{"id":11176,"date":"2025-06-04T14:24:21","date_gmt":"2025-06-04T14:24:21","guid":{"rendered":"https:\/\/novelis.io\/?post_type=research-lab&#038;p=11176"},"modified":"2025-07-07T09:42:49","modified_gmt":"2025-07-07T09:42:49","slug":"adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir","status":"publish","type":"research-lab","link":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/","title":{"rendered":"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir"},"content":{"rendered":"\n<p>Les grands mod\u00e8les de raisonnement (LRMs) ont montr\u00e9 des performances impressionnantes en g\u00e9n\u00e9rant des \u00e9tapes interm\u00e9diaires d\u00e9taill\u00e9es avant de produire une r\u00e9ponse (un processus souvent qualifi\u00e9 de \u00ab r\u00e9flexion \u00bb). Cela am\u00e9liore les r\u00e9sultats sur les t\u00e2ches complexes, mais engendre aussi une lourde charge computationnelle, en particulier pour les probl\u00e8mes simples o\u00f9 une telle r\u00e9flexion est superflue. En r\u00e9sum\u00e9 : ces mod\u00e8les ont tendance \u00e0 <em>sur-r\u00e9fl\u00e9chir<\/em>, m\u00eame pour des questions faciles, gaspillant temps et ressources.<\/p>\n\n\n\n<p><strong>AdaptThink<\/strong> est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement, selon la difficult\u00e9 du probl\u00e8me :<\/p>\n\n\n\n<p><strong>Mode Thinking<\/strong> : engage un raisonnement pas \u00e0 pas<br><strong>Mode NoThinking<\/strong> : saute les \u00e9tapes interm\u00e9diaires et donne directement la r\u00e9ponse<\/p>\n\n\n\n<p>Ce cadre repose sur deux \u00e9l\u00e9ments cl\u00e9s :<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Un <strong>objectif d\u2019optimisation contraint<\/strong>, qui pousse le mod\u00e8le \u00e0 privil\u00e9gier le mode NoThinking pour les t\u00e2ches simples, tout en maintenant une pr\u00e9cision globale ;<\/li>\n<\/ul>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Une <strong>strat\u00e9gie d\u2019\u00e9chantillonnage par importance<\/strong>, garantissant une exposition \u00e9quilibr\u00e9e aux deux modes pendant l\u2019entra\u00eenement, pour un apprentissage efficace et diversifi\u00e9.<\/li>\n<\/ul>\n\n\n\n<p>Les auteurs ont \u00e9valu\u00e9 AdaptThink sur trois jeux de donn\u00e9es de raisonnement math\u00e9matique, en utilisant le mod\u00e8le <strong>DeepSeek-R1-Distill-Qwen-1.5B<\/strong>. Les principaux r\u00e9sultats sont :<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Gains d\u2019efficacit\u00e9<\/strong> : r\u00e9duction de 53 % de la longueur moyenne des r\u00e9ponses, en limitant la r\u00e9flexion inutile<\/li>\n\n\n\n<li><strong>Gains de performance<\/strong> : am\u00e9lioration de 2,4 % de la pr\u00e9cision, prouvant que l\u2019efficacit\u00e9 accrue ne nuit pas \u00e0 la justesse<\/li>\n<\/ul>\n\n\n\n<p>Ces r\u00e9sultats sugg\u00e8rent qu\u2019AdaptThink peut g\u00e9rer efficacement le compromis entre profondeur de raisonnement et efficacit\u00e9 computationnelle.<\/p>\n\n\n\n<p>\u26a0\ufe0f Quelques limites \u00e0 noter : pour l\u2019instant, le cadre n\u2019a \u00e9t\u00e9 test\u00e9 que sur des probl\u00e8mes math\u00e9matiques, donc son potentiel de g\u00e9n\u00e9ralisation \u00e0 d\u2019autres types de t\u00e2ches reste \u00e0 prouver. De plus, comme il repose sur l\u2019apprentissage par renforcement, son entra\u00eenement est plus complexe et co\u00fbteux que les approches classiques.<\/p>\n\n\n\n<p><strong>Pour aller plus loin :<\/strong><br>\ud83d\udcc4 <a class=\"\" href=\"https:\/\/arxiv.org\/abs\/2505.13417\" target=\"_blank\" rel=\"noopener\">AdaptThink: Reasoning Models Can Learn When to Think<\/a><br>\ud83d\udcc4 <a class=\"\" href=\"https:\/\/arxiv.org\/abs\/2505.13379\" target=\"_blank\" rel=\"noopener\">Thinkless: LLM Learns When to Think<\/a><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img fetchpriority=\"high\" decoding=\"async\" width=\"1200\" height=\"1200\" src=\"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/AdaptThink.jpg\" alt=\"\" class=\"wp-image-11170\" style=\"width:569px;height:auto\" srcset=\"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/AdaptThink.jpg 1200w, https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/AdaptThink-600x600.jpg 600w, https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/AdaptThink-250x250.jpg 250w, https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/AdaptThink-768x768.jpg 768w, https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/AdaptThink-30x30.jpg 30w\" sizes=\"(max-width: 1200px) 100vw, 1200px\" \/><\/figure>\n\n\n\n<p><\/p>\n","protected":false},"featured_media":11174,"template":"","categories":[510],"custom_tag":[524,525],"class_list":["post-11176","research-lab","type-research-lab","status-publish","has-post-thumbnail","hentry","category-lab-news-2","custom_tag-lrm-2","custom_tag-machine-learning-2"],"acf":{"externel_link":"","summary":"","filter_opacity":"70","subtitle":"","reading_time":"","authors":"","document_to_download":{"upload_a_file":false,"download_without_form":false,"file":false,"url":""},"show_recent_block_on_the_bottom_of_the_page":false},"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v25.6 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir<\/title>\n<meta name=\"description\" content=\"AdaptThink est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir\" \/>\n<meta property=\"og:description\" content=\"AdaptThink est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement\" \/>\n<meta property=\"og:url\" content=\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/\" \/>\n<meta property=\"og:site_name\" content=\"Novelis innovation\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/novelis.io\" \/>\n<meta property=\"article:modified_time\" content=\"2025-07-07T09:42:49+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"2560\" \/>\n\t<meta property=\"og:image:height\" content=\"1440\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:site\" content=\"@novelis_io\" \/>\n<meta name=\"twitter:label1\" content=\"Dur\u00e9e de lecture estim\u00e9e\" \/>\n\t<meta name=\"twitter:data1\" content=\"2 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/\",\"url\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/\",\"name\":\"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir\",\"isPartOf\":{\"@id\":\"https:\/\/novelis.io\/fr\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg\",\"datePublished\":\"2025-06-04T14:24:21+00:00\",\"dateModified\":\"2025-07-07T09:42:49+00:00\",\"description\":\"AdaptThink est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement\",\"breadcrumb\":{\"@id\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#breadcrumb\"},\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#primaryimage\",\"url\":\"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg\",\"contentUrl\":\"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg\",\"width\":2560,\"height\":1440},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Accueil\",\"item\":\"https:\/\/novelis.io\/fr\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/novelis.io\/fr\/#website\",\"url\":\"https:\/\/novelis.io\/fr\/\",\"name\":\"Novelis innovation\",\"description\":\"Novelis innovation\",\"publisher\":{\"@id\":\"https:\/\/novelis.io\/fr\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/novelis.io\/fr\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"fr-FR\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/novelis.io\/fr\/#organization\",\"name\":\"Novelis innovation\",\"url\":\"https:\/\/novelis.io\/fr\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\/\/novelis.io\/fr\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/novelis.io\/wp-content\/uploads\/2021\/12\/logo-1.png\",\"contentUrl\":\"https:\/\/novelis.io\/wp-content\/uploads\/2021\/12\/logo-1.png\",\"width\":479,\"height\":98,\"caption\":\"Novelis innovation\"},\"image\":{\"@id\":\"https:\/\/novelis.io\/fr\/#\/schema\/logo\/image\/\"},\"sameAs\":[\"https:\/\/www.facebook.com\/novelis.io\",\"https:\/\/x.com\/novelis_io\",\"https:\/\/www.linkedin.com\/company\/novelis-consulting\/\",\"https:\/\/www.youtube.com\/channel\/UCJ5eJR22n2GtfKaTWueWRPQ\"]}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir","description":"AdaptThink est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement","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:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/","og_locale":"fr_FR","og_type":"article","og_title":"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir","og_description":"AdaptThink est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement","og_url":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/","og_site_name":"Novelis innovation","article_publisher":"https:\/\/www.facebook.com\/novelis.io","article_modified_time":"2025-07-07T09:42:49+00:00","og_image":[{"width":2560,"height":1440,"url":"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg","type":"image\/jpeg"}],"twitter_card":"summary_large_image","twitter_site":"@novelis_io","twitter_misc":{"Dur\u00e9e de lecture estim\u00e9e":"2 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/","url":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/","name":"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir","isPartOf":{"@id":"https:\/\/novelis.io\/fr\/#website"},"primaryImageOfPage":{"@id":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#primaryimage"},"image":{"@id":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#primaryimage"},"thumbnailUrl":"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg","datePublished":"2025-06-04T14:24:21+00:00","dateModified":"2025-07-07T09:42:49+00:00","description":"AdaptThink est un cadre bas\u00e9 sur l\u2019apprentissage par renforcement qui apprend aux LRMs \u00e0 choisir dynamiquement entre deux modes de raisonnement","breadcrumb":{"@id":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#breadcrumb"},"inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#primaryimage","url":"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg","contentUrl":"https:\/\/novelis.io\/wp-content\/uploads\/2025\/06\/image-Site-24-scaled.jpg","width":2560,"height":1440},{"@type":"BreadcrumbList","@id":"https:\/\/novelis.io\/fr\/research-lab\/adaptthink-apprendre-aux-modeles-de-raisonnement-quand-reflechir\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Accueil","item":"https:\/\/novelis.io\/fr\/"},{"@type":"ListItem","position":2,"name":"AdaptThink : apprendre aux mod\u00e8les de raisonnement quand r\u00e9fl\u00e9chir"}]},{"@type":"WebSite","@id":"https:\/\/novelis.io\/fr\/#website","url":"https:\/\/novelis.io\/fr\/","name":"Novelis innovation","description":"Novelis innovation","publisher":{"@id":"https:\/\/novelis.io\/fr\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/novelis.io\/fr\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"fr-FR"},{"@type":"Organization","@id":"https:\/\/novelis.io\/fr\/#organization","name":"Novelis innovation","url":"https:\/\/novelis.io\/fr\/","logo":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/novelis.io\/fr\/#\/schema\/logo\/image\/","url":"https:\/\/novelis.io\/wp-content\/uploads\/2021\/12\/logo-1.png","contentUrl":"https:\/\/novelis.io\/wp-content\/uploads\/2021\/12\/logo-1.png","width":479,"height":98,"caption":"Novelis innovation"},"image":{"@id":"https:\/\/novelis.io\/fr\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/novelis.io","https:\/\/x.com\/novelis_io","https:\/\/www.linkedin.com\/company\/novelis-consulting\/","https:\/\/www.youtube.com\/channel\/UCJ5eJR22n2GtfKaTWueWRPQ"]}]}},"_links":{"self":[{"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/research-lab\/11176"}],"collection":[{"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/research-lab"}],"about":[{"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/types\/research-lab"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/media\/11174"}],"wp:attachment":[{"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/media?parent=11176"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/categories?post=11176"},{"taxonomy":"custom_tag","embeddable":true,"href":"https:\/\/novelis.io\/fr\/wp-json\/wp\/v2\/custom_tag?post=11176"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}