A collaborative team of researchers from various institutions has unveiled a groundbreaking paper in the esteemed journal Cell, presenting a comprehensive large model system tailored for aging research. This innovative system is comprised of three pivotal components: the LongevityBench benchmarking platform, which serves as a robust tool for evaluating model capabilities; the Longevity-LLMs series, a collection of large models meticulously trained on aging-related data; and the Longevity Claw research agent, a sophisticated interface that seamlessly connects these large models with cutting-edge aging analysis tools.

LongevityBench, constructed on the foundation of authentic biological data, encompasses an impressive array of 17 distinct categories of aging research tasks. The evaluation outcomes reveal notable disparities in performance among general large models within this specialized domain. Building upon these insights, the research team has engineered five Longevity-LLMs models, each varying in parameter count from a modest 0.6 billion to a substantial 9 billion. Notably, the 9 billion-parameter L-Qwen3.5-9B model has emerged as a standout performer, surpassing even the most advanced general models, such as Gemini 3.1 Pro, in comprehensive assessments and excelling in select specialized tasks.

Through a series of ablation experiments, the researchers have uncovered that the model's decision-making process is heavily reliant on the input biological feature data. However, it currently lacks the capacity to infer causal relationships independently. To address this limitation, the team has ingeniously integrated the L-Qwen3.5-9B model with the Longevity Claw system, enabling it to autonomously explore potential anti-aging targets. Initial screening results have yielded a significant enrichment of known aging-related genes, although further experimental validation is imperative to confirm these findings.

In a move towards fostering transparency and collaboration within the scientific community, the research team has made the benchmarking platform, model architecture, and analysis tools readily accessible to the public.