{"slug":"on-policy-distillation","label":"On-Policy Distillation","item_count":3,"day_count":3,"source_count":3,"first_seen":"2026-10-14T08:12:10+00:00","last_updated":"2026-10-22T17:05:21+00:00","generated_at":"2026-10-26T00:13:14.453929+00:00","sources":["arxiv_cn_open_weight_reports","arxiv_cs_cl","arxiv_cs_lg"],"days":[{"date":"2026-10-14","items":[{"title":"Qwen-MT-Lumen at WMT2026: Instruction-Specialized Translation via Quality-Aware Training and On-Policy Distillation","url":"http://arxiv.org/abs/2610.19758v1","source":"arxiv_cn_open_weight_reports","type":"paper","summary_1line":"We present Lumen, The open-source model is available at, our submission to the WMT 2026 General Machine Translation Shared Task. We focus on instruction-specialized translation, where systems must produce accurate tra...","why_it_matters":"Matches feed focus: eval.","sid":"ee660e8935ab6a11","published":"2026-10-14T08:12:10+00:00","editor_note":"Shows on-policy training in a competitive translation submission."}]},{"date":"2026-10-15","items":[{"title":"OPRD: On-Policy Reminder Distillation for Long-Horizon Coding Agents","url":"http://arxiv.org/abs/2610.21496v1","source":"arxiv_cs_cl","type":"paper","summary_1line":"Reliable repository-level coding requires agents not only to generate correct code but also to preserve constraints and respond to tool feedback across long interactions. However, our diagnostic analysis shows that ag...","why_it_matters":"Matches feed focus: agent, eval.","sid":"bb1f85907195d2e5","published":"2026-10-15T08:32:30+00:00","editor_note":"Extends the idea to long coding-agent sessions with tool feedback."}]},{"date":"2026-10-22","items":[{"title":"When to Follow the Teacher: Advantage-Based On-Policy Distillation","url":"http://arxiv.org/abs/2610.31563v1","source":"arxiv_cs_lg","type":"paper","summary_1line":"Online Policy Distillation (OPD) often outperforms offline methods such as Supervised Fine-Tuning (SFT) by training on trajectories sampled directly from the student. However, standard OPD trains the student to follow...","why_it_matters":"Matches feed focus: evaluation.","sid":"345d8b761233aa20","published":"2026-10-22T17:05:21+00:00","editor_note":"Adds a rule for when the student should defer to the teacher."}]}],"editorial":{"tldr":"Three arXiv papers from Oct 14–22 apply on-policy distillation (training a student on its own sampled trajectories, scored by a teacher) to different problems: machine translation, long-horizon coding agents, and the question of when the teacher should lead.","stale":false,"whats_new":"Oct 22: an arXiv paper proposes advantage-based on-policy distillation, deciding per step when the student should follow the teacher.","why_it_matters":"On-policy distillation is spreading from general post-training into narrow jobs (translation, coding agents), so teams distilling small models should compare it against SFT on their own trajectories.","take_for_builders":"If you distill small models, run an on-policy versus SFT comparison on your own task trajectories before adopting either method; none of these papers has independent replication yet.","beats":[{"kicker":"TRANSLATION","tone":"launch","headline":"Qwen-MT-Lumen applies on-policy training to WMT 2026 translation","sids":["ee660e8935ab6a11"]},{"kicker":"CODING AGENTS","tone":"rising","headline":"OPRD distills constraint reminders into long-horizon coding agents","sids":["bb1f85907195d2e5"]},{"kicker":"NOW","tone":"now","headline":"Advantage-based method decides when to follow the teacher","sids":["345d8b761233aa20"]}],"open_questions":["Does the advantage-based method beat plain on-policy distillation on agentic coding tasks, not only its own benchmarks?","Does any group release code that reproduces the OPRD long-horizon results?"],"generated_at":"2026-10-26T00:12:56.393036+00:00"}}