Aging clocks have become an indispensable tool in gerontology, yet each one is tied to a single type of biological data and a fixed feature set. Adding a new modality means building and validating a separate model, and interpreting the results is left to the researcher's judgment. Neighboring papers in the archive follow the path of specialized predictors: some analyze the methylation landscape, others link aging to loss of meaning in biological texts, still others build maps of brain age. Longevity-LLM stands apart because instead of yet another narrow tool it offers a single language model that absorbs knowledge from many clocks and talks to the researcher in natural language.
The authors assembled a corpus of 766,640 training examples (1.09 billion tokens) covering 38 tasks across four modalities: DNA methylation, clinical biomarkers, transcriptomics and proteomics. The data were converted into structured text prompts with reasoning, then the base model Qwen3-14B underwent two-stage training. The first stage was supervised fine-tuning on 4.5 billion tokens; the second was reinforcement learning on seven regression tasks, where the reward depended on the accuracy of the numerical prediction. Training ran on 24 NVIDIA B200 GPUs, then on 16 A100s.
After training, Longevity-LLM took first place in four of the seven Longevity Bench tasks, beating 15 commercial LLMs. In predicting 10-year mortality from NHANES, the model reached an accuracy of 0.89, and in forecasting cancer survival from TCGA data, 0.77. For predicting age from DNA methylation, the mean absolute error was 4.34 years, which is statistically significantly better than the Horvath clocks (4.61 years). The model also predicted age from proteomic data with an error of 7.9 years and generated plausible proteomic profiles with a Jaccard index of 0.072, outperforming all commercial counterparts.
The work shows that a standard text transformer architecture, with no modifications, can absorb the aging signal from raw omics data. Unlike neighboring studies that build separate models for each data type, here a single model unites several modalities and can answer questions in natural language. This removes the barrier between specialized omics foundation models and conversational LLMs, opening the way to an internally consistent representation of molecular aging.
The next iteration, Longevity-LLM v0.2, will aim for the model not only to predict age but also to explain the biological logic behind its predictions: linking molecular features to pathways and phenotypes. There are plans to strengthen reinforcement learning to get around the bottleneck of manually curated reasoning data. In the long run, such a model could become a companion to the aging researcher rather than just a prediction endpoint.
The approach developed here, distilling knowledge from specialized models into a single language model, applies to any area of biomedicine. In pharma, it is a step toward a monolithic pipeline from prompt to drug, one that would combine target identification, molecule design and clinical trial planning.
The immediate plans include expanding the training corpus, adding new benchmark tasks and scaling up the reinforcement learning stage. The team will also test the hypothesis of cross-modal generalization: whether knowledge from one modality really does improve predictions in another.
For gerontology, the question remains unresolved: why does a patient's biological age exceed their chronological age, and which molecular mechanisms are behind it. In its current version, Longevity-LLM only predicts age but does not explain it, so the core problem of aging clock interpretability persists. Future iterations of the model will need to learn to link deviations from the expected aging trajectory to specific biological pathways.
Before training, the base model Qwen3-14B could not produce a single valid prediction on any task, because it did not know how to format its answers. After fine-tuning, it outperformed commercial systems with hundreds of billions of parameters.
The idea of a single AI that reads biological data and explains aging echoes the medical AI from the series Altered Carbon, where consciousness is transferred between bodies and biological age is controlled by technology.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.