[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-crest-splits-credit-assignment-to-sharpen-tool-using-ai-agents":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},4949,"crest-splits-credit-assignment-to-sharpen-tool-using-ai-agents","CrEST Splits Credit Assignment to Sharpen Tool-Using AI Agents","CrEST splits AI agent training credit between reinforcement learning and a self-teacher, boosting performance on long multi-step tool-use tasks.","A new training recipe wants to fix a blind spot in how AI agents learn: they often get graded only on the final result, not on which step went wrong along the way.\n\nResearchers built CrEST, a training method for AI agents that use tools across multiple turns, like fetching data or running code, then reasoning about the result. Reinforcement learning with verifiable rewards is one way to train these agents, but it typically hands out one reward per full episode, blurring credit for individual turns. A rival approach, on-policy distillation, has the agent copy a stronger teacher model step by step, which gives finer feedback but caps performance at whatever the teacher can do. CrEST tries to combine both: it splits credit by turn using verified rewards, then uses a self-teacher to fine-tune token-level choices within each turn, only stepping in to adjust how strongly a signal counts rather than dictating the move itself.\n\nThat distinction matters because multi-step, tool-using agents are exactly what companies are betting on for coding assistants, research bots, and customer-service tools. Tested on two agent benchmarks, BFCL V3 and WildToolBench, CrEST beat both plain reinforcement learning and pure distillation, with the biggest gains showing up on long conversations and strict pass-fail scoring across a full session.\n\nIt is a benchmark win on paper so far, not a product; the real test is whether this credit-splitting trick survives contact with messier, real-world tool calls outside the lab.","[\"ai\",\"llm-agents\",\"reinforcement-learning\",\"arxiv\"]","2026-08-14T04:00:00.000Z","2026-08-14T20:10:30.877Z","2026-08-14T20:10:42.694Z","published",null,[],"ai",[24,26,27,28],"llm-agents","reinforcement-learning","arxiv",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.13179",0,{"sections":35},[36,40,44,49,54,59,64,69,74,79,84,89,94,99],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",3293,"2026-08-20T04:00:00.000Z",{"name":41,"slug":42,"count":43,"latest_published_at":39},"Security","security",435,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",210,"2026-08-19T09:32:27.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",140,"2026-08-19T18:25:42.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Consumer Tech","consumer-tech",95,"2026-08-18T16:05:00.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Science","science",90,"2026-08-19T18:41:02.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Software","software",73,"2026-08-18T07:51:50.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":78},"Dev Tools","dev-tools",69,"2026-08-18T04:00:00.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Startups","startups",47,"2026-08-19T19:13:46.000Z",{"name":85,"slug":86,"count":87,"latest_published_at":88},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"General","general",33,"2026-08-18T22:18:13.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]