SWE-Pruner vs SWE-Pruner Pro, same problem, different place to put the score: Prior SWE-Pruner: separate pruner model + agent-written goal-hint every turn. Pro: read line importance off the agent…

SWE-Pruner Pro The coder LLM already knows what to prune – reads keep-or-prune signals from its own hidden states, saving up to 39% tokens while preserving task quality.