Are LLM-recalled homologs memorized or generated? UniRef100 comparison says mostly novel
sokrypton · x · 2026-11-20
Part 4 of sokrypton's thread investigates whether Claude Fable 5.2's recalled homologs are pulled from UniProt, memorized, or newly generated. Comparing 500 recalled sequences per family (YaaA, CheY, thioredoxin, DHFR) against UniRef100: median identity to the closest real sequence is 68–75%, only 0–8% exceed 95% identity, and 0–1% are exact matches — indicating the model generates novel but plausible sequences rather than reciting the database.
More from Research
- Image Models Now Ship Fine-tuned LLM Prompt Upsamplers — Why Not Fuse Them? — ostrisai · 2026-11-20
- PyTorch's ezyang on formalizing a trivial in-place kernel under PTX semantics — ezyang · 2026-11-20
- Scientist: Gemini 4 Neon designed innovative CAR receptors to kill cancer cells — DeryaTR_ · 2026-11-20
- New arXiv paper: Bregman power divergences turn KL queries into Euclidean NN in d+1 — FrnkNlsn · 2026-11-20
- Untrained Neural Networks Hit 86.4% AUROC on ADBench Anomaly Detection — ideal-lab-sejong · 2026-11-20
- NAVER's Unfinished-Bench: Persistent Agents Break Hidden Rules 43.6% of the Time — naver-ai · 2026-11-20