SWE-Pruner Pro: Context Pruning Using Internal Agent Representations

For efficient context management in coding agents, instead of the conventional approach using a separate classifier, we propose SWE-Pruner Pro, which leverages the agent's own internal representations.

The core mechanism is as follows:

• Based on the internal representations generated when the agent reads tool outputs, a small head is used to decide whether to keep or drop each line.
• A length-aware embedding is applied, matched to the number of lines in each tool output.

The key results are as follows:

• Tested on two open-weight backbones and 4 multi-turn benchmarks, it achieved up to a 39% reduction in prompt and completion tokens while maintaining task quality.
• It maintained performance while keeping inference overhead limited.
• Notably, on the MiMo-V2-Flash model, it improved the SWE-Bench Verified resolve rate by 3.8% and Oolong accuracy by 2.2 points.