{"paper":{"title":"Reasoning Can Be Restored by Correcting a Few Decision Tokens","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Base LLMs lose most reasoning ability at a few early planning tokens that can be fixed by brief stronger-model intervention.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"An Zhang, Changshuo Shen, Leheng Sheng, Xiang Wang, Yuxin Chen","submitted_at":"2026-05-16T08:33:31Z","abstract_excerpt":"Large reasoning models (LRMs) substantially outperform their base LLM counterparts on challenging reasoning benchmarks, yet it remains poorly understood where base models go wrong during token-by-token generation and how to narrow this gap efficiently. We study the base-reasoning gap through quantifying token-level distributional disagreement between a base model and a stronger reasoning model using likelihood-based divergences. Across benchmarks, we find that the reasoning advantage is highly sparse and concentrates on a small set of early, planning-related decision tokens. For instance, on Q"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across benchmarks, the reasoning advantage is highly sparse and concentrates on a small set of early, planning-related decision tokens. For instance, on Qwen3-0.6B, only ~8% of generated tokens account for the salient disagreement, and these tokens concentrate early in the response, are strongly enriched in planning-related decisions (17x), and coincide with high base-model uncertainty.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the positions of high likelihood-based distributional disagreement between base and reasoning models are precisely the causal points where the base model’s early planning errors derail the subsequent reasoning trajectory, rather than merely correlated symptoms.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Reasoning gaps between base LLMs and LRMs concentrate on ~8% of early planning tokens; intervening with the reasoning model only at high-disagreement positions recovers performance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Base LLMs lose most reasoning ability at a few early planning tokens that can be fixed by brief stronger-model intervention.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f781aae1e3143eebab73738df5644d515e07334a7b61102f7bf2c0d2c02506dc"},"source":{"id":"2605.16874","kind":"arxiv","version":1},"verdict":{"id":"0fc56498-5930-4705-bf68-ead1ca1ba92b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:52:50.412735Z","strongest_claim":"Across benchmarks, the reasoning advantage is highly sparse and concentrates on a small set of early, planning-related decision tokens. For instance, on Qwen3-0.6B, only ~8% of generated tokens account for the salient disagreement, and these tokens concentrate early in the response, are strongly enriched in planning-related decisions (17x), and coincide with high base-model uncertainty.","one_line_summary":"Reasoning gaps between base LLMs and LRMs concentrate on ~8% of early planning tokens; intervening with the reasoning model only at high-disagreement positions recovers performance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the positions of high likelihood-based distributional disagreement between base and reasoning models are precisely the causal points where the base model’s early planning errors derail the subsequent reasoning trajectory, rather than merely correlated symptoms.","pith_extraction_headline":"Base LLMs lose most reasoning ability at a few early planning tokens that can be fixed by brief stronger-model intervention."},"integrity":{"clean":false,"summary":{"advisory":1,"critical":0,"by_detector":{"doi_compliance":{"total":1,"advisory":1,"critical":0,"informational":0}},"informational":0},"endpoint":"/pith/2605.16874/integrity.json","findings":[{"note":"DOI in the printed bibliography is fragmented by whitespace or line breaks. A longer candidate (10.1109/TIT.2009.2027527.URL) was visible in the surrounding text but could not be confirmed against doi.org as printed.","detector":"doi_compliance","severity":"advisory","ref_index":10,"audited_at":"2026-05-19T21:01:24.900423Z","detected_doi":"10.1109/TIT.2009.2027527.URL","finding_type":"recoverable_identifier","verdict_class":"incontrovertible","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T21:01:24.900423Z","status":"completed","version":"1.0.0","findings_count":1},{"name":"doi_title_agreement","ran_at":"2026-05-19T21:01:19.213886Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.295789Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.372472Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"001554fc071bcef2dbad9ab2d4cbe1380ebca1053576159ee2b79183bc7bdd8a"},"references":{"count":33,"sample":[{"doi":"","year":2005,"title":"Language Models are Few-Shot Learners","work_id":"214732c0-2edd-44a0-af9e-28184a2b8279","ref_index":1,"cited_arxiv_id":"2005.14165","is_internal_anchor":true},{"doi":"","year":null,"title":"arXiv preprint arXiv:2510.00553 , year=","work_id":"a6fda285-7688-48ed-9537-cbaa641673ef","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Cascade speculative drafting for even faster llm inference,","work_id":"a42682ab-46cc-4a73-823d-d74a2fe116e9","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","ref_index":4,"cited_arxiv_id":"2110.14168","is_internal_anchor":true},{"doi":"","year":null,"title":"Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs","work_id":"f65a84bf-c5b4-4491-a618-18bb263c60e5","ref_index":5,"cited_arxiv_id":"2503.01307","is_internal_anchor":true}],"resolved_work":33,"snapshot_sha256":"19f1cf9b040ccd4069233bab69759a120d96a88c0291a629b1ea91f008d01313","internal_anchors":15},"formal_canon":{"evidence_count":2,"snapshot_sha256":"e5de90e08a366df0720a68a411b87a1a69cdf948a4230f53c5f6224a4599c976"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}