AI BriefingByteDance Seed paper reveals phase sensitivity in DeepSeek V4 KV-cache compression causes periodic retrieval failures
·2026.11.13 16:50
Key point
A ByteDance Seed paper reveals that DeepSeek V4's chunked KV-cache compression creates periodic retrieval failures, with base model accuracy varying by 40.2 percentage points depending on token position.
Details
- ByteDance Seed paper (arXiv 2609.36322) identifies "phase sensitivity" in DeepSeek V4's chunked KV-cache compression, where retrieval accuracy oscillates based on a token's offset relative to the compression window start.
- In a code completion test, V4-Flash-Base answered incorrectly with 71.3% probability when prompt length left a remainder of 0 or 1 modulo 4, versus 91.5% correct probability at remainders 2 or 3.
- Base model accuracy varied by 40.2 percentage points across phases in 128K needle-in-a-haystack tests. After post-training, the spread narrows to 19.1 points on Flash-0731, 14.8 on Pro-0813, and 6.1 on V4.1-Flash-0910.
- Serving-side variables including checkpoint drift, quantization differences (fp4 vs fp8), and reasoning budget allocation contribute to perceived model degradation, with phase sensitivity acting as a deterministic multiplier.
- Recommended mitigation strategies include pinning specific model IDs, placing critical constraints at the end of prompts, and avoiding decorative formatting like long runs of equals signs.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
ggml v1.0.0 Released with FP8 Support and Backend Optimizations