# BEAVER: A Training-Free Hierarchical Prompt Compression Method via Structure-Aware Page Selection

Source: [BEAVER: A Training-Free Hierarchical Prompt Compression Method via Structure-Aware Page Selection](<https://www.sota2.com/research/paper/69c0c8d70636e5bf4afc394a>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-26T06:00:01.930Z

Authors: Zhengpei Hu, Kai Li, Dapeng Fu, Chang Zeng, Yue Li, Yuanhao Tang, Jianqiang Huang

Published: 2026-03-20T04:35:04.000Z

## Abstract

The exponential expansion of context windows in LLMs has unlocked capabilities for long-document understanding but introduced severe bottlenecks in inference latency and information utilization. Existing compression methods often suffer from high training costs or semantic fragmentation due to aggressive token pruning. In this paper, we propose BEAVER, a novel training-free framework that shifts compression from linear token removal to structure-aware hierarchical selection. BEAVER maximizes hardware parallelism by mapping variable-length contexts into dense page-level tensors via dual-path pooling, and preserves discourse integrity through a hybrid planner combining semantic and lexical dual-branch selection with sentence smoothing. Extensive evaluations on four long-context benchmarks demonstrate that BEAVER achieves comparable performance to state-of-the-art (SOTA) methods like LongLLMLingua. Notably, on the RULER benchmark, BEAVER maintains high fidelity in multi-needle retrieval where baselines deteriorate. Regarding efficiency, BEAVER reduces latency by 26.4x on 128k contexts, offering a scalable solution for high-throughput applications. Our code is available at https://cslikai.cn/BEAVER/.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Prompt Compression on L-Eval (test)](<https://www.sota2.com/research/sota/prompt-compression-on-l-eval-test.md>): Prompt Compression; L-Eval (test); Coursera QA Accuracy: 64.4.

- [Long-context Question Answering on L-Eval](<https://www.sota2.com/research/sota/long-context-question-answering-on-l-eval.md>): Long-context Question Answering; L-Eval; Coursera QA: 28.3.

- [Long-context retrieval and reasoning on RULER 16k context 3k budget](<https://www.sota2.com/research/sota/long-context-retrieval-and-reasoning-on-ruler-16k-context-3k-budget.md>): Long-context retrieval and reasoning; RULER 16k context 3k budget; S-1 Score: 100.

- [Zero-shot long-context reasoning on ZeroSCROLLS](<https://www.sota2.com/research/sota/zero-shot-long-context-reasoning-on-zeroscrolls.md>): Zero-shot long-context reasoning; ZeroSCROLLS; Average Score: 32.4.

- [Long-context Language Understanding on LongBench](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench.md>): Long-context Language Understanding; LongBench; M-Avg: not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.19635>)

[PDF](<https://arxiv.org/pdf/2603.19635.pdf>)

[Implementation](<https://github.com/JusperLee/BEAVER>)