[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-teaching-code-ai-to-think-like-an-architect":10,"sections":40},{"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":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},4450,"teaching-code-ai-to-think-like-an-architect","Teaching Code AI to Think Like an Architect","A new fine-tuning pipeline beats the base Qwen3 model by 540% on a standard software benchmark by training it to judge architectural quality.","Researchers found that code-generating AI gets better when you teach it to care about architecture, not just passing tests.\n\nA team published a method for labeling training data using two AI judges instead of human experts. The first judge estimates how much architectural understanding a given coding task actually requires. The second checks whether a proposed code patch follows the conventions of the specific repository it belongs to — not just whether it works in isolation. Fine-tuning Qwen3 models ranging from 8B to 32B parameters on 3,360 curated examples produced a resolved rate of up to 27.2% on SWE-bench Verified. That 27.2% represents a 540% improvement over the base Qwen3 model and a 256% improvement over a version fine-tuned on unfiltered data — meaning both the starting point and the filtering method matter, and the two numbers measure different things.\n\nMost code AI benchmarks reward models for producing patches that pass tests. That misses something: real codebases have conventions, abstractions, and structural patterns that a correct-but-oblivious patch can quietly erode. This research makes architectural conformance a first-class training signal, which is harder to game than a test suite.\n\nThe catch is that the judging pipeline itself relies on a strong LLM, so the quality of the whole system is upstream-dependent — a limitation the paper acknowledges but cannot fully resolve.","[\"ai\",\"code\",\"llms\",\"benchmarks\"]","2026-07-08T04:00:00.000Z","2026-07-08T09:42:23.140Z","2026-07-08T09:42:25.952Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"The lede and dek frame the 256% figure as the headline result, but the source reports 540% over the base model as the larger gain — the draft buries the stronger number and misrepresents which comparison is the 'best result'; also, '540% improvement' and '256% over unfiltered fine-tuning' are improvements over different baselines and the draft conflates them in a way that will confuse readers, requiring a clearer explanation of what each percentage actually compares.","resolved","ai",[30,32,33,34],"code","llms","benchmarks",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.14948",0,{"sections":41},[42,46,51,56,61,66,71,76,81,86,91,95,100,105],{"name":43,"slug":30,"count":44,"latest_published_at":45},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":92,"slug":93,"count":89,"latest_published_at":94},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":101,"slug":102,"count":103,"latest_published_at":104},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":106,"slug":107,"count":108,"latest_published_at":109},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]