2026-05-10 / Signal #1
LLMs Silently Corrupt Documents Over Long Delegated Workflows (DELEGATE-52 benchmark)
8.5Weirdness
Why It Matters
Needs editorial pass.
Evidence
Microsoft Research arXiv paper (Apr 2026, HN frontpage ~400-600 pts in last 24h); frontier models (Claude 4.6, Gemini 3.1, GPT-5.4) corrupt ~25% of content in simulated professional editing across 52 domains (code, crystallography, music notation); errors are sparse/severe, compound with length/distractors; agentic tools don't fix it.
Signal Read
Novelty: 8Receipts: 9Story voltage: 9Heat: 8
Source Trail
Daily scan: 2026-05-10