SWE-Pruner Pro：利用Agent内部表示自动剪枝工具输出

The paper proposes using the coding agent’s own internal representations to remove irrelevant tool-output lines without a separate pruning model.

SWE-Pruner Pro reads the model’s last-layer hidden states and labels every tool-output line as keep or remove.

Coding agents repeatedly read files, logs, and search results, then carry large amounts of unused text into later turns.

SWE-Pruner Pro adds a small classifier that reads hidden states, the model’s internal summaries, and labels each line keep or prune.

A length-aware signal makes it gentler with short outputs, while a balanced training loss protects rare but important lines.

The team trained only this added head on 22,609 labeled tool responses, leaving the coding LLM itself unchanged.

Across 2 open models and 4 multi-turn benchmarks, it cut token use by up to 39% while mostly preserving quality.

On MiMo-V2-Flash, it also improved one coding benchmark by 3.8% and one long-context test by 2.2 points.

The result suggests coding agents already contain enough relevance information to manage context without a separate pruning model.

– arxiv. org/abs/2607.18213

Title: "SWE-Pruner Pro: The Coder LLM Already Knows What to Prune"