| title | LongAttnComp: Cross-Family Context Compression for Long-Context Reasoning |
|---|---|
| authors | Mengmeng Ji, Ravi Shanker Raju, Jonathan Lingjie Li, Chen Wu |
| date | 2026-05-31 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2606.01336 |
| affiliation | SambaNova |
| upvotes | 3 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf |
| abstract | As real-world applications increasingly require processing inputs of 100k+ tokens, the gap between context length and inference efficiency has become a critical bottleneck. Context compression offers a way to reduce prefill costs while preserving task accuracy. However, existing training-free attention-based methods leave substantial gaps in demanding long-context tasks such as code reasoning. We present LongAttnComp, a long-context adaptation of AttnComp that fine-tunes a lightweight cross-attention scoring layer and introduces tokenlevel chunking, a token-budget top-p algorithm, positional reordering, and a formatagnostic query parser. We further design a two-stage fine-tuning recipe for the compressor: Stage 1 builds a general retrieval foundation from NIAH-style data, and Stage 2 extends it with multi-hop and reasoning data for broader long-context task coverage. On InfiniteBench Code-Debug, LongAttnComp matches or exceeds full-context accuracy, substantially outperforms training-free baselines, and transfers across four target models from three families. On LongBench v2, the two-stage recipe largely closes the Stage 1 gap on multi-document reasoning while preserving Code-Debug performance. |
cite
@misc{ji2026longattncomp,
title={LongAttnComp: Cross-Family Context Compression for Long-Context Reasoning},
author={Mengmeng Ji and Ravi Shanker Raju and Jonathan Lingjie Li and Chen Wu},
year={2026},
eprint={2606.01336},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2606.01336}
}