[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-algorithm-makes-differentiable-top-k-selection-fast":10,"sections":35},{"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":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},9239,"new-algorithm-makes-differentiable-top-k-selection-fast","New Algorithm Makes Differentiable Top-k Selection Fast","A new GPU algorithm picks exact top-k elements from 100 million scores in milliseconds, making differentiable sparsity practical at scale.","A new algorithm lets AI systems pick the exact best k items out of a hundred million in milliseconds, without breaking the smooth math that models need to learn.\n\nResearchers introduced Fast LapSum, a GPU solver for differentiable top-k selection that keeps an exact selection budget of k while remaining fully differentiable. Instead of sorting every score, it uses probabilistic bracketing: it narrows the search to a thin band around the likely cutoff using noised samples, then runs a certification pass to verify the result, falling back to a full sort only in the worst case. The certified GPU implementation processes a million scores in 0.92 milliseconds, ten million in 1.39 milliseconds, and a hundred million in 7.24 milliseconds. The team tested it on two problems: generating sparse adversarial image attacks that perturb only a small fraction of pixels, and trimming the number of 3D Gaussians needed for 3D Gaussian splatting renders.\n\nTop-k selection quietly underpins a lot of large-scale AI infrastructure - deciding which experts to activate in a mixture-of-experts model, which tokens to route, or which attention weights to keep. Picking a hard, countable subset usually breaks the smooth gradients that training depends on, forcing a choice between exactness and trainability. Fast LapSum's pitch is that it removes that tradeoff at the scale where routing and attention pruning actually happen in production models.\n\nOn the adversarial-image task, it beat existing state-of-the-art methods by an order of magnitude - a sizable claim for something that is, at its core, sorting with extra steps.","[\"ai-research\",\"gpu-computing\",\"sparse-models\",\"3d-gaussian-splatting\"]","2026-10-01T04:00:00.000Z","2026-10-02T03:29:24.103Z","2026-10-02T03:29:28.475Z","published",null,[],"ai",[26,27,28,29],"ai-research","gpu-computing","sparse-models","3d-gaussian-splatting",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.06912",0,{"sections":36},[37,41,45,49,54,58,62,67,72,77,82,87,92,97],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",6001,"2026-10-02T04:00:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":40},"Security","security",843,{"name":46,"slug":47,"count":48,"latest_published_at":40},"Policy","policy",438,{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":40},"Hardware","hardware",199,{"name":59,"slug":60,"count":61,"latest_published_at":40},"Science","science",175,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":71},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]