LongevityBench tests AI models on aging biology, finding uneven performance

Researchers have introduced an open toolkit for evaluating how well artificial intelligence interprets aging-related biological data, addressing a practical question for longevity science: whether models can extract useful information from measurements rather than simply produce convincing scientific prose.

The study, published in Cell’s September 17 issue and announced that day by Insilico Medicine, introduces LongevityBench, specialized language models and a research interface called Longevity Claw. Its immediate contribution is a shared framework for testing computational tools. It does not establish that an AI system can improve human healthspan or identify an effective treatment.

Testing predictions across biological datasets

According to the study abstract, LongevityBench comprises 17 tasks across five biological data domains. The researchers evaluated 18 frontier AI systems from six developer teams and found that no single system performed best across every task. Predicting age from molecular measurements remained particularly difficult, regardless of model size.

This was a computational benchmarking study using existing datasets, rather than a prospective clinical trial. The outcomes were model predictions assessed against defined answers, not changes in disease, physical function or survival following treatment.

Liquid AI, a collaborator, describes the benchmark as containing 25,457 prompts. Its source material includes clinical records from the National Health and Nutrition Examination Survey, DNA methylation studies, gene-expression profiles, plasma-protein measurements and genetic evidence. Biological records are converted into structured text that a language model can process.

Tasks include identifying which of two participants is older, assigning an age group and estimating age numerically. These are related but distinct problems: success at comparing two profiles does not automatically establish accurate numerical prediction.

The collaborators report separating training and evaluation data according to the underlying source—for example, by survey wave for clinical records, by study for methylation data and by individual for gene-expression and protein data. Those distinctions matter when assessing whether performance extends beyond examples encountered during training.

Specialized models showed an advantage

The investigators also trained five compact Longevity-LLMs on aging-specific data. Ranging from 0.6 billion to 9 billion parameters, the models matched or exceeded much larger frontier systems on the benchmark, according to the paper’s abstract.

Insilico reports that performance also depended on how questions were phrased. That sensitivity is a practical limitation for research workflows: a model’s apparent competence can change with the presentation of the same scientific problem.

The findings suggest that training tailored to biological data can improve selected computational tasks. They leave open how reliably those advantages transfer to unfamiliar populations, laboratories or research questions.

The toolkit also includes Longevity Claw, which Insilico describes as an interface combining models with analytical and evidence-retrieval tools. Its proposed role is to support research workflows and prioritize candidate targets. A computational nomination remains a hypothesis requiring biological testing.

Research utility still needs validation

Liquid AI’s published model documentation explicitly limits the intended use to aging research and interpretation of molecular data. It says predictions require experimental validation and that performance is strongest for data types represented in training.

That boundary is central to interpreting the announcement. Predicting age from a blood or tissue profile does not demonstrate that the model has identified a cause of aging. Nor does better benchmark performance show that following its suggestions would prevent disease.

Both Insilico and Liquid AI participated in developing the tools, so their technical accounts are developer reports rather than independent replication. Public availability gives outside researchers an opportunity to test robustness and usefulness. For longevity research, the meaningful next step is determining whether these tools produce reproducible insights that survive experimental investigation.