Open benchmark tests whether AI can reason across aging biology – and finds specialist models can outperform far larger systems.
Artificial intelligence has become adept at reading the scientific literature, retrieving facts and sounding convincingly at home in the language of biology. Asking it to reason directly from the molecular messiness of aging, however, is a rather different test.
A study published in Cell has introduced LongevityBench, an open benchmark designed to test AI systems across 17 tasks spanning clinical data, genetics, epigenetics, transcriptomics and proteomics. The researchers evaluated 18 frontier AI systems and developed five compact Longevity-LLMs trained specifically on aging data; despite containing just 0.6 billion to 9 billion parameters, the specialist models matched or exceeded much larger general-purpose systems across the benchmark [1].
Longevity.Technology: A 0.6-billion-parameter model giving its vastly larger cousins a run for their money makes for an appealing David-and-Goliath result; more importantly, however, it raises a consequential question about what AI for longevity actually needs. Aging biology is not simply another corpus to be read – it is a messy, multidimensional system to be reasoned over, and fluency in the literature is not the same as being able to extract biological meaning from methylation patterns, transcriptomes, proteomes and clinical data. LongevityBench rather neatly exposes that distinction: frontier models that can speak impressively about aging falter when asked to work directly from some of its underlying molecular measurements, while much smaller models trained specifically on aging data can compete with, and sometimes surpass, them.
That does not mean AI has learned how aging works – still less how to stop it. Conventional machine learning remains superior for some tasks, biological age prediction remains stubbornly difficult and Longevity Claw’s 328 proposed targets are hypotheses requiring experimental validation, not 328 geroprotectors waiting patiently to be bottled. But that may be precisely why this work matters. Longevity has no shortage of data and increasingly little shortage of AI capable of talking about it; what it needs are systems capable of turning heterogeneous biological measurements into hypotheses that survive contact with experimental biology. Making the benchmark, models and research platform openly available also means those claims can now be tested rather than merely admired. In a field currently enamored of ever larger models, the lesson may be pleasingly inconvenient – when it comes to understanding aging, knowing more biology could matter rather more than simply having more parameters.
Putting AI to the biological test
Most biomedical AI benchmarks test knowledge that may already exist somewhere in a model’s training material. LongevityBench attempts something different: prompts are generated from primary biological measurements, asking models to infer conclusions such as which individual is older from a methylation profile, predict chronological age from plasma proteomics or estimate the lifespan effect of genetic perturbations. The complete evaluation comprises 25,457 prompts.
This matters because, as the authors note, LLMs can produce fluent answers without an internal mechanism capable of separating established biological fact from “pattern-matched correlation.” A model that has memorized what the literature says about a gene is not necessarily capable of recognizing an aging-related signal buried within an unfamiliar omics profile [1].
The distinction appeared starkly in a supplementary test based on CellAge. Frontier models performed extremely well when asked whether published evidence indicated that particular genes promoted or inhibited senescence – Gemini-3.1-Pro scored 95.9% and GPT-5 94.9% – yet their performance on tasks built from primary omics measurements was generally weaker than that of the specialized models. The authors conclude that interpreting experimental data represents a capability distinct from knowledge retrieval [1].
Smaller models, narrower expertise
No frontier model dominated across all five biological domains and several struggled particularly with raw molecular measurements. Among the specialist systems, however, L-Qwen3.5-9B achieved the highest overall benchmark rank; even L-Qwen3-0.6B, the smallest Qwen-based specialist model, placed sixth among the 26 models included in the main comparison [1].
The implication is not simply that small is beautiful. Domain-specific training appears able to compensate for scale when the problem is sufficiently specialized; the authors argue that, for certain biological patterns, models below three billion parameters can match systems many times their size.
Yet there were stubborn limits. Predicting age numerically from omics remained difficult, while conventional statistical approaches sometimes retained the advantage. In proteomic age regression, for example, the best specialist LLM achieved a mean absolute error of 5.7 years, compared with 3.9 years for an elastic-net clock trained on the same protein panel [1].
Alex Zhavoronkov, founder and co-CEO of Insilico Medicine, said: “Our work with rentosertib demonstrated that an AI-discovered and AI-designed drug can influence biological aging signatures in patients. Now, we are opening the tools that can help the global scientific community discover the next generation of longevity therapeutics. LongevityBench establishes a rigorous standard for measuring whether AI can truly reason about aging biology, while our specialized models and Longevity Claw turn that intelligence into an autonomous discovery engine.”
From answers to targets
Benchmarking is only part of the project. The researchers also placed L-Qwen3.5-9B inside Longevity Claw, an agentic platform combining the model with tools for aging-clock calculation, gene-set enrichment and population profiling, then asked it to search across 14 categories spanning the 12 hallmarks of aging for potential therapeutic targets.
The exercise produced 328 candidate genes. Twelve overlapped with a previously published set of 300 aging targets identified using Insilico’s PandaOmics platform, representing a statistically significant 2.4-fold enrichment; two candidates, CXCL12 and ADAMTS14, were among nine experimentally validated targets in that earlier work [1].
Intriguingly, KDM1A repeatedly emerged from the new system – and had already been independently validated in a 2023 study as an aging and cancer target whose modulation extended lifespan in C elegans [2]. Proof of therapeutic utility, certainly not – and the authors explicitly caution that the exercise does not replace experimental target validation.
A benchmark that can move
Perhaps the most consequential feature of LongevityBench is less glamorous than autonomous discovery: it is open. The datasets, trained models, evaluation code and Longevity Claw have been publicly released, allowing other researchers to test models against the same biological problems and extend the benchmark.
That creates the possibility of measuring progress in AI for aging against biological competence rather than eloquence – and of watching the benchmark evolve as models improve. In longevity research, where the datasets are getting larger and the questions stubbornly remain larger still, that may prove a useful discipline.
[1] https://pubmed.ncbi.nlm.nih.gov/42753698/
[2] https://onlinelibrary.wiley.com/doi/10.1111/acel.14017
