[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-tiny-network-replaces-some-prompt-examples-not-all":10,"sections":35},{"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":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},9638,"tiny-network-replaces-some-prompt-examples-not-all","Tiny Network Replaces Some Prompt Examples, Not All","Researchers trained a small network to replace few-shot prompts and found it beats them on grammar rules but stalls on arbitrary word pairs like antonyms.","A new study shows you can swap out few-shot prompting for a tiny trained module, but only when the task follows a rule instead of a memorized fact.\n\nThe researchers trained a 2.6 million parameter network to read the few-shot examples normally stuffed into a prompt, then inject an equivalent signal directly into the activations of a frozen GPT-2-large or GPT-2-XL model, skipping the examples entirely at inference time. They tested it on eight grammatical inflection tasks, like changing verb tense, plus one task based on antonym pairs. On straightforward tense changes, the module matched standard few-shot prompting while costing nothing extra per query. On lemmatization, reducing a word to its root form, it did significantly better than prompting, scoring up to 72 percentage points higher, because ten examples in a prompt often fail to convey that particular rule even though the underlying model can execute it. On antonyms, the module plateaued at roughly half of what prompting could achieve, regardless of how it was scaled up or retrained.\n\nThis matters for anyone building systems that cache or compress prompts to save on inference costs. The split is clean: operations with an underlying rule, like verb conjugation, compress into a small reusable signal, sometimes better than prompting itself conveys them, while operations built on memorized, arbitrary pairings, like antonyms, don't compress at all.\n\nIn other words, the trick works when there's actually a trick to find. Ask a model to recall a fact rather than apply a rule, and no amount of clever engineering around its internals will make that lookup any cheaper.","[\"in-context-learning\",\"language-models\",\"ai-research\",\"gpt-2\"]","2026-10-02T04:00:00.000Z","2026-10-03T04:09:45.998Z","2026-10-03T04:09:52.349Z","published",null,[],"ai",[26,27,28,29],"in-context-learning","language-models","ai-research","gpt-2",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.00526",0,{"sections":36},[37,40,44,48,53,57,61,66,71,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",5976,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",842,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",438,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":18},"Hardware","hardware",199,{"name":58,"slug":59,"count":60,"latest_published_at":18},"Science","science",173,{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]