[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-protein-ai-learns-to-revisit-its-own-mistakes":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},8177,"protein-ai-learns-to-revisit-its-own-mistakes","Protein AI Learns to Revisit Its Own Mistakes","Spectral Feedback lets protein diffusion models re-edit weak token choices instead of committing to them, lifting stable-protein yields up to 32.3%.","A new technique called Spectral Feedback lets AI models designing proteins go back and fix their own bad guesses instead of committing to them forever.\n\nResearchers built Spectral Feedback for discrete diffusion models used in protein inverse folding, the task of finding an amino acid sequence that folds into a target shape. Instead of only nudging token choices forward during generation, the method re-masks and re-samples specific positions after the fact, similar to how image-editing tools reintroduce noise and rerun a diffusion process. The hard part is picking which positions to revisit, since editing one token changes the value of editing another. The team found that these edit-position value functions have a sparse structure that makes them cheap to learn and optimize.\n\nThis matters because most alignment tricks for diffusion models assume you only get one shot at each token, then patch quality afterward with sampling tricks like Best-of-N or heavier fine-tuning. Spectral Feedback works on top of any of those setups, whether pretrained, test-time-aligned, or RL fine-tuned, without touching the underlying generative process, and it delivered gains in every case: 32.3% more stable proteins for a pretrained model, 24.8% for Best-of-10 sampling, and 5.8% on top of an already RL fine-tuned model.\n\nThat last number is the honest one. A 5.8% bump on a model already tuned with reinforcement learning suggests the easy gains in protein diffusion alignment are shrinking, even as the technique itself looks genuinely useful.","[\"ai\",\"protein-diffusion\",\"alignment\",\"biotech\"]","2026-09-28T04:00:00.000Z","2026-09-28T12:53:02.057Z","2026-09-28T12:53:09.285Z","published",null,[],"ai",[24,26,27,28],"protein-diffusion","alignment","biotech",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.30456",0,{"sections":35},[36,39,43,48,53,58,62,67,72,77,82,87,91,96],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",4799,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",762,{"name":44,"slug":45,"count":46,"latest_published_at":47},"Policy","policy",399,"2026-09-27T18:39:02.000Z",{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",261,"2026-09-27T15:30:35.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",188,"2026-09-27T20:46:36.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":18},"Science","science",151,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",135,"2026-09-26T14:30:00.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":71},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Dev Tools","dev-tools",84,"2026-09-26T04:20:58.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Startups","startups",76,"2026-09-25T18:33:59.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":88,"slug":89,"count":85,"latest_published_at":90},"General","general","2026-09-26T17:02:42.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",30,"2026-09-24T20:07:31.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]