Tag

Frontier Models

1 take tagged with this topic.

May 15, 2026 at 7:25am ET

25% document corruption. Even from the best AI models.

Even the best frontier AI models corrupted roughly a quarter of a document during a 20-step delegated workflow. Across all models tested, average degradation was 50%.

Source: VentureBeat / Microsoft Research

AI Productivity Frontier Models

DELEGATE-52 benchmark table from Microsoft Research showing round-trip relay results for 19 LLMs across workflow lengths from 2 to 20 interactions. All models accumulate errors; even the strongest (Gemini 3.1 Pro) retains only ~81% document integrity after 20 steps, while weaker models drop below 20%.