EngramEdit: near-perfect fact editing in LLMs via conditional memory, 3x baseline CoT accuracy
teortaxesTex · x · 2026-10-09
arXiv paper 2610.10533 proposes EngramEdit, which decouples factual knowledge updates from the Transformer backbone by editing the embeddings of conditional memory architectures (n-gram lookup tables, as in DeepSeek Engram). Key challenge: different phrasings of a fact activate different n-gram embeddings, and shared embeddings risk collateral changes. The method first computes target memory representations so the model predicts the updated fact across multiple expressions, then jointly updates shared embeddings while penalizing edits to frequently reused ones. Experiments show near-perfect editing success, generalization to unseen expressions and multi-hop reasoning, and nearly 3x the strongest baseline's accuracy under chain-of-thought prompting. The author frames it as a new axis of PEFT.
Related event: EngramEdit Enables Decoupled Fact Editing in LLMs via Conditional Memory(3 posts)→
More from Models
- Claim that "Gemini is just an agent harness adding Claude models" called out as misleading — prajdabre · 2026-10-09
- Claude claims discovery of new binary red dwarf pair ~500 light-years away — nitarshan · 2026-10-09
- Training a personal persona model on 11GB+16GB GPUs from your full internet history — ailee43 · 2026-10-09
- Reddit users complain all closed models got simultaneously dumber — No_Vehicle7826 · 2026-10-09
- Users complain OpenAI is silently swapping models mid-chat even on the Pro plan — Slow_Ad1827 · 2026-10-09
- GPT-6 reportedly rolling out with Intelligent UI — answers embed interactive tools — winer666 · 2026-10-09