[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-grokking-is-conditional-and-fragile-a-small-study-finds":10,"sections":45},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":34,"tags":35,"sources":40,"feedback":44,"feedback_at":22,"cost_usd":44,"total_tokens":44},4012,"grokking-is-conditional-and-fragile-a-small-study-finds","Grokking Is Conditional and Fragile, a Small Study Finds","Researchers using a tiny ~11,856-parameter model find grokking is a fragile phase transition, not a reliable sign of deep generalization.","A new paper argues that \"grokking\" is a fragile statistical artifact, not a reliable indicator of genuine learning.\n\nResearchers studied Glimmer-1-Base, which the paper describes as a publicly released Llama-style transformer with approximately 11,856 parameters. That is small enough to enumerate every weight, every attention pattern, and the complete input-output map. Instead of reporting a single dramatic training run, the team measured grokking as a multi-seed rate across modular arithmetic tasks. The approach overturned three striking single-run results in their own dataset, each one a seed confound.\n\nThe most unsettling finding is numerical. Two changes to the floating-point environment, switching between CPU and GPU or varying the CPU thread count, each flipped a minority of same-seed outcomes without any detectable shift in the aggregate rate. That means published grokking results can quietly change depending on hardware, and the field would have no way of knowing because multi-seed reporting has not been standard. The coverage threshold that gates grokking tracks output cardinality, specifically the modulus in modular arithmetic, more than task structure.\n\nYears of theorizing about late generalization may rest partly on results that were accidents of hardware configuration and lucky random seeds.","[\"grokking\",\"machine learning\",\"neural networks\",\"research\"]","2026-07-07T04:00:00.000Z","2026-07-07T14:50:35.471Z","2026-07-07T14:50:38.291Z","published",null,[24,30],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"The article names 'Glimmer-1-Base' as a publicly released model, but this name cannot be confirmed as a real, publicly released offering — it should be attributed as named in the paper with a hedge or citation, or the editors must verify it before publication.","resolved",{"id":31,"reviewer":26,"round":32,"reason":33,"status":29},"editor-r2",2,"The open concern [editor-r1] is still unresolved: the article names 'Glimmer-1-Base' as a publicly released model without a hedge or citation, and while the source abstract references it as 'publicly released,' the draft must either attribute this directly to the paper or add a hedge until independent verification is possible; additionally, the body states the model has 'roughly 12,000 parameters' while the source specifies '~11,856 parameters' — the rounding is not flagged as approximate, creat","ai",[36,37,38,39],"grokking","machine learning","neural networks","research",[41],{"name":42,"url":43},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.05104",0,{"sections":46},[47,51,56,61,66,71,76,81,86,90,95,99,104,109],{"name":48,"slug":34,"count":49,"latest_published_at":50},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":18},"Dev Tools","dev-tools",59,{"name":91,"slug":92,"count":93,"latest_published_at":94},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":96,"slug":97,"count":93,"latest_published_at":98},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":105,"slug":106,"count":107,"latest_published_at":108},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":110,"slug":111,"count":112,"latest_published_at":113},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]