AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

MeSH: Memory-as-State-Highways for Recursive Transformers

Chengting Yu, Xiaobo Shu, Yadao Wang, Yizhen Zhang, Haoyi Wu, Jiaang Li, Rujiao Long, Ziheng Chen, Yuchi Xu, Wenbo Su, Bo Zheng

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse efficient transformers for long context · 6 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
add MeSH memory buffer and routers to recursive transformerPythia-1.4B recursive (4+8R2+4) · Task average 0-shotaverage accuracy: 48.89 -> 50.56Improved
add MeSH to recursive transformerPythia-1.4B recursive (4+8R2+4) · Pile validationperplexity: 7.63 -> 7.39Improved
add MeSH to recursive transformerPythia-1B recursive (3+5R2+3) · Pile validationperplexity: 8.20 -> 7.90Improved
add MeSH to recursive transformerPythia-160M recursive (2+4R2+2) · Pile validationperplexity: 11.79 -> 11.37Improved
add MeSH to recursive transformerPythia-410M recursive (3+6R3+3, 50% fewer params) · Task average 0-shotaverage accuracy: 41.94 -> 43.53Improved
add MeSH to recursive transformerPythia-410M recursive (3+6R3+3, 50% fewer params) · Lambada Standardperplexity: 81.75 -> 52.07Improved
add anchor connection to recursive transformerPythia-1.4B recursive · Task average 0-shotaverage accuracy: 48.89 -> 49.39No clear effect
add residual connection to recursive transformerPythia-1B recursive · Task average 0-shotaverage accuracy: 45.72 -> 46.19Improved
add residual connection to recursive transformerPythia-410M recursive (3+6R3+3) · Pile validationperplexity: 9.65 -> 9.69Hurt
use MeSH recursive model instead of larger VanillaPythia-1.4B · Task average 0-shotaverage accuracy: 49.50 -> 50.56Improved
use MeSH recursive model instead of larger VanillaPythia-160M · Task average 0-shotaverage accuracy: 39.88 -> 39.41Hurt
add MeSH to recursive transformerPythia-1.4B recursive · Task average 5-shotaverage accuracy: 50.99 -> 52.79Improved

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