{"paper":{"title":"CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing","license":"http://creativecommons.org/licenses/by/4.0/","headline":"CRANE injects selected reasoning directions from thinking checkpoints into instruct models by editing their parameter nullspace, raising code agent success rates on benchmarks while keeping tool-use efficiency.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.SE","authors_text":"Michele Merler, Mingzhi Zhu, Raju Pavuluri, Stacy Patterson","submitted_at":"2026-05-13T20:09:35Z","abstract_excerpt":"Code agents must both reason over long-horizon repository state and obey strict tool-use protocols. In paired Instruct/Thinking checkpoints, these capabilities are complementary but misaligned. The Instruct model is concise and tool-disciplined, whereas the Thinking model offers stronger planning and recovery behavior but often over-deliberates and degrades agent performance. We present CRANE (Constrained Reasoning Injection for Code Agents via Nullspace Editing), a training-free parameter-editing method that treats the Thinking-Instruct delta as a directional pool of candidate reasoning edits"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By merging paired Instruct and Thinking checkpoints, CRANE delivers strong gains over either individual model while preserving Instruct-level efficiency: on Roo-Eval it achieves pass1 of 66.2% (+19.5%) for Qwen3-30B-A3B and 81.5% (+8.7%) for Qwen3-Next-80B-A3B; on SWE-bench-Verified it resolves up to 14 additional instances at both scales (122/500 and 180/500); and on Terminal-Bench v2 it improves pass1/pass5 by up to 2.3%/7.8%, reaching 7.6%/17.9% and 14.8%/30.3%, respectively, consistently outperforming alternative merging strategies across all three benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the Thinking-Instruct delta vector contains directional information that can be selectively transferred via magnitude thresholding, Conservative Taylor Gate, and Graduated Sigmoidal Projection without degrading tool-use protocols or introducing new failure modes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CRANE merges Instruct and Thinking model checkpoints via constrained nullspace editing to improve code agent reasoning and benchmark performance without retraining.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"CRANE injects selected reasoning directions from thinking checkpoints into instruct models by editing their parameter nullspace, raising code agent success rates on benchmarks while keeping tool-use efficiency.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"616447d40d0cd1ec1642b09cc23906aca6d39df32be06afc62f5732e36c6ad77"},"source":{"id":"2605.14084","kind":"arxiv","version":1},"verdict":{"id":"df310626-5be2-45b9-8a4e-e48893b0fea6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T04:42:27.537662Z","strongest_claim":"By merging paired Instruct and Thinking checkpoints, CRANE delivers strong gains over either individual model while preserving Instruct-level efficiency: on Roo-Eval it achieves pass1 of 66.2% (+19.5%) for Qwen3-30B-A3B and 81.5% (+8.7%) for Qwen3-Next-80B-A3B; on SWE-bench-Verified it resolves up to 14 additional instances at both scales (122/500 and 180/500); and on Terminal-Bench v2 it improves pass1/pass5 by up to 2.3%/7.8%, reaching 7.6%/17.9% and 14.8%/30.3%, respectively, consistently outperforming alternative merging strategies across all three benchmarks.","one_line_summary":"CRANE merges Instruct and Thinking model checkpoints via constrained nullspace editing to improve code agent reasoning and benchmark performance without retraining.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the Thinking-Instruct delta vector contains directional information that can be selectively transferred via magnitude thresholding, Conservative Taylor Gate, and Graduated Sigmoidal Projection without degrading tool-use protocols or introducing new failure modes.","pith_extraction_headline":"CRANE injects selected reasoning directions from thinking checkpoints into instruct models by editing their parameter nullspace, raising code agent success rates on benchmarks while keeping tool-use efficiency."},"references":{"count":21,"sample":[{"doi":"","year":2026,"title":"Qwen3-Coder-Next Technical Report","work_id":"ad966e68-641d-4b33-a9da-57cf741f35a6","ref_index":1,"cited_arxiv_id":"2603.00729","is_internal_anchor":true},{"doi":"","year":null,"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","ref_index":2,"cited_arxiv_id":"2501.12948","is_internal_anchor":true},{"doi":"","year":null,"title":"Lorahub: Efficient cross-task generalization via dynamic lora composition","work_id":"66ee831a-2b33-423d-87ec-ba52b341a21b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"SWE-bench: Can Language Models Resolve Real-World GitHub Issues?","work_id":"d0effe15-a689-441a-8e3f-ea35f1c4e4b1","ref_index":4,"cited_arxiv_id":"2310.06770","is_internal_anchor":true},{"doi":"","year":null,"title":"arXiv preprint arXiv:2505.22113 , year=","work_id":"1eb7cd22-ed7f-48ef-be86-59775a274675","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":21,"snapshot_sha256":"8e50bd1b8a5fe7bf2a0bf23fcc1a9212cdf947d8adc8cc87d7fdd9996534443c","internal_anchors":7},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}