[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-goldimask-improves-how-diffusion-language-models-learn-to-fill-blanks":10,"sections":34},{"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":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},8864,"goldimask-improves-how-diffusion-language-models-learn-to-fill-blanks","GoldiMask Improves How Diffusion Language Models Learn to Fill Blanks","GoldiMask, a new fine-tuning method, picks mask tokens more carefully, lifting accuracy on reasoning and code tasks while cutting decoding steps.","Researchers have found a better way to train diffusion language models by being pickier about which words they hide during practice.\n\nDiffusion language models learn by having some words in a response masked out, then trying to guess them from whatever is still visible. Most training setups mask tokens randomly, which the researchers argue is wasteful: a random mask might hide an easy word while leaving a genuinely hard one in plain sight for free. The new method, called GoldiMask, uses a scoring approach to decide which tokens to reveal as context versus which to mask as a prediction target, weighing how useful each word is for understanding against how useful it is as something to guess. It then adjusts how much training weight each masked word gets based on how much the model still has to learn from guessing it. Tested across three different model backbones and three training datasets, GoldiMask beat standard random masking on accuracy in most settings, including reasoning problems and code generation, and let models hit the same confidence level in fewer decoding steps on the GSM8K and MATH-500 math benchmarks.\n\nDiffusion language models are the main challenger to the token-by-token autoregressive models that power most chatbots today, generating text by filling in masked blanks in parallel rather than one word at a time. Training them well has been fiddly, and this work suggests the fix is less about scale and more about being deliberate on what the model practices guessing - a cheap lever other diffusion model builders could try.\n\nIt is one paper with no claim of beating top autoregressive models outright, but it's a useful reminder that how you quiz a model can matter as much as how big the class is.","[\"ai\",\"language-models\",\"research\",\"diffusion-models\"]","2026-10-01T04:00:00.000Z","2026-10-01T08:13:06.571Z","2026-10-01T08:13:11.949Z","published",null,[],"ai",[24,26,27,28],"language-models","research","diffusion-models",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.38385",0,{"sections":35},[36,39,44,49,54,59,64,69,74,78,83,88,93,98],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",5269,{"name":40,"slug":41,"count":42,"latest_published_at":43},"Security","security",801,"2026-09-30T22:18:23.000Z",{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",429,"2026-10-01T02:26:17.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",298,"2026-09-30T21:00:26.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",196,"2026-09-30T13:00:00.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Science","science",157,"2026-09-30T15:00:56.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Consumer Tech","consumer-tech",149,"2026-09-30T22:57:11.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Dev Tools","dev-tools",93,"2026-10-01T02:30:48.000Z",{"name":75,"slug":76,"count":72,"latest_published_at":77},"Software","software","2026-09-30T21:41:11.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",84,"2026-09-30T20:39:09.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"Gaming","gaming",51,"2026-09-30T16:24:30.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]