{"type":"post","id":"rethinking-cross-tokenizer-on-policy-distillation-from-alignment-5d9f1e","slug":"rethinking-cross-tokenizer-on-policy-distillation-from-alignment-5d9f1e","title":"Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability","description":"Cross-tokenizer distillation may benefit more from reliable supervision at aligned positions than from maximizing coverage.","retrieval_nugget":"Cross-tokenizer distillation may benefit more from reliable supervision at aligned positions than from maximizing coverage.","published_at":"2026-10-10","updated_at":"2026-10-10","record_date":"2026-10-10","date_kind":"published_at","topics":[],"source_urls":["https://arxiv.org/abs/2610.08448"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:rethinking-cross-tokenizer-on-policy-distillation-from-alignment-5d9f1e","status":"published","visuals":[],"editorial_provenance":{"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:ccc645754a73079a852fc04365f67f6fbfaf22cca2fe0392b3c36885040ce670","reviewed_at":"2026-10-10T09:04:10.120Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":1,"related_record_count":0},"analysis":{"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Cross-tokenizer distillation may benefit more from reliable supervision at aligned positions than from maximizing coverage.","observed_facts":[{"text":"Across three heterogeneous teacher–student pairs on mathematical reasoning and code generation, strict one-to-one groups covered most student-generated tokens.","source_urls":["https://arxiv.org/abs/2610.08448"]},{"text":"The paper reports that shared vocabulary retained nearly all teacher and student probability mass at strictly aligned positions on average.","source_urls":["https://arxiv.org/abs/2610.08448"]},{"text":"Adding span supervision provided complete coverage but reduced accuracy in the reported experiments.","source_urls":["https://arxiv.org/abs/2610.08448"]}],"mechanism":"Adding span log-probability supervision in mismatch groups produced gradients with weak or negative directional agreement with strict-loss gradients, potentially introducing conflicting training signals.","why_now":"The paper makes an empirical case against treating coverage as a sufficient proxy for useful training information in cross-tokenizer OPD.","implications":["In the evaluated settings, compact supervision at strict positions can perform as well as or better than broader coverage.","Gradient agreement and downstream accuracy are relevant checks when adding supervision for mismatched spans."],"evidence_boundary":"These are results reported in one arXiv paper across three heterogeneous pairs and math and code-generation tasks; they do not establish that strict-only supervision is preferable across all architectures, datasets, or training regimes.","watch_conditions":["Replication across additional model pairs and tasks would show whether the reported gradient conflict and accuracy trade-off generalize."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied for this specific cross-tokenizer distillation result."},"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/rethinking-cross-tokenizer-on-policy-distillation-from-alignment-5d9f1e/","markdown":"https://newruntime.com/posts/rethinking-cross-tokenizer-on-policy-distillation-from-alignment-5d9f1e.md","json":"https://newruntime.com/posts/rethinking-cross-tokenizer-on-policy-distillation-from-alignment-5d9f1e.json"},"next_reads":[{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/source-ledger.json","url":"https://newruntime.com/source-ledger.json","title":"Source Ledger","reason":"Inspect the public source evidence graph.","media_type":"application/json"}]}