{"paper":{"title":"Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Fine-tuning data with ambiguous decision points causes reasoning models to lose coverage.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Chandan K Reddy, Khoa D Doan, Nan Zhang, Ngoc-Hieu Nguyen, Parshin Shojaee, Phuc Minh Nguyen, Rui Zhang","submitted_at":"2026-05-16T14:55:12Z","abstract_excerpt":"Recent progress in large language models has led to the emergence of reasoning models, which have shown strong performance on complex tasks through specialized fine-tuning procedures. While these methods reliably improve pass@1 accuracy, prior works have observed that they show a coverage shrinkage behavior, where pass@k degrades relative to the base model. In this paper, we investigate the reasoning shrinkage arise under SFT-based post-training. We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or \"forks in the road\" sce"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or 'forks in the road' scenarios where model faces indecipherable patterns with multiple valid reasoning paths.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The controlled case studies using graph branching and reasoning modes accurately capture the decision-point dynamics present in real fine-tuning datasets for reasoning models.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Coverage shrinkage after SFT in reasoning models correlates with prevalence of decision-point scenarios in data and can be partially mitigated by targeted data synthesis and diversity-aware decoding.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Fine-tuning data with ambiguous decision points causes reasoning models to lose coverage.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"41483c2b9196a4088ad8efd4f8907e158f1fd4815a3d7ef31f64c7073f1297c7"},"source":{"id":"2605.17026","kind":"arxiv","version":1},"verdict":{"id":"4e675cf6-3703-446b-a1ff-9b7bc3112c4e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:28:40.256953Z","strongest_claim":"We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or 'forks in the road' scenarios where model faces indecipherable patterns with multiple valid reasoning paths.","one_line_summary":"Coverage shrinkage after SFT in reasoning models correlates with prevalence of decision-point scenarios in data and can be partially mitigated by targeted data synthesis and diversity-aware decoding.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The controlled case studies using graph branching and reasoning modes accurately capture the decision-point dynamics present in real fine-tuning datasets for reasoning models.","pith_extraction_headline":"Fine-tuning data with ambiguous decision points causes reasoning models to lose coverage."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.17026/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:01:19.007508Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T20:40:42.638744Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"citation_quote_validity","ran_at":"2026-05-19T19:49:44.685131Z","status":"skipped","version":"0.1.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T18:51:57.643361Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.177593Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:23.494370Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"27b76f4a138faead9813f5402b880aaad5552ffd2df01777d3b1b69f902826e1"},"references":{"count":19,"sample":[{"doi":"10.1038/s41586-025-09422-z","year":2021,"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","ref_index":1,"cited_arxiv_id":"2110.14168","is_internal_anchor":true},{"doi":"","year":null,"title":"Substitutep=l+7 into the target expression, yieldings=l+18","work_id":"2e758a34-c102-4f90-b378-c27e479f1674","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Substitutel=m+5 into the target expression, yieldings=m+23","work_id":"ab49658f-2a2f-4199-a457-12ca2d68223d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Substitutem=f+9 into the target expression, yieldings=f+32","work_id":"536dd746-212d-446f-99d1-bd89c04145cc","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Substitutef=g+11 into the target expression, yieldings=g+43","work_id":"0164f060-f692-44ca-a74f-c3eadd23e986","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":19,"snapshot_sha256":"19342224d08837088c7bbe151169a74f4f89bce03a87e0ace330765a72fd7108","internal_anchors":1},"formal_canon":{"evidence_count":2,"snapshot_sha256":"3f685701b55407544a51549971253411cb6fca79692b4747d6b14b187d8ed0fa"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}