{"paper":{"title":"OGLS-SD: On-Policy Self-Distillation with Outcome-Guided Logit Steering for LLM Reasoning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Outcome-guided logit steering calibrates teacher responses in on-policy self-distillation by contrasting successful and failed trajectories, reducing reflection bias for better LLM reasoning.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Weitong Zhang, Xiaoyun Wang, Yuxiao Yang","submitted_at":"2026-05-12T17:00:53Z","abstract_excerpt":"We study on-policy self-distillation (OPSD), where a language model improves its reasoning ability by distilling privileged teacher distributions along its own on-policy trajectories. Despite its promise, OPSD can suffer from training instability due to a pattern mismatch between teacher and student responses. Self-reflected teacher responses may introduce reflection-induced biases and response templates that miscalibrate token-level supervision, ultimately harming the student's reasoning ability. To mitigate this issue, we propose OGLS-SD, an outcome-guided logit-steering framework that lever"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By combining outcome-level correctness with dense token-level guidance through logit steering, OGLS-SD stabilizes self-distillation and improves reasoning performance over standard OPSD and other variants across diverse benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That verifiable outcome rewards can reliably contrast successful and failed on-policy trajectories to calibrate teacher logits and mitigate reflection-induced bias without introducing new calibration issues or depending on tasks where outcomes are easily verifiable.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OGLS-SD improves LLM reasoning by using verifiable outcome rewards to guide logit steering that calibrates teacher distributions in on-policy self-distillation, addressing reflection-induced mismatches.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Outcome-guided logit steering calibrates teacher responses in on-policy self-distillation by contrasting successful and failed trajectories, reducing reflection bias for better LLM reasoning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a34ed2143aab1182847d78b1fc606f13e67ee4443a9575f7535b78c8165d112d"},"source":{"id":"2605.12400","kind":"arxiv","version":2},"verdict":{"id":"3c816f6c-41af-4df3-8077-fa5a10a1946d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T05:14:01.558810Z","strongest_claim":"By combining outcome-level correctness with dense token-level guidance through logit steering, OGLS-SD stabilizes self-distillation and improves reasoning performance over standard OPSD and other variants across diverse benchmarks.","one_line_summary":"OGLS-SD improves LLM reasoning by using verifiable outcome rewards to guide logit steering that calibrates teacher distributions in on-policy self-distillation, addressing reflection-induced mismatches.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That verifiable outcome rewards can reliably contrast successful and failed on-policy trajectories to calibrate teacher logits and mitigate reflection-induced bias without introducing new calibration issues or depending on tasks where outcomes are easily verifiable.","pith_extraction_headline":"Outcome-guided logit steering calibrates teacher responses in on-policy self-distillation by contrasting successful and failed trajectories, reducing reflection bias for better LLM reasoning."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.12400/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-26T14:40:00.497927Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-20T13:31:24.992235Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T09:37:46.618193Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T22:21:57.872449Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"4ccc60c6e5dbd90b21ad39be1748bbd4cf3e8a42ec904c34d6b7d211b50113cc"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"bd398c76ef47e7aeaca8a0939b173d541fae73171e53f59e79da53c3bf777d43"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}