2026-09-18
Insilico and Liquid AI release LongevityBench (17 tasks, 25,457 prompts) in Cell; a 9B domain model takes aggregate rank 4.4, ahead of Gemini 3.1 Pro and 15 other frontier systems.
Aging biology already measures people across DNA methylation, transcriptomes, plasma proteomes, and clinical labs. The software stack does not. Horvath-style clocks eat CpG beta values. Proteomic clocks eat an Olink panel. A transcriptomic model retrains when the tissue changes. A general-purpose LLM could, in principle, verbalize any of those tables and keep one set of weights. Nobody had stress-tested that claim on real records.
Quiz-style biology questions measure literature recall. What a wet-lab group actually needs is a model that reads a sample's omics profile and returns age, mortality risk, or the direction of a gene's age effect. This Cell paper, led by Insilico Medicine with Liquid AI, the Buck Institute, and Harvard Medical School, turns that requirement into an open benchmark and a family of 0.6B-9B domain models.
LongevityBench rewrites public datasets as ChatML prompts that fit a 30K-token window. The official suite is 17 tasks and 25,457 prompts across five domains: NHANES clinical records, GEO DNA methylation, GTEx transcriptomes, Olink plasma proteomics, and genetic evidence from OpenGenes, CellAge, and SynergyAge. The same underlying records are asked four ways (binary, pairwise, multiclass, regression) so format skill can be separated from data skill.
Splits follow biological units, not random rows: NHANES by survey wave, methylation by GEO study, GTEx and Olink by donor, OpenGenes by protein family, SynergyAge by species. The 17 evaluation tasks were held out of fine-tuning.
Eighteen frontier systems from six labs were queried at temperature 0. o3 and o4-mini delivered parseable answers on only 0.77 and 0.70 of prompts, below the 80% cutoff, so they left the main table. The roster is 26 systems: 16 frontier models, five Longevity-LLMs, and five untuned bases. Invalid replies were dropped from the denominator in the main text; an invalid-as-wrong ranking sits in the supplement.
Five compact models (0.6B-9B) were then fine-tuned on the same aging corpus. L-Qwen3.5-9B is LoRA on Qwen3.5-9B, a hybrid that interleaves Gated DeltaNet linear attention with standard self-attention at 3:1. Training saw 1.4B unique tokens and 4B tokens in total, two LoRA stages, LLaMA-Factory plus DeepSpeed ZeRO-2 on eight RTX PRO 6000 GPUs. The other four (L-Qwen3-0.6B/1.7B, L-LFM2-1.2B/2.6B) used full-parameter supervised fine-tuning. Reasoning traces came from a Qwen3-235B-A22B teacher that never saw the gold answer. LFM2 trained three mixtures (longevity only, plus 10% chat, plus 20% chat) and merged them at equal weight so dialogue skill would not collapse.
Longevity Claw wraps L-Qwen3.5-9B with MSigDB enrichment, published aging clocks, and cohort profiling. Prompts covered the 12 hallmarks of aging plus two extra categories, each asking for 30 novel druggable genes, repeated independently six times.
Lower aggregate rank is better. L-Qwen3.5-9B posted 4.4 (bootstrap 95% CI 3.5-5.1), an interval that does not overlap any other model. The next two slots are also L-LLMs: L-LFM2-2.6B at 7.6 and L-Qwen3-1.7B at 7.8. The strongest frontier system, Gemini-3.1-Pro, sits at 8.2 (CI 6.7-9.5), overlapping Claude Opus-4.6, Opus-4.5, Sonnet-4.6, and Kimi-K2.5. There is no lone commercial winner. The 0.6B L-Qwen3 placed sixth of 26, behind only Gemini-3.1-Pro and Claude Opus-4.6, and ahead of GPT-5.2, Grok-4.3, and Kimi-K2.5.
| Model | Metric | Result |
| L-Qwen3.5-9B | Aggregate rank | 4.4 (CI 3.5-5.1) |
| Gemini-3.1-Pro | Aggregate rank (best frontier) | 8.2 (CI 6.7-9.5) |
| L-Qwen3.5-9B | GEO DNAm pairwise concordance | 0.868 vs 0.685 best frontier |
| L-Qwen3-0.6B | Olink proteomic-age MAE | 5.7 yr vs 10.1 yr best frontier |
| L-Qwen3-0.6B | NHANES 10-year mortality balanced accuracy | 0.890 |
L-Qwen3.5-9B was the only system to beat its runner-up on three tasks: GEO DNAm pairwise, GEO DNAm regression, and GTEx pairwise age. Scale is not monotone. L-LFM2-2.6B outranked the 9B model on five tasks. On masked OpenGenes binary classification, L-LFM2-1.2B reached 78.0% versus 71.0% for the 9B model (McNemar p=0.004).
Sample-level age regression stays hard. A purpose-built proteomic clock hit 0.749 concordance on Olink pairwise age; the best frontier model, Gemini, reached 0.703, and GPT-4.1 was near chance. L-LLMs cut proteomic-age MAE to 5.7 years, roughly matching that specialist clock, while a de novo supervised fit on the same features still won. On plasma-protein generation, frontier Jaccard scores sat near 0 and covered about 47% of the protein vocabulary; L-LFM2-2.6B reached mean Jaccard 0.17, 2,583 unique proteins, and 91% coverage.
A knowledge-retrieval control splits the two skills. Gemini-3.1-Pro scored 95.9% and GPT-5 94.9% on literature-style items; L-LLMs averaged 71.4%, below the frontier mean of 86.0% (Mann-Whitney p=0.0073). Recalling papers is a different job from reading an array.
Longevity Claw emitted 328 unique gene targets. Against the 300 aging targets from Pun et al. 2022 (PandaOmics), 12 overlapped (Fisher p=0.004, 2.4-fold). The druggable subset was 2.8-fold enriched. Two of nine experimentally validated genes in that paper, CXCL12 and ADAMTS14, appeared (p=0.045, 5.6-fold). That is statistical overlap, not a wet-lab confirmation.
AI-for-science groups now have a public ruler that scores whether the omics record in the prompt was actually used, rather than whether the chat voice sounds like a biologist. Domain data plus a 0.6B-9B full-parameter or LoRA pass can beat general models two orders of magnitude larger on structured omics, and the compact weights can stay on-prem when patient-level data cannot leave the lab.
The durable assets are the benchmark and the split rules, not one leaderboard snapshot. Single-task aging clocks are not obsolete: a from-scratch regressor on the same features still wins. The 328 Claw genes are a hypothesis generator, not a target list.
The authors say so themselves: the suite tests prediction and format robustness, not mechanistic reasoning or causal inference; contamination cannot be ruled out for commercial models with opaque corpora; omics tasks were not empirically probed for text leakage; cross-domain reuse of genes, CpG sites, or subjects was not audited; chemical screens, cross-species translation, epigenetic reprogramming, and single-cell assays sit outside the current scope.
A few more discounts apply. The Friedman critical difference is CD=9.6, so mid-pack aggregate ranks are noisy; the claim that holds is the 9B model's non-overlapping CI. Dropping invalid answers from the denominator inflates models that refuse hard prompts. The 5.6-fold enrichment rests on 2 hits among 9 validated genes at p=0.045. Insilico built both the models and the benchmark; external checks depend on the public longebench release. The dataset is CC-BY-NC while the paper is CC-BY, which matters for commercial evaluation.