SWE-Pruner Pro: The Coder LLM Already Knows What to Prune
The pruning signal comes from the coding agent’s own activations, not a separate classifier.
SWE-Pruner Pro adds a small head that turns the agent’s internal representations into line-by-line keep-or-prune decisions for tool outputs. The method also uses a length-aware embedding tied to each output’s line count. Across two open-weight backbones and four multi-turn benchmarks, the paper reports up to 39% token savings while preserving task quality. On MiMo-V2-Flash, it reports a +3.8% SWE-Bench Verified resolve-rate gain and +2.2 points on long-context Oolong accuracy. HF Daily Papers' note
SWE-Pruner Pro adds a small head that turns the agent’s internal representations into line-by-line keep-or-prune decisions for tool outputs. The method also uses a length-aware embedding tied to each output’s line count. Across two open-weight backbones and four multi-turn benchmarks, the paper reports up to 39% token savings while preserving task quality. On MiMo-V2-Flash, it reports a +3.8% SWE-Bench Verified resolve-rate gain and +2.2 points on long-context Oolong accuracy. HF Daily Papers' note
score 5