AI for Bio · Model

RetroDFM-R

Reasoning-style retrosynthesis LLM trained with verifiable chemistry rewards, emitting precursors with a reasoning trace

Small moleculesAcademic labChinaActive中文

Overview

NameRetroDFM-R · aliases: RetroDFM-R, RetroDFM-R-8B
Full nameThe paper uses the model name only; no expanded name
Ambiguous names (not auto-linked)ChemDFM
FamilyNot a family entry. The sibling chemistry LLM line has no site entity record, so no family entry
OrganizationShanghai Jiao Tong University (X-LANCE Lab; School of Chemistry and Chemical Engineering), Suzhou Laboratory
Organization typeAcademic lab
CountryChina
Model classRetrosynthesis (chemistry LLM + reinforcement-learning reasoning)
DomainSmall molecules (also: Text, LLMs & agents)
First release · paper revision
Basis: arXiv 2507.17448 v1 submitted
Note: The site previously cited another paper's identifier (2510.16588); corrected source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 (v2 on 2026-08-24) source
  • Code release: † · Code repository commits including the licence; no release object exists source
  • Weights release: † · First 8B weights repository created (README announcement 2025-11-22) source
Latest update† partly verified · new version
Basis: Code repository commits (no release object)
Code repository commits, alongside the paper's v2 (2026-08-24) and the creation of a new 8B weights repository (2026-08-20). No journal version was found and whether the newer weights change results is unverified. This is the latest sourced event
Note: Previously 2026-08-24; corrected to 2026-08-25 per the repository commits source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Paper revision, weights and code events within 12 months
OpennessCode: Open · GPL-3.0
Weights: Open · GPL-3.0 (per the group's model pages)
Weights and code are GPL-3.0 per the group's public pages; the hosting namespace is the lab's open-source space, not a company
Benchmarks reported1 counted (self 1)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[0], release_milestones.weights_release, release_milestones.code_release, latest_update
As of2026-11-16

What it does

Starts from the group's chemistry LLM, continues pretraining on retrosynthesis text, cold-starts with distilled chain-of-thought and then applies reinforcement learning with verifiable signals such as reaction plausibility and product match, emitting precursors together with a reasoning trace. Unlike the pretraining route of Uni-Mol3 or the sequence-to-sequence route of Chemformer, the reasoning text is part of the output, which is its interpretability claim.

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 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
07 USPTO family (+PaRoutes)Developer report, official protocol—Self-reportedUSPTO-50K/FULL; with test-time augmentation noted · evidence

Other evaluations (outside the 44)

  • Double-blind preference study (Own task set) Self-reported: A double-blind comparison of the model's precursors against ground-truth reactants, with most responses rating the model output as better or equal; disagreements cluster on halogen substituent swaps, and the study was run by the authors
  • Multi-step route planning case studies (Case study) Self-reported: Several multi-step synthesis route case studies with no common metric

Developers report: reinforcement learning with verifiable chemistry rewards beats existing sequence and edit-based methods on the official retrosynthesis splits while producing a readable reasoning trace.

Independent evidence: No independent evidence found. No third-party re-run or independent evaluation was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesContinued pretraining on retrosynthesis chemistry data, including structure-name pairs from a public compound library and public retrosynthesis prediction data source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoThe paper reports no wet-lab validation source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic lab: a cross-media language intelligence lab in a Chinese university's computer science school, working with the same university's chemistry school and a regional laboratory; its open-source namespace is not a company. The lab also runs a joint lab with a speech-technology company.
  • Funding: Acknowledgements list a national key programme, three national natural science foundation grants and two municipal science projects; training used eight datacentre-class GPUs.
  • Products and partnerships: Sibling work includes a chemistry LLM with multimodal and reasoning variants (three public papers); on-site competitors are Chemformer and Uni-Mol3. No company partnerships disclosed.
  • Related models: Chemformer, Uni-Mol3

Reading notes

  1. The site cited the wrong paper identifier; the correct source is arXiv 2507.17448 and the other identifier belongs to a different paper.
  2. The site's unknown-class figure is unverified and the abstract gives a different value, so always go back to the abstract's framing.
  3. With and without test-time augmentation are two settings with a visible gap; the figures once recorded as an abstract-versus-text conflict were a single reaction class, not the overall framing.
  4. USPTO data are frozen and leak templates, so high scores are not synthetic feasibility; there is no wet-lab or route-level independent validation.

Caveats

  1. All benchmark results are author-reported with no board entry.
  2. Weight dates rely on host-side repository creation; whether the two 8B weight releases differ in results is unverified.
  3. The reasoning trace aids readability, but its correctness rests on self-assessment only.
  4. The code repository has no release objects, so versioning can only be read from commits.

Sources

  1. Model paper, arXiv 2507.17448 (correct source) (preprint)
  2. Paper v1 full text (preprint)
  3. Code repository (repository)
  4. First 8B weights repository (package registry)
  5. Newer 8B weights repository (package registry)

Corrections

4 corrections
FieldWasNowBasisFound
sourcesCited arXiv 2510.16588 as this model's paperThe correct id is arXiv 2507.17448; the earlier id belongs to another papersource2026-11-16
benchmarks_reportedThe "unknown class" retrosynthesis figure shown on the siteThe figure is not confirmed; recorded per the abstract, and this page does not transcribe the scoresource2026-11-16
release_milestones.weights_release8B weights released 2026-01-12First 8B weights repository created 2025-11-20 (announced 2025-11-22); 2026-01-12 was only a README edit; another 8B repository was created 2026-08-20source2026-11-16
latest_update2026-08-24 (paper v2)2026-08-25 (code repository commit, no release object)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.