SWE-Pruner Pro: The Coder LLM Already Knows What to Prune
The pruning signal is pulled from the coding agent’s own hidden state, not a separate classifier.
SWE-Pruner Pro adds a small head that labels each tool-output line as keep or prune using the agent’s internal representations. The paper says a length-aware embedding helps account for the number of lines in each tool result. Across two open-weight backbones and four multi-turn benchmarks, it cuts token use by up to 39% while preserving task quality. On MiMo-V2-Flash, it also reports gains on SWE-Bench Verified and Oolong. ArXiv · AI/CL/LG's note
SWE-Pruner Pro adds a small head that labels each tool-output line as keep or prune using the agent’s internal representations. The paper says a length-aware embedding helps account for the number of lines in each tool result. Across two open-weight backbones and four multi-turn benchmarks, it cuts token use by up to 39% while preserving task quality. On MiMo-V2-Flash, it also reports gains on SWE-Bench Verified and Oolong. ArXiv · AI/CL/LG's note
score 5