CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing

CRANE (Constrained Reasoning Injection for Code Agents via Nullspace Editing) introduces a training-free parameter-editing method that optimally merges Instruct and Thinking model checkpoints to enhance code agent performance. By leveraging techniques such as magnitude thresholding and Graduated Sigmoidal Projection, CRANE achieves significant improvements in reasoning and tool-use efficiency, yielding pass rates of 66.2% on Roo-Eval for Qwen3-30B-A3B and 81.5% for Qwen3-Next-80B-A3B, while also enhancing performance across multiple benchmarks. This approach is crucial for practitioners seeking to optimize the functionality of code agents without extensive retraining, allowing for more effective long-horizon reasoning and adherence to tool-use protocols.