[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-new-way-to-train-ai-agents-that-use-multiple-tools":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},11360,"a-new-way-to-train-ai-agents-that-use-multiple-tools","A New Way to Train AI Agents That Use Multiple Tools","PACT trains tool-using AI agents on expert examples only during training, not live use, to improve multi-turn tool-calling without copying fixed scripts.","AI agents that chain together multiple tool calls over several conversation turns are notoriously hard to train well, and a new paper proposes a fix called PACT.\n\nResearchers behind PACT, short for Privileged Trace Co-Training, say today's two main training methods both have flaws: reinforcement learning on these agents suffers from sparse rewards and weak credit assignment, while supervised fine-tuning on expert examples can lock a model into copying fixed scripts instead of learning to adapt. PACT tries to get the best of both by using expert traces only as a training-time signal, never as a hint during actual use. It combines a trace-conditioned RL surrogate that judges prompt-only rollouts against expert context, a component-aware supervised loss over reasoning and tool-call steps, and a prompt-only anchoring step that keeps the model grounded in standard RL updates. The team tested PACT on three agent benchmarks, FTRL, BFCL, and ToolHop, and reports it beat strong SFT and RL baselines on each.\n\nThat gap between training-time supervision and runtime behavior is the real story here. Multi-turn tool-use agents, the kind expected to search, query APIs, and chain actions across a conversation, keep tripping on the same problem: learn too rigidly from demonstrations and they can't adapt to new observations, or learn purely from reward and get little useful signal until the very end of a long trajectory. Splitting the difference, rather than picking one, is a reasonable approach.\n\nIt is worth remembering this is a single arXiv paper with self-reported benchmark wins, not a deployed system, and the listing is itself a replace-cross update to an earlier version.","[\"ai agents\",\"reinforcement learning\",\"tool use\",\"llm training\"]","2026-10-12T04:00:00.000Z","2026-10-12T10:12:36.623Z","2026-10-12T10:12:42.693Z","published",null,[],"ai",[26,27,28,29],"ai agents","reinforcement learning","tool use","llm training",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.16215",0,{"sections":36},[37,41,46,51,56,61,66,71,76,81,86,91,96,101],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",6967,"2026-10-12T15:20:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",954,"2026-10-12T14:20:37.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Policy","policy",496,"2026-10-11T20:06:43.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",492,"2026-10-12T14:55:53.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Hardware","hardware",237,"2026-10-12T13:24:45.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Science","science",196,"2026-10-11T15:34:40.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Consumer Tech","consumer-tech",192,"2026-10-12T14:07:13.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Startups","startups",121,"2026-10-12T13:22:34.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Software","software",115,"2026-10-11T07:34:45.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Dev Tools","dev-tools",111,"2026-10-12T11:22:24.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",71,"2026-10-12T16:10:40.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Gaming","gaming",59,"2026-10-09T11:43:43.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",35,"2026-10-12T14:00:00.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]