CRANE: Training-Free Method Merges Reasoning and Tool-Use in Code Agents
Researchers have unveiled a novel technique named CRANE, which stands for Constrained Reasoning Injection for Code Agents via Nullspace Editing. This method enables the modification of parameters without requiring training, combining paired Instruct and Thinking checkpoints specifically for code agents. The Instruct model is known for its efficiency and tool proficiency, while the Thinking model is great at strategizing but often overanalyzes. CRANE identifies the differences between these two models as opportunities for reasoning improvements, applying methods like magnitude thresholding, a Conservative Taylor Gate, and Graduated Sigmoidal Projection to maintain beneficial edits. Essentially, this approach aims to align the strengths of both models that are usually out of sync. You can check out the paper on arXiv under the identifier 2605.14084.
Key facts
- CRANE stands for Constrained Reasoning Injection for Code Agents via Nullspace Editing.
- It is a training-free parameter-editing method.
- It merges paired Instruct and Thinking checkpoints.
- Instruct model is concise and tool-disciplined.
- Thinking model offers stronger planning but over-deliberates.
- CRANE uses magnitude thresholding to denoise the delta.
- Conservative Taylor Gate retains jointly beneficial edits.
- Graduated Sigmoidal Projection suppresses format-critical updates.
- Paper available on arXiv: 2605.14084.
Entities
Institutions
- arXiv