Teaching a large language model new knowledge currently has two mainstream paths: stuff documents into the context window, or fine-tune the weights. The first is expensive and forgetful; the second touches billions of parameters, so writing in one new fact may quietly push out three old ones. A single-author paper submitted to arXiv on October 22, titled Delta N-gram, offers a third path: leave the weights alone and patch the model native n-gram memory table with a detachable overlay.
No Weight Edits, Just a Table
The author earlier analysis found that the model native hashed n-gram table is where continual updates disturb the base model least, but previous writes required backpropagation and modified the pretrained table itself. Delta N-gram removes both constraints: the pretrained table is never modified; new evidence goes into a sparse patch that only covers the rows it addresses. Writing is done by a 4.1M-parameter writer that emits a target and a gate per position-hash-head event, while a block-averaged delta rule keeps each row a convex combination of its old value and the gated targets. The writer is trained offline on raw text with the pretraining objective; at use time only forward passes are needed, and zeroing the patch instantly restores the original model.
In one sentence: knowledge updates shift from retraining a layer to editing a rollback-able table.
The Numbers: 16/16 Versus 2-4/16
The comparison is striking. A 0.6B prototype trained to answer questions keeps 16/16 facts after a second write, while online gradient writes retain only 2-4/16 in the same setting. On Qwen3.8-Flash-Next, the plain writer memory is document-specific: across 64 held-out news events it lowers continuation NLL by 0.068 nats/token versus empty memory and by 0.024 versus another document memory, still at 0.063 and 0.013 after 16 intervening articles. The delta writer, which reads its current memory while encoding, is more resistant to distraction, retaining 70% of its memory versus 54% for the plain version.
Stability is covered too: over 35,000 sequential writes on a seven-task chain, both writers remain stable, with later-task accuracy changing by only -0.26 and -0.52 percentage points, against -7.29 points for the gradient-update control group. Base-model drift is likewise smaller.
Cold Water Included
The paper does not hide the flaws: acquisition is negative, meaning more writes do not strengthen general capability; format rows read by every sample grow to 4-12 times their pretrained norm, causing up to 32.8% drift on general inputs. The redeeming part is interpretability: zeroing eight such rows removes over 98% of layer-2 drift on math inputs, together with their NLL gain. Controllable and diagnosable is worth more than a headline score.
Scale deserves sobriety as well: the prototype is 0.6B, main validation runs on Qwen3.8-Flash-Next, and production-grade models remain far away. The paper is 33 pages, the code is open-sourced, and the author list has exactly one name.
So What
If detachable memory holds up, the minimal unit of model updates drops from a training run to a single forward write, with rollback cost reduced to zero. The next question is sharper: when patches can be applied and cleared at any time, does the boundary of what a model knows live in its weights, or in its table? (Paper)