AnalysisAI ModelsOctober 9, 2026

ByteDance Seed finds cause of DeepSeek long-context retrieval gaps

Read original source →technode.com

Chunked KV-cache compression makes retrieval accuracy swing by up to 40 percentage points depending on where information falls in a compression window, a pattern ByteDance Seed calls "phase sensitivity." The team reproduced it in models trained from scratch and found different attention components specialize in different positions.

1 source

More stories today

Open the live feed