Skip to content
NewA calmer, faster ZeroHour
StoryAI researchinfo

EngramEdit Updates LLM Facts in Conditional Memory

2 sources · 2 articlesFirst reported Updated ()2 · 1 read
What's new

First coverage.

Merged summary · grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

EngramEdit revises facts in DeepSeek Engram conditional memory while leaving the Transformer fixed, with near-perfect success.

EngramEdit is a method for updating facts stored in conditional n-gram memory, such as DeepSeek Engram, without changing the Transformer backbone. It computes target memory representations so a revised fact is predicted across multiple expressions, then jointly edits shared n-gram embeddings while penalizing changes to frequently reused ones. Experiments report near-perfect edit success, use of revised knowledge on unseen expressions and in multi-hop reasoning, and nearly three times the strongest baseline's accuracy under chain-of-thought prompting. Unrelated knowledge is largely preserved; the later arXiv report also states that general capabilities remain largely intact as edits accumulate. The two sources agree on the method and headline results and do not report conflicting figures.

Key facts
  • EngramEdit updates factual knowledge in conditional n-gram memory such as DeepSeek Engram while the Transformer backbone stays fixed.
  • It computes target memory representations so a revised fact is predicted across multiple expressions.
  • Shared n-gram embeddings are updated jointly, with stronger penalties on frequently reused embeddings to limit collateral change.
  • Edited facts transfer to unseen phrasing and multi-hop chain-of-thought reasoning.
  • Reported edit success is near-perfect and nearly three times the strongest baseline's accuracy under chain-of-thought prompting.
  • Unrelated knowledge is largely preserved; the later arXiv report also says general capabilities remain largely intact as edits accumulate.
  • Hugging Face daily papers dated the item 2026-10-06; the arXiv cs.AI/cs.LG/cs.CL listing is 2026-10-07.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat
Organizations
DeepSeek
AI models
DeepSeek Engram

Coverage

oldest first · one row per article · UTC
  1. · 19d ago Hugging Face daily papersFirst report

    EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

    EngramEdit updates facts in DeepSeek Engram conditional memory while keeping the Transformer backbone fixed.

  2. · 18d ago arXiv cs.AI / cs.LG / cs.CL

    EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

    EngramEdit updates LLM facts in conditional memory while keeping the Transformer backbone fixed.