{"id":6152,"date":"2026-10-19T16:49:30","date_gmt":"2026-10-19T16:49:30","guid":{"rendered":"https:\/\/i10xnews.kinsta.cloud\/?p=6152"},"modified":"2026-10-19T17:22:12","modified_gmt":"2026-10-19T17:22:12","slug":"on-policy-vs-off-policy-distillation-dynamics","status":"publish","type":"post","link":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics","title":{"rendered":"On-Policy Distillation Isn&#8217;t the Magic \u2014 Learning Rate and KL Direction Are"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u26a1 Quick Take<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>The finding:<\/strong> In strong-to-weak distillation, on-policy rollouts offer no consistent accuracy edge: best means are 72% (OnPD) vs 73% (OffPD).<\/li><li><strong>KL direction dominates:<\/strong> Forward KL holds 71\u201373% across settings; reverse KL ranges from 35% to 72% and is learning-rate sensitive.<\/li><li><strong>LR governs forgetting:<\/strong> OOD drops \u22641.3 pp at low LR but 11.2\u201314.0 pp at high LR; update sparsity falls from ~85\u201390% to ~52\u201360%.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\">On-policy learning is often credited with less forgetting, sparser updates and better generalization. Cambridge researchers note that prior SFT-vs-RL comparisons entangle many factors. They isolate rollout policy in controlled distillation on Llama 3 and Qwen2.5 families.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1344\" height=\"768\" src=\"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-kl.webp\" alt=\"Grouped bars showing forward KL accuracy between 71 and 73 percent versus reverse KL spanning 35 to 72 percent.\" class=\"wp-image-6151\" srcset=\"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-kl.webp 1344w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-kl-300x171.webp 300w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-kl-1024x585.webp 1024w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-kl-768x439.webp 768w\" sizes=\"auto, (max-width: 1344px) 100vw, 1344px\" \/><figcaption class=\"wp-element-caption\">i10X chart based on data from Piskorz et al. (2026), arXiv:2609.35259, main accuracy summary (CC BY 4.0).<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udd2c What the study shows<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Independently vary rollout policy, token-level KL direction and learning rate (1e-5 or 5e-5) across scientific, medical and arithmetic reasoning tasks. Also sweep a continuous student\u2013teacher rollout mixture via \u03bb.<\/p>\n\n\n<ul class=\"wp-block-list\"><li><strong>Forward KL is robust to rollout policy:<\/strong> Performance stays strong despite policy changes.<\/li><li><strong>Reverse KL prefers student-like rollouts:<\/strong> More sensitive overall.<\/li><li><strong>Sparsity tracks LR almost perfectly:<\/strong> R\u00b2\u22481.00 between LR and update sparsity on Countdown-3.<\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\u2705 Why it matters: stop over-crediting on-policy<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Teams pay for on-policy sampling assuming it is the key ingredient. This study says check KL direction and learning rate first.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83e\uddea Method in brief<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Controlled factorial design on Llama 3 and Qwen2.5; KL gradient analysis; continuous rollout spectrum.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u2753 What remains open<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>Distillation setting:<\/strong> Strong-to-weak; may differ from RLVR from scratch.<\/li><li><strong>Task set:<\/strong> Reasoning domains tested.<\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\ud83e\udded Context<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Related distillation dynamics in <a href=\"\/news\/category\/studies\/efficiency-training\">Efficiency &amp; Training<\/a> and <a href=\"\/news\/category\/studies\/reasoning\">Reasoning<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udee0\ufe0f What it means for distillation pipelines<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>Default to forward KL:<\/strong> More robust.<\/li><li><strong>Keep LR low if OOD matters:<\/strong> Forgetting is LR-driven.<\/li><li><strong>Do not assume on-policy is required:<\/strong> Measure it.<\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\u270d\ufe0f About the analysis<\/h2>\n\n\n<p class=\"wp-block-paragraph\">This article is based on the full text of the preprint &#8220;On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics&#8221; (arXiv:2609.35259v1, submitted September 28, 2026). All numbers are taken from the paper&#8217;s text, tables and figures; i10X did not rerun the experiments. The chart was redrawn by i10X from main accuracy summary of the paper, which is published under a CC BY 4.0 license.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udd2d i10x Perspective<\/h2>\n\n\n<p class=\"wp-block-paragraph\">A careful negative result is still a result. On-policy distillation&#8217;s folklore advantages largely fail to appear once other knobs are controlled. Sometimes the boring hyperparameters were the story all along.<\/p>\n\n","protected":false},"excerpt":{"rendered":"<p>Cambridge study: OnPD vs OffPD show nearly identical best accuracies (72% vs 73%). Forward KL stays at 71\u201373%; reverse KL swings 35\u201372%. Higher LR drives forgetting and denser updates.<\/p>\n","protected":false},"author":2,"featured_media":6302,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"paper_title":"On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics","paper_url":"https:\/\/arxiv.org\/abs\/2609.35259","pdf_url":"https:\/\/arxiv.org\/pdf\/2609.35259v1","doi":"","arxiv_id":"2609.35259","authors":"Julianna Piskorz, Antonin Berthon, Mihaela van der Schaar","institutions":"University of Cambridge","paper_date":"2026-09-28","venue":"arXiv (cs.LG), version 1","peer_reviewed":false,"code_url":"","data_url":"","funding_coi":"Academic study.","figure_license":"Chart drawn by i10X from main accuracy summary (CC BY 4.0)","paper_license":"CC BY 4.0","replication":"no","study_models":"Llama 3 and Qwen2.5 families (strong-to-weak distillation)","study_sample":"Scientific, medical and arithmetic reasoning tasks","footnotes":"","_yoast_wpseo_primary_category":""},"categories":[57,52],"tags":[],"companies":[42],"products":[],"class_list":["post-6152","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-efficiency-training","category-studies","company-meta"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.7 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>On-Policy or Off-Policy? What Actually Drives Distillation - i10X News<\/title>\n<meta name=\"description\" content=\"Controlled distillation study: rollout policy matters less than KL direction and learning rate. Forward KL is robust; reverse KL is fragile.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"On-Policy or Off-Policy? What Actually Drives Distillation - i10X News\" \/>\n<meta property=\"og:description\" content=\"Controlled distillation study: rollout policy matters less than KL direction and learning rate. Forward KL is robust; reverse KL is fragile.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics\" \/>\n<meta property=\"og:site_name\" content=\"i10X News\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-19T16:49:30+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-10-19T17:22:12+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-title.webp\" \/>\n\t<meta property=\"og:image:width\" content=\"1344\" \/>\n\t<meta property=\"og:image:height\" content=\"768\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/webp\" \/>\n<meta name=\"author\" content=\"Christopher Ort\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@i10x_ai\" \/>\n<meta name=\"twitter:site\" content=\"@i10x_ai\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Christopher Ort\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"2 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"NewsArticle\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics\"},\"author\":{\"name\":\"Christopher Ort\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/person\\\/d410bb7ee72febefca2c9c369939219a\"},\"headline\":\"On-Policy Distillation Isn&#8217;t the Magic \u2014 Learning Rate and KL Direction Are\",\"datePublished\":\"2026-10-19T16:49:30+00:00\",\"dateModified\":\"2026-10-19T17:22:12+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics\"},\"wordCount\":388,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/on-policy-vs-off-policy-distillation-dynamics-title.webp\",\"articleSection\":[\"Efficiency &amp; Training\",\"Studies\"],\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics\",\"name\":\"On-Policy or Off-Policy? What Actually Drives Distillation - i10X News\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/on-policy-vs-off-policy-distillation-dynamics-title.webp\",\"datePublished\":\"2026-10-19T16:49:30+00:00\",\"dateModified\":\"2026-10-19T17:22:12+00:00\",\"description\":\"Controlled distillation study: rollout policy matters less than KL direction and learning rate. Forward KL is robust; reverse KL is fragile.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#primaryimage\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/on-policy-vs-off-policy-distillation-dynamics-title.webp\",\"contentUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/on-policy-vs-off-policy-distillation-dynamics-title.webp\",\"width\":1344,\"height\":768,\"caption\":\"Illustration: i10X\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/on-policy-vs-off-policy-distillation-dynamics#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/news.i10x.ai\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"On-Policy Distillation Isn&#8217;t the Magic \u2014 Learning Rate and KL Direction Are\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#website\",\"url\":\"https:\\\/\\\/news.i10x.ai\\\/\",\"name\":\"i10X News\",\"description\":\"Stay updated with the latest news and announcements from i10X.\",\"publisher\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/news.i10x.ai\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#organization\",\"name\":\"i10X\",\"url\":\"https:\\\/\\\/news.i10x.ai\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/i10x-logo-180.png\",\"contentUrl\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/i10x-logo-180.png\",\"width\":180,\"height\":180,\"caption\":\"i10X\"},\"image\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/x.com\\\/i10x_ai\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/i10x-ai\",\"https:\\\/\\\/instagram.com\\\/i10x.ai\",\"https:\\\/\\\/www.youtube.com\\\/@i10X_ai_official\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/person\\\/d410bb7ee72febefca2c9c369939219a\",\"name\":\"Christopher Ort\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"contentUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"caption\":\"Christopher Ort\"},\"description\":\"Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU). He is the author of \u201cThe Style Penalty\u201d, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/author\\\/christopher-ort\",\"sameAs\":[\"https:\\\/\\\/i10x.ai\\\/blog\\\/author\\\/christopher-ort\"],\"worksFor\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#organization\"},\"alumniOf\":{\"@type\":\"EducationalOrganization\",\"name\":\"Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU)\",\"url\":\"https:\\\/\\\/www.fau.eu\"}}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"On-Policy or Off-Policy? What Actually Drives Distillation - i10X News","description":"Controlled distillation study: rollout policy matters less than KL direction and learning rate. Forward KL is robust; reverse KL is fragile.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics","og_locale":"en_US","og_type":"article","og_title":"On-Policy or Off-Policy? What Actually Drives Distillation - i10X News","og_description":"Controlled distillation study: rollout policy matters less than KL direction and learning rate. Forward KL is robust; reverse KL is fragile.","og_url":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics","og_site_name":"i10X News","article_published_time":"2026-10-19T16:49:30+00:00","article_modified_time":"2026-10-19T17:22:12+00:00","og_image":[{"width":1344,"height":768,"url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-title.webp","type":"image\/webp"}],"author":"Christopher Ort","twitter_card":"summary_large_image","twitter_creator":"@i10x_ai","twitter_site":"@i10x_ai","twitter_misc":{"Written by":"Christopher Ort","Est. reading time":"2 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"NewsArticle","@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#article","isPartOf":{"@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics"},"author":{"name":"Christopher Ort","@id":"https:\/\/news.i10x.ai\/#\/schema\/person\/d410bb7ee72febefca2c9c369939219a"},"headline":"On-Policy Distillation Isn&#8217;t the Magic \u2014 Learning Rate and KL Direction Are","datePublished":"2026-10-19T16:49:30+00:00","dateModified":"2026-10-19T17:22:12+00:00","mainEntityOfPage":{"@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics"},"wordCount":388,"commentCount":0,"publisher":{"@id":"https:\/\/news.i10x.ai\/#organization"},"image":{"@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#primaryimage"},"thumbnailUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-title.webp","articleSection":["Efficiency &amp; Training","Studies"],"inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#respond"]}]},{"@type":"WebPage","@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics","url":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics","name":"On-Policy or Off-Policy? What Actually Drives Distillation - i10X News","isPartOf":{"@id":"https:\/\/news.i10x.ai\/#website"},"primaryImageOfPage":{"@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#primaryimage"},"image":{"@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#primaryimage"},"thumbnailUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-title.webp","datePublished":"2026-10-19T16:49:30+00:00","dateModified":"2026-10-19T17:22:12+00:00","description":"Controlled distillation study: rollout policy matters less than KL direction and learning rate. Forward KL is robust; reverse KL is fragile.","breadcrumb":{"@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#primaryimage","url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-title.webp","contentUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/on-policy-vs-off-policy-distillation-dynamics-title.webp","width":1344,"height":768,"caption":"Illustration: i10X"},{"@type":"BreadcrumbList","@id":"https:\/\/i10x.ai\/news\/on-policy-vs-off-policy-distillation-dynamics#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/news.i10x.ai\/"},{"@type":"ListItem","position":2,"name":"On-Policy Distillation Isn&#8217;t the Magic \u2014 Learning Rate and KL Direction Are"}]},{"@type":"WebSite","@id":"https:\/\/news.i10x.ai\/#website","url":"https:\/\/news.i10x.ai\/","name":"i10X News","description":"Stay updated with the latest news and announcements from i10X.","publisher":{"@id":"https:\/\/news.i10x.ai\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/news.i10x.ai\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/news.i10x.ai\/#organization","name":"i10X","url":"https:\/\/news.i10x.ai\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/news.i10x.ai\/#\/schema\/logo\/image\/","url":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/i10x-logo-180.png","contentUrl":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/i10x-logo-180.png","width":180,"height":180,"caption":"i10X"},"image":{"@id":"https:\/\/news.i10x.ai\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/i10x_ai","https:\/\/www.linkedin.com\/company\/i10x-ai","https:\/\/instagram.com\/i10x.ai","https:\/\/www.youtube.com\/@i10X_ai_official"]},{"@type":"Person","@id":"https:\/\/news.i10x.ai\/#\/schema\/person\/d410bb7ee72febefca2c9c369939219a","name":"Christopher Ort","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","contentUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","caption":"Christopher Ort"},"description":"Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU). He is the author of \u201cThe Style Penalty\u201d, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.","url":"https:\/\/i10x.ai\/news\/author\/christopher-ort","sameAs":["https:\/\/i10x.ai\/blog\/author\/christopher-ort"],"worksFor":{"@id":"https:\/\/i10x.ai\/news\/#organization"},"alumniOf":{"@type":"EducationalOrganization","name":"Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU)","url":"https:\/\/www.fau.eu"}}]}},"_links":{"self":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/6152","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/comments?post=6152"}],"version-history":[{"count":2,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/6152\/revisions"}],"predecessor-version":[{"id":6303,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/6152\/revisions\/6303"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/media\/6302"}],"wp:attachment":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/media?parent=6152"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/categories?post=6152"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/tags?post=6152"},{"taxonomy":"company","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/companies?post=6152"},{"taxonomy":"product","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/products?post=6152"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}