[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-teach-ai-to-highlight-evidence-before-summarizing":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},8068,"researchers-teach-ai-to-highlight-evidence-before-summarizing","Researchers Teach AI to Highlight Evidence Before Summarizing","A new open-source model highlights the relevant evidence in long documents before summarizing it, then beats a much larger rival on a long-context benchmark.","A new open-source AI model skims a long document for the relevant parts before it writes an answer.\n\nResearchers built a system called Highlight-Then-Summarize, or H2S, that separates reading from answering. It first pulls out the sentences and passages in a document that actually bear on the question, then compresses them into a short, question-focused summary, and only then generates a final answer. The team trained the approach on a new dataset of 6,647 examples pulled from 11 existing benchmarks, with documents averaging 43.9K tokens long, using a reinforcement-learning method that rewards good evidence selection and summary writing, not just a correct final answer. On a seven-task long-context test suite, the 14-billion-parameter version, H2S-14B, scored 32.60 on average under a 128K-token input and 4K-token output budget, 10.17 points ahead of Qwen3.8-27B, a model roughly twice its size, and the best score among the open-source systems tested.\n\nThat's the interesting part: a smaller model beat a considerably larger one by getting pickier about what it reads, not by getting bigger. H2S-14B also held onto 97.1% of its performance when its own output budget was cut from 16K tokens down to 4K, suggesting the gains come from smarter evidence handling rather than just more room to write. For anyone running long-context AI at scale, that matters: shorter outputs mean cheaper inference without giving up much accuracy.\n\nIt's a reminder that most 'long-context' benchmarks are really testing needle-in-haystack retrieval dressed up as reasoning, and that throwing a bigger context window at a model isn't the same as teaching it what to ignore.","[\"ai\",\"llms\",\"research\",\"long-context\"]","2026-09-28T04:00:00.000Z","2026-09-28T08:08:27.341Z","2026-09-28T08:08:34.348Z","published",null,[],"ai",[24,26,27,28],"llms","research","long-context",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.31382",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",4750,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",759,{"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"]