-3 only in a figure caption, a modified protocol if kept; paper identity unresolved (two same-name papers). †disputed · evidence |
| 18 LAB-Bench / LABBench2 | Unconfirmed (not counted) | — | Unconfirmed source | Not in the paper text †disputed · no source link |
| 23 BioASQ | Unconfirmed (not counted) | — | Unconfirmed source | Not found in the paper text. †disputed · no source link |
| 29 BixBench | Unconfirmed (not counted) | — | Unconfirmed source | Not found in the paper text †disputed · no source link |
| 41 BioML-bench | Unconfirmed (not counted) | — | Unconfirmed source | The paper text does not mention this benchmark; the only evidence is a site page. †disputed · no source link |
| 42 BioMysteryBench | Unconfirmed (not counted) | — | Unconfirmed source | The paper text does not mention this benchmark; the only evidence is a site page. †disputed · no source link |
Other evaluations (outside the 44)
- A large self-selected task set (Own task set) Self-reported: The paper reports 86 evaluations with best results on 67, with baselines run by the authors
- Public tasks such as virtual screening and localisation (Public benchmark not among the 44) Self-reported: Compared with existing methods on a virtual-screening set of over a hundred targets, subcellular localisation, RNA tasks and materials properties
- Double-blind expert assessment (Case study) Self-reported: The authors report experts judging its reasoning chains better than or comparable to a frontier general model in the great majority of cases
Developers report: encoding three-dimensional structure as shared tokens lets reasoning cite structural evidence, yielding best results on most of a large self-selected task set, with reasoning chains judged better than or comparable to a frontier general model in a double-blind expert assessment.
Independent evidence: No independent evidence found. Weights are unreleased, no third-party evaluation was found and there is no record of official blind-test participation
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Dozens of public archives spanning protein sequences and structures, literature, chemistry and bioactivity, reactions, materials and nucleic acids, with proteins homologous to the test set and reactions overlapping the reaction test set removed from pretraining source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The lead organisation is a government-backed new-type research institution in Shanghai whose lines include a general large-model series, an evaluation suite and an open data platform; collaborators include several Chinese universities, a pharmaceutical research institute and two overseas universities.
- Funding: Supported by public and municipal funding with no commercial financing.
- Products and partnerships: The same institute runs a general science base model line recorded here as Intern-S; comparisons for this model include the chemistry specialists ether0 and TxGemma.
- Related models: Intern-S, ether0, TxGemma
Reading notes
- The claim of best on most tasks rests on a self-selected task set with baselines the authors ran, which is not a standing on third-party boards.
- Gene-ontology prediction uses a third-party test collection rather than the official blind test, which requires submitting to the competition track.
- The comparison with general large models shows such tasks go badly without structural tokens; it does not show strength in conversational biomedical question answering.
- The two same-named papers must be disambiguated first, or neither first release nor latest update holds.
Caveats
- The identity between two same-named papers is unresolved and the page must state which is meant.
- No weights or licence were visible, so nothing can be rechecked externally.
- Six benchmark pairings rest on this site's own pages and must display as unconfirmed.
- The gene-ontology work is not official blind-test participation.
- The paper contains no wet-lab validation.
Sources
- Preprint 2026 (preprint)
- Preprint full text (preprint)
- Earlier paper with the same name, 2025 (preprint)
- Code repository (repository)
- Project site (official site)
Corrections
None recorded.
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.