{"paper":{"title":"GRACE: Gradient-aligned Reasoning Data Curation for Efficient Post-training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"GRACE scores each reasoning step by its alignment with the answer gradient and trajectory consistency to select data subsets that match or exceed full performance with 5-20 percent of the samples.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Jianghong Ma, Junjie Li, Ningxuan Ma, Xiaofeng Zhang, Ziao Wang","submitted_at":"2026-05-13T07:55:39Z","abstract_excerpt":"Existing reasoning data curation pipelines score whole samples, treating every intermediate step as equally valuable. In reality, steps within a trace contribute very unevenly, and selecting reasoning data well requires assessing them individually. We present GRACE, a gradient-aligned curation method that views each reasoning trace as a sequence of optimization events and scores every step by two complementary signals: its alignment with the answer-oriented gradient direction, and its consistency with the preceding reasoning trajectory. Step-level scores are aggregated into a sample-level valu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Post-training Qwen3-VL-2B-Instruct on MMathCoT-1M, GRACE reaches 108.8% of the full-data performance with 20% of the data and retains 100.2% with only 5%, with subsets that transfer effectively across model backbones.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the representation-level gradient proxy accurately captures step-level alignment with the answer-oriented gradient and that the two signals (alignment and consistency) reliably identify valuable reasoning steps without external reward models or step annotations.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GRACE scores reasoning steps via gradient alignment and trajectory consistency to select data subsets that match full performance with 5% of the data on Qwen3-VL-2B-Instruct.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"GRACE scores each reasoning step by its alignment with the answer gradient and trajectory consistency to select data subsets that match or exceed full performance with 5-20 percent of the samples.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ffc8bbbaaaf523aafc28b6ebae70aa304b391d0a969e5b3b832073fae678949a"},"source":{"id":"2605.13130","kind":"arxiv","version":1},"verdict":{"id":"504e6685-81d7-4c5f-83b5-34d3cfa40f77","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:30:21.939493Z","strongest_claim":"Post-training Qwen3-VL-2B-Instruct on MMathCoT-1M, GRACE reaches 108.8% of the full-data performance with 20% of the data and retains 100.2% with only 5%, with subsets that transfer effectively across model backbones.","one_line_summary":"GRACE scores reasoning steps via gradient alignment and trajectory consistency to select data subsets that match full performance with 5% of the data on Qwen3-VL-2B-Instruct.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the representation-level gradient proxy accurately captures step-level alignment with the answer-oriented gradient and that the two signals (alignment and consistency) reliably identify valuable reasoning steps without external reward models or step annotations.","pith_extraction_headline":"GRACE scores each reasoning step by its alignment with the answer gradient and trajectory consistency to select data subsets that match or exceed full performance with 5-20 percent of the samples."},"references":{"count":37,"sample":[{"doi":"","year":2022,"title":"Chain-of-thought prompting elicits reasoning in large language models","work_id":"3c1378f6-2da6-4177-a291-457b1d9a8feb","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Unlocking mul- timodal mathematical reasoning via process reward model","work_id":"60e34229-23f0-4f66-baf4-54f53416ec28","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"LIMA: less is more for alignment","work_id":"b86e15f2-318c-4e3d-bf76-7d2e60e38819","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V . Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. InThe Elev","work_id":"4c6bab41-df70-4184-9335-14ec662bf3bd","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Mille","work_id":"69ae1ccd-58f3-4ccc-875e-b88f2e78e809","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":37,"snapshot_sha256":"515f22013a59ae50b0be7a1c3c4e4f9b7a7d5587198bca1c0c8aad8770aa8900","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"bb03cad62e01311a18ce1ffdf7520798f0b60b56b075330ecb761b567e5a37d9"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}