AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Cartridges: Lightweight and general-purpose long context representations via self-study

Sabri Eyuboglu, Ryan Saul Ehrlich, Simran Arora, Neel Guha, Dylan Zinsley, Emily Ruoyu Liu, Atri Rudra, James Zou, Azalia Mirhoseini, Christopher Re

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 69 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Initialize with real token KV vectors instead of random vectors (ablation)Cartridge · LongHealthaccuracy: 29.9% -> 51.3%Improved
Initialize with first p corpus tokens rather than random tokens (ablation)Cartridge · LongHealthaccuracy: 51.3% -> 55.3%Improved
Sample from five seed prompt types instead of one (ablation)Self-Study · MTOBchrF: 24.1 -> 32.0Improved
Sample from five seed prompt types instead of one (ablation)Self-Study · LongHealthaccuracy: 43.6 -> 48.4Improved
Sample from five seed prompt types instead of one (ablation)Self-Study · QASPERperplexityNo clear effect
Use context distillation instead of next-token prediction (ablation)Self-Study · MTOBchrF: 24.9 -> 33.5Improved
Use context distillation instead of next-token prediction (ablation)Self-Study · LongHealthaccuracy: 47.6 -> 51.3Improved
Use prefix-tuning instead of LoRA (ablation)Cartridge · MTOBchrF: +4.5 chrFImproved
Use prefix-tuning instead of LoRA (ablation)Cartridge · LongHealthaccuracy: 47.25% -> 55.6%Improved
Use prefix-tuning instead of LoRA at large size (ablation)Cartridge · MMLU (out-of-domain)accuracy: 45.3 -> 54.3Improved
Freeze the attention sink (first token KV) (ablation)Cartridge · MMLUaccuracy: accuracy dips below 30% when trainableImproved
Compose two independently trained CartridgesCartridge · multi-document 10-K questionsaccuracyImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy