CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing

CRANE is a training-free parameter-editing method that merges paired Instruct and Thinking model checkpoints to create superior code agents. By selectively combining reasoning capabilities from Thinking models with the tool-discipline of Instruct models, CRANE achieves significant performance gains—66.2% pass rate on Roo-Eval (+19.5%) and resolves 14 additional instances on SWE-bench—while maintaining computational efficiency.

CRANE addresses a fundamental tension in large language model design: code agents require both precise tool-use adherence and sophisticated reasoning over complex repository states, yet models optimized for one capability often sacrifice the other. The Instruct checkpoint prioritizes conciseness and protocol compliance, while Thinking variants excel at planning and recovery but introduce computational overhead and format degradation. Rather than retraining, CRANE extracts the performance delta between these paired models and intelligently applies it to the Instruct backbone through parameter editing.

This approach reflects the broader AI trend toward parameter-efficient adaptation. As model sizes grow beyond billion-parameter scales, training-free methods that unlock emergent capabilities without full retraining become increasingly valuable. CRANE's three-stage mechanism—magnitude thresholding to filter noise, Conservative Taylor Gating to balance competing objectives, and Graduated Sigmoidal Projection to protect format-critical parameters—demonstrates sophisticated understanding of how model representations encode different competencies.

For the AI development community, CRANE's results carry tangible implications. Code generation and repository reasoning are critical bottlenecks in autonomous development workflows. The demonstrated improvements across three different benchmarks (Roo-Eval, SWE-bench-Verified, Terminal-Bench v2) suggest the method generalizes beyond specific model architectures or task distributions. The efficiency gains matter particularly for production systems where inference cost scales with model capacity.

The work establishes parameter editing as a viable strategy for capability merging. Future research likely explores whether similar delta-extraction approaches generalize to other complementary model pairs—reasoning versus efficiency, instruction-following versus creativity—and whether cascading multiple deltas compounds gains or introduces instability.

• →CRANE merges Instruct and Thinking checkpoints via training-free parameter editing without sacrificing inference efficiency
• →Performance improves 19.5% on Roo-Eval for Qwen3-30B and resolves 14 additional SWE-bench instances through selective delta application
• →The method uses magnitude thresholding and Conservative Taylor Gating to balance reasoning transfer with tool-use protocol preservation
• →Graduated Sigmoidal Projection prevents format degradation by protecting parameters critical to correct tool invocation
• →Results generalize across three independent benchmarks, suggesting broad applicability to code agent architecture design