---
id: 20261029-T0-04
title: "DeepSeek Engram条件记忆层量化掉点：去掉它PopQA直降36.8分"
title_en: "Quantizing DeepSeek's Engram Memory Layer Costs 36.8 PopQA Points"
url: https://ai.daily.yangsir.net/daily/20261029-T0-04
issue_date: 2026-10-29
publish_date: 2026-10-28T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.35985
---

# DeepSeek Engram条件记忆层量化掉点：去掉它PopQA直降36.8分

arXiv新论文研究条件记忆层的量化问题。以DeepSeek的Engram为例，这类层为语言模型挂载大型哈希n-gram表，从DeepSeek-V4.1-Flash中移除后PopQA下降36.8分。低比特量化会优先丢失长尾事实，标准补救方案效果有限。论文提出"让门控决定"的思路，为每次查表单独设定精度，以在压缩记忆层的同时保住长尾知识。

## English Version

**Quantizing DeepSeek's Engram Memory Layer Costs 36.8 PopQA Points**

A new arXiv paper examines quantization of conditional memory layers such as DeepSeek's Engram, which attach large hashed n-gram tables to language models. Removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the standard remedy falls short. The paper proposes letting the gate decide per-lookup precision, aiming to compress these memory layers while preserving long-tail knowledge.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.35985)

**详情页**：https://ai.daily.yangsir.net/daily/20261029-T0-04

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*