2026-09-19
LongevityBench scores LLMs on 17 aging-omics tasks; compact 0.6B–9B Longevity-LLMs outrank Gemini 3.1 Pro, while omics age prediction stays hard at any scale.
Aging research sits on DNA methylation, transcriptomes, plasma proteomes, clinical blood panels, and genetic perturbations, yet no shared test asked a language model a blunt question: given a primary measurement, can you reach a phenotypic conclusion. Most biomedical benchmarks reward literature recall or sequence-to-function mapping, where the answer often already lives in pretraining text. Aging clocks return a number, but they are stuck on one modality, a frozen feature set, and no conversation.
Insilico Medicine, with Liquid AI, the Buck Institute, and Harvard Medical School, put LongevityBench on the cover of Cell: 17 tasks, five biodata domains, 25,457 prompts. They also release five Longevity-LLMs from 0.6B to 9B parameters, plus Longevity Claw, an agent interface that wires the models to enrichment and clock tools.
Tasks cover NHANES clinical records, GEO methylation, GTEx transcriptomes, Olink plasma proteomics, and genetics from OpenGenes and SynergyAge. The same underlying records are rendered as binary classification, pairwise comparison, multiclass, and regression, so format effects can be separated from domain skill. Evaluation splits are cut by survey wave, GEO series, subject, protein family, or species. The 17 ranked tasks were held out of fine-tuning.
Prompts are programmatically built structured text, not paper abstracts. A methylation item can exceed 26,000 tokens at the median. OpenGenes masks the gene as gene-X and supplies function and tissue context, which blocks a lookup shortcut.
Eighteen frontier systems from six labs were queried. o3 and o4-mini delivered parseable answers on under 80% of prompts and were dropped from the main comparison, leaving 16 frontier models, five L-LLMs, and five base checkpoints, 26 systems in total. The headline score is mean rank across the five biodata domains with equal weight, so a task-heavy domain cannot dominate.
L-Qwen3.5-9B was LoRA-tuned on Qwen3.5-9B in two stages, seeing about 4B tokens of which 1.4B were unique, on eight RTX PRO 6000 GPUs. The other four members, Qwen3-0.6B and 1.7B plus LFM2-1.2B and 2.6B, trained on the same collection. LFM2 variants trained on domain data alone, mixed with 10% chat, and mixed with 20% chat were then merged at equal weight so conversational skill would not collapse.
No frontier model won every task. Omics age prediction was the hardest, and scale did not rescue it. Gemini-3.1-Pro led the commercial pack, but its 95% CI of 6.7 to 9.5 overlapped Claude Opus-4.6, Opus-4.5, Sonnet-4.6, and Kimi-K2.5, so there is no separable frontier champion. On compact clinical prompts, elastic-net and Cox baselines still sit above most LLMs. On GEO methylation pairwise comparison, naive Bayes reached concordance 0.61 and the best frontier model, Kimi K2.5, only 0.69. On GTEx pairwise age, no frontier model exceeded 0.56. On Olink pairwise age a purpose-built protein clock hit 0.749; Gemini-3.1-Pro reached 0.703, and GPT-4.1 was near chance.
Once the L-LLMs enter, they take the top three ranks. L-Qwen3.5-9B sits at 4.4 (95% CI 3.5 to 5.1), overlapping nobody, and is the only system that significantly beats its runner-up on three tasks: GEO methylation pairwise and regression, and GTEx pairwise. L-Qwen3-0.6B placed sixth of 26, ahead of Claude Opus-4.5, Kimi-K2.5, GPT-5.2, and Grok-4.3.
| Metric | L-LLM | Comparator |
| Aggregate rank | L-Qwen3.5-9B at 4.4 | Best frontier Gemini-3.1-Pro, CI 6.7–9.5 |
| Olink age MAE | L-Qwen3-0.6B at 5.7 years | Best frontier Kimi-K2.5 at 10.1; OrganAge clock 6.8; same-panel elastic net 3.9 |
| Olink generation Jaccard | L-LFM2-2.6B at 0.17 | Best frontier GPT-4.1-mini at 0.04 |
| GTEx generation Jaccard | L-LFM2-2.6B at 0.35 | Best frontier Gemini-2.5-Pro at 0.15; age-blind frequency baseline 0.54 |
| CellAge senescence role | L-LLM mean 71.4% | Gemini-3.1-Pro 95.9%, GPT-5 94.9% |
Ablations point to distributed biological signal: dropping the entire NHANES blood panel flipped 47.4% of previously correct “who is older” answers. Generation tasks were demoted to ancillary metrics because they mostly reward familiarity with a measurement panel, not age inference.
Longevity Claw asked L-Qwen3.5-9B for 30 gene targets in each of 14 hallmark categories, six independent runs, 328 unique candidates after dedup. Twelve overlapped the 300 aging targets from Pun et al. 2022, a 2.4-fold enrichment (p=0.004). CXCL12 and ADAMTS14 sat among that paper's nine experimentally validated genes. The authors label this an illustration, not a validated discovery pipeline.
AI-for-science groups get a reproducible ruler: single turn, no tools, primary measurements in the prompt. The hard claim is that on omics age, domain data beats scale. A 0.6B model that can run locally is the more useful artifact for labs that cannot send patient rows to an external API.
Claw's target list tastes like existing AI target tools; overlap plus enrichment is not a new drug target. The CellAge control makes the split explicit: frontier models recite published senescence roles, then fall behind once the input is raw omics.
The suite scores prediction and format robustness, not mechanistic reasoning or causality. Contamination cannot be ruled out for commercial models with closed corpora; the authors did not empirically test text leakage on the omics tasks. Splits are within-domain. Cross-domain reuse of genes, CpGs, or subjects was not audited. Chemical screens, cross-species translation, epigenetic reprogramming, and single-cell assays are out of scope. Cost, latency, and fairness across sex, age, and ancestry were not measured. Train and eval still come from the same public sources, so independence is incomplete even with a clean split.
SynergyAge pairwise comparison has 26 prompts, so genetic-perturbation ranks are noisy. L-LLMs approach specialist clocks on a fixed feature panel and lose that edge when features rotate, which means fine-tuning did not produce a general “read any CpG set” skill.