[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-tool-checks-which-ai-image-features-can-actually-be-steered":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},9162,"new-tool-checks-which-ai-image-features-can-actually-be-steered","New Tool Checks Which AI Image Features Can Actually Be Steered","D-Scope tests whether sparse-autoencoder features in diffusion image models can actually steer generation, not just explain it.","Researchers built a way to check whether the \"interpretable\" features found inside AI image generators actually do anything when you pull their levers.\n\nA team introduced D-Scope, a framework for testing sparse autoencoders (SAEs), a popular tool for finding human-readable patterns inside diffusion transformers, the architecture behind many modern image generators. Until now there was no reliable way to confirm that a given feature could actually control what the model generates, rather than just correlate with it. D-Scope matches target text descriptions against clusters of image patches that most strongly activate each feature, using SigLIP 2 embeddings to find candidates without hand-labeling every one, then tests the underlying \"decoder direction\" by masking parts of an image and nudging that direction at different strengths under fixed generation settings. The team ran this across 150 SAEs spanning two model families and five layers, and built a benchmark of 100 target concepts tested across ten contexts each.\n\nThe finding worth noting is a gap between how a feature looks and what it does: models can reconstruct images with high fidelity while using only a fraction of their learned feature \"dictionary,\" and many features have thin visual evidence behind their supposed meaning. That undercuts the assumption that an SAE feature readout tells you what a model actually knows. Steering by directly retrieving a feature that matches a word also proved less reliable than contrastive retrieval, which produced bigger gains in the targeted region, though neither approach reliably left the rest of the image alone.\n\nInterpretability research keeps promising a dashboard of control knobs for generative models. D-Scope is mainly useful for showing how many of those knobs are still just painted on.","[\"ai\",\"diffusion-models\",\"interpretability\",\"sparse-autoencoders\"]","2026-10-01T04:00:00.000Z","2026-10-01T23:17:18.554Z","2026-10-01T23:17:19.961Z","published",null,[],"ai",[24,26,27,28],"diffusion-models","interpretability","sparse-autoencoders",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.39625",0,{"sections":35},[36,39,43,47,52,57,61,66,71,75,80,85,90,95],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",5597,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",815,{"name":44,"slug":45,"count":46,"latest_published_at":18},"Policy","policy",430,{"name":48,"slug":49,"count":50,"latest_published_at":51},"Deals","deals",298,"2026-09-30T21:00:26.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":56},"Hardware","hardware",196,"2026-09-30T13:00:00.000Z",{"name":58,"slug":59,"count":60,"latest_published_at":18},"Science","science",163,{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",149,"2026-09-30T22:57:11.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Dev Tools","dev-tools",93,"2026-10-01T02:30:48.000Z",{"name":72,"slug":73,"count":69,"latest_published_at":74},"Software","software","2026-09-30T21:41:11.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Startups","startups",84,"2026-09-30T20:39:09.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"Gaming","gaming",51,"2026-09-30T16:24:30.000Z",{"name":86,"slug":87,"count":88,"latest_published_at":89},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":91,"slug":92,"count":93,"latest_published_at":94},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]