SWE-Pruner is a self-adaptive context pruning framework specifically designed for coding agents. It addresses the challenges of long interaction contexts, such as high API costs and latency, by performing task-aware adaptive pruning.
Inspired by how human programmers selectively skim code, SWE-Pruner enables agents to formulate explicit goals (e.g., "focus on error handling") which guide a lightweight neural skimmer (0.6B parameters). This skimmer dynamically selects relevant lines from the surrounding context, preserving critical implementation details while significantly reducing token usage.
Evaluations across benchmarks show that SWE-Pruner achieves 23-54% token reduction on agent tasks like SWE-Bench Verified and up to 14.84x compression on single-turn tasks like LongCodeQA with minimal performance impact.
Given that we have made significant modifications to the model, its dual-head architecture and the complex compression head service code will be rather complex.
Therefore, we recommend that you use the version we have released on
GitHub instead of attempting to use the original model on your own.
1@misc{wang2026sweprunerselfadaptivecontextpruning,
2 title={SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents},
3 author={Yuhang Wang and Yuling Shi and Mo Yang and Rongrui Zhang and Shilin He and Heng Lian and Yuting Chen and Siyu Ye and Kai Cai and Xiaodong Gu},
4 year={2026},
5 eprint={2601.16746},
6 archivePrefix={arXiv},
7 primaryClass={cs.SE},
8 url={https://arxiv.org/abs/2601.16746},
9}
