frontier models
Facts Moved Into an Editable Memory Layer
EngramEdit changed conditional n-gram memory while leaving the Transformer backbone fixed and penalizing broadly shared embeddings.
Summary
EngramEdit changed conditional n-gram memory while leaving the Transformer backbone fixed and penalizing broadly shared embeddings.
Conditional-memory models look up learned embeddings from input n-grams, creating a possible place to update facts without rewriting a model's general computation. EngramEdit first finds target memory representations that produce the new fact across several phrasings, then jointly updates the shared embeddings while penalizing changes to frequently reused entries. The authors report near-perfect edit success, transfer to unseen expressions and multi-hop reasoning, and almost three times the strongest comparison's chain-of-thought accuracy. Unrelated knowledge and general capabilities were largely preserved as edits accumulated in the reported tests. That does not settle open-world model maintenance, but it turns a scaling component into a concrete interface for decoupled knowledge updates.
Why it matters
EngramEdit changed conditional n-gram memory while leaving the Transformer backbone fixed and penalizing broadly shared embeddings.
Limits and context
- That does not settle open-world model maintenance, but it turns a scaling component into a concrete interface for decoupled knowledge updates.
Key claims
EngramEdit changed conditional n-gram memory while leaving the Transformer backbone fixed and penalizing broadly shared embeddings.
Qualification: That does not settle open-world model maintenance, but it turns a scaling component into a concrete interface for decoupled knowledge updates.
Evidence: source-2026-10-08-006
Sources
- arXiv preprint 2610.10533arXiv · primary research
Corrections
No corrections have been recorded for this story.