Researchers eye open-model training data attribution as LLMs shown memorizing Cas9 and Alu sequences
anshulkundaje · x · 2026-10-23
A discussion sparked by genomics researcher Anshul Kundaje argues that training data attribution on open models would be highly valuable.
LLMs have already been shown to memorize ubiquitous proteins like Cas9 and GFP as well as DNA sequences like Alu and Shine-Dalgarno, raising research questions about memorization vs. generalization and the case for data provenance transparency.
More from Research
- Microsoft's step-by-step diagnostic AI lands Science cover, code released — erichorvitz · 2026-10-23
- Sophont's Holon-S1 predicts genes and biomarkers from routine pathology slides — iScienceLuvr · 2026-10-23
- Glasgow study: raised eyebrows and slow blinks make speakers look confident — Limor_Raviv · 2026-10-23
- IFM open-sources RL360, the agentic RL infra behind K2 Horizon's training — rupspace · 2026-10-23
- Wayfarer discovers RL options from a single experience stream, SOTA on hard Atari games — MarlosCMachado · 2026-10-23
- PhysVidBench accepted at EMNLP 2026: do video gen models grasp physical commonsense? — aykuterdemml · 2026-10-23