← News

Models & toolsSources: arXiv · 19 Nov 2026

MoE-Context paper masks unneeded context blocks in a single KV cache

An arXiv preprint describes a system that hides irrelevant context blocks behind a mask before each step, keeping everything recallable instead of deleted.

In plain words

A research preprint proposes that language models should not reread their entire conversation history at every step. Instead, the context is split into blocks, and a quick pass hides the blocks not needed for the current step. Hidden blocks stay in memory as short notes and can be brought back, so nothing is lost.

What was announced

The authors state that AI agents typically read their whole history at every step, whereas people keep only what the moment needs. In their design, the context is divided into blocks, and before each step a fast yes-or-no pass decides which blocks that step requires. The rest remain as one-line notes that can be recalled later, and nothing is deleted. In the version they built, one model both makes the selection decision and answers, working over a single key-value cache in which hidden blocks are masked rather than removed.

The reported evidence is limited to a laptop running Qwen3-1.7B on a synthetic lookup task, where showing only the needed block raised accuracy from 93 to 119 out of 120, and a masked block was as hidden as a deleted one. The authors identify the model's own block picking as the weak part: a supervised linear probe written into spare output rows lifted its accuracy from 80 to 110 of 120, but simple keyword search picked perfectly on this task, and the probe did not transfer to a second task. They also describe two cheap usage patterns the approach enables: undoing a failed attempt and running subagents without a new context.

What it does not show

  • Evidence comes only from the paper's abstract, which reports a single synthetic lookup task on one small model (Qwen3-1.7B); no independent confirmation or broader task results are available.
  • The selection probe did not transfer to a second task, and simple keyword search matched or beat it on the tested task, so the selection mechanism's general value is not established.

For security and IT teams

Scope
The reported results cover a synthetic lookup task on a laptop with Qwen3-1.7B, not production agent workloads.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.