InfoFlow KV: Information-Flow-Aware KV Recomputation for Long Context
Xin Teng, Canyu Zhang, Shaoyi Zheng, Danyang Zhuo, Tianyi Zhou, Shengjie Wang · 2026
Read InfoFlow KV: Information-Flow-Aware KV Recomputation for Long Context by Xin Teng, Canyu Zhang, Shaoyi Zheng, Danyang Zhuo, Tianyi Zhou, Shengjie Wang (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-2603-05353.txt.
Description
Xin Teng and the rest of the team cast selective recomputation as information flow. Long-context answering bottlenecks on prefilling large contexts. Precomputed per-document caches with a small recomputed token subset restore causal dependencies, but current methods use heuristics without building token influence. A simple attention-norm query signal spots relevant, well-positioned tokens for propagating information, under RoPE shape. Global positions and information-flow-guided chunk reordering follow. Benchmarks show gains under comparable latency.
Catalog record
| CREATOR | Xin Teng, Canyu Zhang, Shaoyi Zheng, Danyang Zhuo, Tianyi Zhou, Shengjie Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43064 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.05353