Context Compression via Explicit Information Transmission
Jiangnan Ye, Hanqi Yan, Zhenyi Shen, Heng Chang, Ye Mao, Yulan He · 2026
Read Context Compression via Explicit Information Transmission by Jiangnan Ye, Hanqi Yan, Zhenyi Shen, Heng Chang, Ye Mao, Yulan He (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03784.txt.
Description
Long-running agents drown in their own context, and compressing it with another model call usually loses too much. Jiangnan Ye and 5 co-authors diagnose structural bottlenecks in those compressors in a 2026 paper. Compression tokens coordinate poorly while gathering information, and signals from middle layers dilute as they pass upward.
Their framework, ComprExIT, transmits information explicitly instead. It selects features adaptively across the frozen model's layers, then moves information from anchors to compression slots under a globally coordinated transport plan. Across twelve datasets it beats strong soft-compression baselines, improving average F1 by up to about eighteen percent. It adds only about one percent trainable parameters and compresses more than twice as fast as the quickest baseline.
Catalog record
| CREATOR | Jiangnan Ye, Hanqi Yan, Zhenyi Shen, Heng Chang, Ye Mao, Yulan He |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-03784