Pruning Tool Outputs for Coding Agents

Summary & Methodology Analysis

SWE-Pruner Pro operates by utilizing the coding agent's internal state to identify and discard redundant tool output. The agent backbone prefills history and tool calls into its KV cache, which is the memory structure used in transformers to store previous tokens. During a single prefill pass, the system captures the last-layer hidden states, which are the numerical representations of data within the model layers, to determine token importance. A lightweight head then maps these hidden states to keep or prune logits using a length-aware embedding and a two-block feed-forward classifier, which is a simple neural network layer that processes inputs without recurrent connections. These per-token decisions are aggregated into per-line actions via majority vote, with pruned content replaced by a placeholder before the next turn. This approach allows the model to manage its own context window dynamically. The technique is currently restricted to open-weight models because it requires direct access to the internal hidden states of the transformer architecture. While the methodology is language-agnostic, existing benchmarks like SWE-Bench Verified are primarily Python-centric. Deployments in production environments should be validated for specific safety-critical or novel language scenarios. The paper does not specify the exact memory footprint reduction or the specific impact on architectures outside of those evaluated.