[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-teaching-ai-to-learn-from-noisy-human-feedback":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},3141,"teaching-ai-to-learn-from-noisy-human-feedback","Teaching AI to Learn From Noisy Human Feedback","Researchers have built the first provably robust offline RLHF methods, designed to recover reliable AI behavior even when training data is partially corrupted.","New research tackles a weak point in how AI systems learn from people: what happens when the feedback itself is bad?\n\nA paper published on arXiv introduces the first algorithms with provable guarantees for training AI policies using corrupted human preference data — a setting researchers call corruption robust offline reinforcement learning with human feedback (RLHF). The setup is realistic: given a dataset of trajectory pairs with human preference labels, some fraction of those labels may be flipped or the trajectory features manipulated, either by an adversary or simply by inconsistent human raters. The researchers designed methods that can still identify a near-optimal policy from that noisy data, under varying assumptions about how the training data was collected.\n\nThis matters because RLHF is the backbone of most modern large language model alignment pipelines — it is how models like GPT-4 and Claude are steered toward useful, safe behavior. If a fraction of preference labels are wrong, either through data poisoning or rater fatigue, the resulting model can drift in unpredictable directions. Prior theoretical work handled corrupted rewards or clean RLHF separately; this is the first to address both at once.\n\nThe approach works by learning a reward model with confidence sets, then deriving a pessimistic policy — a standard offline RL trick, applied here in a corruption-aware way. The authors show this can be done using different types of optimization oracles depending on data coverage assumptions. Notably, the guarantees are provable, not just empirical, which is a higher bar than most published RLHF work clears. Whether the assumed corruption fractions map to real-world rater noise in production pipelines is a question the paper leaves for future work.","[\"ai\",\"machine-learning\",\"rlhf\",\"security\"]","2026-07-01T04:00:00.000Z","2026-07-01T08:14:08.265Z","2026-07-01T08:14:14.108Z","published",null,[],"ai",[24,26,27,28],"machine-learning","rlhf","security",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.06734",0,{"sections":35},[36,40,44,49,54,59,64,69,74,79,84,88,93,98],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":41,"slug":28,"count":42,"latest_published_at":43},"Security",294,"2026-07-15T19:59:48.000Z",{"name":45,"slug":46,"count":47,"latest_published_at":48},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":78},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":85,"slug":86,"count":82,"latest_published_at":87},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]