STAMP gives AI models a structured way to edit molecules
Researchers at the National University of Singapore have introduced STAMP, a molecular representation designed for learning and editing. The preprint, submitted on October 13, gives language models explicit molecular fragments to select and change, with a separate compiler carrying out the edit.
The work reports improvements on several computational benchmarks. It does not demonstrate a new medicine, a successful laboratory synthesis or a clinical outcome. Its central contribution is a way to connect a model’s proposed change to a defined operation on a molecular structure.
Making the target of an edit explicit
The paper separates a molecule into a tree of fragments. Different tokens describe the fragments, their attachment sites and their branching relationships. Each fragment has an address, and each descendant subtree occupies a contiguous section of the representation.
For instruction guided editing, the language model chooses a target and an operation, such as adding or replacing a fragment. The compiler checks the proposed connections and constructs the resulting molecular graph. This reduces the need for the model to rewrite the entire molecular string while also preserving the parts that should stay unchanged.
The distinction is important when interpreting the results. A model that selects an operation with structured chemical information receives a different task from one asked to generate a complete product string. The experiment measures the combined representation and execution interface, rather than proving that a tokenizer alone makes a model better at chemistry.
Editing scores improve, with important exceptions
The authors tested two frozen language models on six S2 Bench subtasks, with 500 examples in each subtask. Across three molecular editing tasks, mean success for Qwen 2.5 7B Instruct rose from 29.3% with complete SMILES rewriting to 62.0% with STAMP actions. The second tested model, identified in the paper as DeepSeek V4 Pro, rose from 76.2% to 85.4%.
Those averages do not describe a clean sweep. On the substitution task, the DeepSeek model scored 83.0% with STAMP actions, below its 86.6% result with SMILES rewriting. Additional controls using atom indexed SMILES actions narrowed differences and sometimes reversed the result. The paper therefore supports a more specific conclusion about useful interfaces, not universal superiority over other molecular representations.
In a separate property prediction comparison, the STAMP encoder reduced error on 23 of 30 targets against a matched SMILES encoder. A classical support vector machine reference still had the stronger aggregate result. The distinction between beating one neural baseline and beating the strongest reference in the comparison matters here.
The definition of success limits the claim
The evaluation methods define molecular editing success through changes in specified functional group counts. Property optimization succeeds when a calculated property moves in the requested direction. That test does not require a minimum improvement or a molecular similarity threshold.
A small numerical change can therefore count as success in the property task. That is useful for testing whether an instruction was executed, but it cannot establish that the resulting molecule is worth developing. The study also reports docking based optimization, which remains a computational comparison rather than experimental evidence of biological activity.
The frozen model evaluations used one complete run per condition. Several other experiments used multiple seeds, while some baseline figures came from prior publications. These designs should be kept distinct when judging repeatability and comparison strength. ByteForward has not reproduced the experiments, and no peer reviewed publication is identified in the inspected arXiv record.
Public code comes with access and licensing limits
The authors link a public STAMP repository containing the representation software, vocabulary tools, benchmark components and research workflow scripts. Its documentation describes separate training and generation workflows, environment specifications and cluster launch recipes. Those are research materials, not evidence of a supported hosted product.
The repository explicitly says it has no selected top level software license. It also describes the matching data snapshot as private and says datasets, checkpoints and generated results are kept outside the repository. We could not retrieve the linked dataset page. Publicly visible code should therefore not be presented as a fully licensed, independently reproducible package.
Researchers considering a replication should first establish which data and checkpoints they can obtain, which revision matches the paper and what reuse permissions apply. The repository also warns that many cluster scripts contain site specific defaults. A clean installation and a successful published benchmark are separate milestones.
The next question is how much the interface contributes
The most informative follow up would hold the model, chemical information and execution support constant, then compare the fragment interface with equally capable alternatives. Repeating the tests across runs would help distinguish a dependable improvement from an advantage confined to one setup.
STAMP offers a concrete research direction for making model outputs easier to execute and inspect. Its practical value will depend on reproducing those gains and showing that they survive more demanding objectives. For now, the evidence supports a computational method worth investigating, with clear limits around access, evaluation and real world chemistry.
Research summary adapted from the October 13 preprint by Yuxuan Ren, Jiatong Li, Ziqiao Meng and Yatao Bian, published under CC BY 4.0. Wording and explanatory framing are ByteForward’s.
Illustrative photograph published in 2025 by Chidera Faustina Okeke, used under the Unsplash License. It does not depict the STAMP team or an experiment from the paper.
