[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273358-105":59,"doc-detail-273358-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","scale-or-reason-a-compute-equivalent-analysis-of-reasoning-distillation","SCALE OR REASON - A COMPUTE-EQUIVALENT ANALYSIS OF REASONING DISTILLATION","","Distilling reasoning traces from strong teacher models is a common recipe for training capable small language models, but reasoning traces are 5–20× longer than standard IFT outputs, increasing compute consumption. A controlled experiment compares IFT and reasoning supervision by toggling only the teacher’s reasoning mode while keeping prompts identical. Students trained at five scales (0.5B–14B) are evaluated on 18 benchmarks to assess compute-efficient trade-offs across tasks.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/scale-or-reason-a-compute-equivalent-analysis-of-reasoning-distillation/273358/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/scale-or-reason-a-compute-equivalent-analysis-of-reasoning-distillation/273358.png","ImageObject",300,407,{"name":92,"@type":93},"Logic","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-27","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":44},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What problem does the document study about reasoning distillation?","Question",{"text":112,"@type":113},"It studies whether investing compute in reasoning distillation is worthwhile compared with training a larger standard instruction fine-tuning (IFT) model under the same compute budget.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How is the comparison between IFT and reasoning distillation controlled?",{"text":117,"@type":113},"A single teacher generates paired IFT and reasoning outputs for identical prompts by toggling only its reasoning mode, isolating supervision format as the sole variable.",{"name":119,"@type":110,"acceptedAnswer":120},"What general finding is reported at matched compute (FLOPs)?",{"text":121,"@type":113},"At matched training and inference FLOPs, IFT models lie on or near the Pareto frontier across most configurations, while reasoning approaches the frontier mainly on open-ended tasks at 7B and above.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},273358,1789471048,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":44,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":129,"read_time":144,"preview_page_count":4},1099513958762,"https://ap-avatar.wpscdn.com/avatar/1000023916a998db790?x-image-process=image/resize,m_fixed,w_180,h_180&k=1784791008015729253","arXiv : 2509 .22 193v 3 [cs .CL] 2 1 Aug 2026  \nSCALE OR REASON?  \nA COMPUTE-EQUIVALENT ANALYSIS OF REASONING DISTILLATION  \nNicolas Boizard1,5 Hippolyte Gisserot-Boukhlef2,5  \nKevin El Haddad 1,4 Céline Hudelot5 Pierre Colombo3  \n1Diabolocom 2Artefact Research Center 3 Cohere 4ISIA Lab, University of Mons  \n5MICS, CentraleSupélec, Université Paris-Saclay  \nAbstract: Distilling reasoning traces from strong teacher models has become the standard recipe for building capable small language models. Yet reasoning traces are 5-20 × longer than standard instruction fine-tuning (IFT) outputs, meaning every practitioner who chooses reasoning distillation implicitly forgoes training a larger IFT model on the same compute budget. Whether this trade-off is worthwhile remains unaddressed. We study it with a controlled experiment: a single teacher generates paired IFT and reasoning outputs for identical prompts by toggling only its reasoning mode, isolating supervision format as the sole variable. Training students at five scales (0.5B to 14B) and evaluating on 18 benchmarks, we find that at matched FLOPs, IFT lies on or near the Pareto frontier across the majority of configurations. Reasoning reaches the Pareto frontier only on open-ended tasks at 7B and above. Even there, a sequential curriculum mixing just 25-50% reasoning data with IFT captures most of the accuracy benefit at far lower compute cost.  \nCorrespondence: [nicolas.boizard@centralesupelec.fr](nicolas.boizard@centralesupelec.fr)  \nWebsite: [https://huggingface.co/scale-or-reason](https://huggingface.co/scale-or-reason)  \nDate: September 26, 2025  \n1 Introduction  \nReasoning distillation has become the dominant training recipe for capable small language models (Guo et al., 2025 ; Yang et al., 2025 ; Rastogiet al., 2025) . Yet every reasoning trace is 5-20 × longer than a standard IFT output (Figure 1), and this length gap has a direct cost: training on reasoning data consumes proportionally more compute. A practitioner with a fixed budget who chooses reasoning distillation implicitly gives up the option of training a larger IFT model instead, an alternative that standard scaling laws suggest is a highly efficient use of additional compute (Kaplan et al., 2020 ; Hoffmann et al., 2022) . Whether reasoning justifies this cost has not been studied: DeepSeekR1 (Guo et al., 2025), s1 (Muennighoffet al., 2025), and Snell et al. (2024) all evaluate at fixed model sizes, leaving the compute-allocation question entirely open. This leads directly to our question:  \nFor a given compute budget, should a practitioner invest in reasoning distillation, or train a larger IFT model?  \nFigure 1: Compute overhead of reasoning distillation. Each point is a benchmark-model-size pair. X-axis: extra-token factor when switching from IFT to reasoning supervision. Y-axis: accuracy gain (%) . The overhead is large (5-20 ×) and heterogeneous across tasks.  \nTo isolate the effect of supervision format from all other variables, we use a single teacher that generates paired IFT and reasoning outputs for identi-  \ncal prompts by toggling its reasoning mode on and off. Both formats are thus distilled from the same teacher weights in a controlled setup, ensuring that supervision format is the sole variable. We train Qwen2 .5 students at five scales (0 .5B to 14B), track FLOPs precisely across all configurations, and evaluate on 18 benchmarks spanning four task families. We replicate all key findings using a second teacherstudent pair (Nemotron-Super-49B with Gemma-3 ; Appendix G), confirming that results are not specific to a single architecture or teacher. This yields three findings:  \n• Scaling IFT is compute-efficient by default. At matched training and inference FLOPs, IFT models lie on or near the Pareto frontier across the majority of configurations. Reasoning only approaches the frontier on open-ended tasks and at larger scales (§3) .  \n• Format and capacity jointly determine the value of reasoning. Outpu","cbCaihbC1wxNjeYM","https://ap.wps.com/l/cbCaihbC1wxNjeYM","pdf",4037979,25,"English","# Introduction\n## Experimental Setup\n### Distillation\n### Students\n## Training","[{\"question\":\"What problem does the document study about reasoning distillation?\",\"answer\":\"It studies whether investing compute in reasoning distillation is worthwhile compared with training a larger standard instruction fine-tuning (IFT) model under the same compute budget.\"},{\"question\":\"How is the comparison between IFT and reasoning distillation controlled?\",\"answer\":\"A single teacher generates paired IFT and reasoning outputs for identical prompts by toggling only its reasoning mode, isolating supervision format as the sole variable.\"},{\"question\":\"What general finding is reported at matched compute (FLOPs)?\",\"answer\":\"At matched training and inference FLOPs, IFT models lie on or near the Pareto frontier across most configurations, while reasoning approaches the frontier mainly on open-ended tasks at 7B and above.\"}]","SCALE OR REASON - A COMPUTE-EQUIVALENT ANALYSIS OF REASONING DISTILLATION | PDF",63]