{"paper":{"title":"UniSD: Towards a Unified Self-Distillation Framework for Large Language Models","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"A unified framework makes self-distillation a reliable way to adapt large language models without stronger teachers.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"B. Aditya Prakash, Haoxin Liu, Jindong Wang, Josiah Hester, Lucheng Fu, Srijan Kumar, Yijia Xiao, Yinyi Luo, Yiqiao Jin, Yiyang Wang","submitted_at":"2026-05-07T17:22:11Z","abstract_excerpt":"Self-distillation (SD) offers a promising path for adapting large language models (LLMs) without relying on stronger external teachers. However, SD in autoregressive LLMs remains challenging because self-generated trajectories are free-form, correctness is task-dependent, and plausible rationales can still provide unstable or unreliable supervision. Existing methods mainly examine isolated design choices, leaving their effectiveness, roles, and interactions unclear. In this paper, we propose UniSD, a unified framework to systematically study self-distillation. UniSD integrates complementary me"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"UniSDfull, an integrated pipeline that combines complementary components, achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the listed mechanisms (multi-teacher agreement, EMA, token contrastive learning, feature matching, divergence clipping) reliably address supervision instability in free-form self-generated trajectories and that their interactions can be isolated and combined without post-hoc selection bias affecting the reported gains.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"UniSD unifies complementary self-distillation mechanisms for autoregressive LLMs and achieves up to +5.4 point gains over base models and +2.8 over baselines across six benchmarks and six models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A unified framework makes self-distillation a reliable way to adapt large language models without stronger teachers.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"43f0993ffab088ed0c3a24d3818d0127f6c5b5d58b2ff8847fd9c854ec42b87a"},"source":{"id":"2605.06597","kind":"arxiv","version":2},"verdict":{"id":"c6a1c7da-9796-4d61-893c-c22d84397521","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T09:52:45.380637Z","strongest_claim":"UniSDfull, an integrated pipeline that combines complementary components, achieves the strongest overall performance, improving over the base model by +5.4 points and the strongest baseline by +2.8 points.","one_line_summary":"UniSD unifies complementary self-distillation mechanisms for autoregressive LLMs and achieves up to +5.4 point gains over base models and +2.8 over baselines across six benchmarks and six models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the listed mechanisms (multi-teacher agreement, EMA, token contrastive learning, feature matching, divergence clipping) reliably address supervision instability in free-form self-generated trajectories and that their interactions can be isolated and combined without post-hoc selection bias affecting the reported gains.","pith_extraction_headline":"A unified framework makes self-distillation a reliable way to adapt large language models without stronger teachers."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.06597/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T12:22:03.785641Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-20T07:37:52.552364Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T18:01:19.364843Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T12:33:31.546986Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"162ddda1bbe6f2db18cebf495266973fcf3756ba02a57ed1f4fd0e4d62967ed0"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"}