An open benchmark and language models for AI in aging biology

Alex Zhavoronkov and colleagues introduced LongevityBench, an open benchmark with 17 tasks across five biodata domains, and tested 18 frontier AI systems from six developer teams. They also fine-tuned five aging-focused Longevity-LLMs ranging from 0.6 to 9 billion parameters. No single model dominated every task. Omics-based age prediction was the hardest category regardless of model scale. The compact Longevity-LLMs matched or exceeded much larger general-purpose systems on the benchmark. The work matters as infrastructure for longevity research: it creates a reproducible way to compare models on biological-age prediction and structured aging data rather than relying on general language benchmarks. It does not discover a longevity molecule, validate a biomarker clinically, or show that AI-generated hypotheses improve human health. Benchmark performance can also reflect training-data overlap, task construction, and the choice of metrics. The commercial interests are material. Alex Zhavoronkov founded Insilico Medicine, several authors work there, and the benchmark, model weights, and research interface are linked to commercial drug-discovery activities. The paper is best read as a tool and methods report. Independent replication, held-out datasets, prospective predictions, and experimental validation are needed before model performance is treated as evidence about aging biology.

The compact (0.6B-9B parameters) Longevity-LLMs matched or exceeded far larger frontier systems on LongevityBench, showing that general-purpose language models can be adapted to structured-omics tasks. We publicly release the benchmark, models, and Longevity Claw, an agentic research interface for aging researchers.