[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-method-keeps-ai-unlearning-intact-after-compression":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},7555,"new-method-keeps-ai-unlearning-intact-after-compression","New Method Keeps AI Unlearning Intact After Compression","A new technique targets the specific weights that let AI models quietly relearn forgotten data once they're compressed for deployment.","A new technique aims to stop AI unlearning from quietly failing once a model gets compressed for deployment.\n\nUnlearning is the process of scrubbing an LLM's memory of specific training data, usually to satisfy copyright takedowns or privacy requests, without retraining the whole model from scratch. Researchers found that when unlearned models go through quantization, the standard compression step used to shrink models for real-world use, the forgetting effect degrades faster than the model's general usefulness does. Their fix locates the specific weights responsible for that fragility through a curvature analysis of the loss landscape, then adds targeted noise to push those weights toward a flatter, more stable minimum. A second piece, called forget-critical optimization, updates only the layers tied to the forgotten data instead of touching the whole network. Tested on the MUSE and TOFU unlearning benchmarks across several existing unlearning algorithms, the combined approach held up better after quantization while keeping model utility intact.\n\nThe gap matters because quantization is not an edge case. Most LLMs deployed on phones, laptops, or cost-conscious servers run in 4-bit or 8-bit form, not the full-precision version they were unlearned in. A company could believe it has honored a deletion request while the compressed model actually in production has not.\n\nWorth remembering: none of this deletes data from the model in any literal sense. It just tunes the weights so the information is harder to surface, which means robust to quantization is not the same as robust to someone trying harder to get it out.","[\"ai\",\"machine-unlearning\",\"quantization\",\"llm-safety\"]","2026-09-24T04:00:00.000Z","2026-09-24T06:22:25.188Z","2026-09-24T06:22:31.414Z","published",null,[],"ai",[24,26,27,28],"machine-unlearning","quantization","llm-safety",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.27355",0,{"sections":35},[36,39,43,48,53,58,63,68,73,78,83,88,93,98],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",4424,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",724,{"name":44,"slug":45,"count":46,"latest_published_at":47},"Policy","policy",380,"2026-09-23T22:53:43.000Z",{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",221,"2026-09-24T08:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",173,"2026-09-23T23:42:10.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Science","science",136,"2026-09-24T09:00:00.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",116,"2026-09-24T00:51:49.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",85,"2026-09-23T20:00:00.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":77},"Dev Tools","dev-tools",79,"2026-09-22T22:21:13.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",66,"2026-09-23T17:28:38.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"Gaming","gaming",45,"2026-09-22T15:35:06.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",43,"2026-09-21T23:48:56.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",27,"2026-09-22T13:00:00.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]