200M-parameter GPN-Star beats 40B Evo 2 by encoding evolutionary priors
bravo_abad · x · 2026-09-10
Ye and coauthors built evolutionary knowledge directly into GPN-Star: ordinary self-attention for DNA sequence context plus phylogeny-informed cross-attention that weights information from other species by evolutionary distance.
The payoff is striking: with only 200M parameters, GPN-Star outperforms much larger single-sequence genomic language models — including the 40B-parameter Evo 2 — across several coding and non-coding variant-effect benchmarks.
The insight: models shouldn't spend billions of parameters rediscovering evolutionary constraints we already know. Encoding that prior in the architecture can beat scaling.
Related event: GPN-Star hits Nature, beating 40B Evo 2 with 200M parameters(3 posts)→
More from Research
- Open Models Match GPT-6 on MedMCQA but Trail 17 Points on Hard Clinical Cases — keunwoochoi · 2026-10-27
- Health-Focused Benchmark Reranks LLMs: Muse Spark Open Jumps from #9 to #1 — keunwoochoi · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- Goodfire commits $10M to a 'Human Genome Project for AI' to reverse-engineer frontier models — CatAstro_Piyush · 2026-10-27
- π0.5 study: raising demos from 10 to 438 lifts Place Ball in Bowl success from 7% to 83% — stingning · 2026-10-27
- HandsOn benchmark: 66 verified tasks and 18,000 human demos for dexterous manipulation — stingning · 2026-10-27