[{"data":1,"prerenderedAt":204},["ShallowReactive",2],{"project-44535078":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":33,"aspects":34,"strip":40,"post_text":42,"what":43,"said":70,"maker_in_thread":127,"shape":137,"grid":140,"attempts":166,"same_week":167,"maker_shows":41,"maker_hidden":196,"reader_version":197,"card":198},"44535078","RULER","Easily apply RL to any agent","title","Show HN: RULER – Easily apply RL to any agent","https:\u002F\u002Fopenpipe.ai\u002Fblog\u002Fruler","openpipe.ai","kcorbitt",1752256056,2025,81,11,882,3837,23,"notable","ai","reinforcement learning","library",[21,24,25,26,27],"reward function","grpo","llm judge","agent training","curious",0.333,{"positive":31,"neutral":31,"mixed":32,"critical":32},3,0,6,{"idea":35,"docs":37,"design":38,"performance":39},{"positive":36,"critical":32},1,{"positive":36,"critical":32},{"positive":32,"critical":32},{"positive":32,"critical":32},[41,32,31,32,32,31,32,41,32,31,41],5,"Hey HN, Kyle here, one of the co-founders of OpenPipe.\n\nReinforcement learning is one of the best techniques for making agents more reliable, and has been widely adopted by frontier labs. However, adoption in the outside community has been slow because it's so hard to implement.\n\nOne of the biggest challenges when adapting RL to a new task is the need for a task-specific \"reward function\" (way of measuring success). This is often difficult to define, and requires either high-quality labeled data and\u002For significant domain expertise to generate.\n\nRULER is a drop-in reward function that works acr",{"problem":44,"audience":50,"pricing":55,"motivation":56,"time_to_build":55,"stack":60,"openness":69,"stage":69,"team_size":69},{"value":45,"source":46,"evidence":47},"Defining task-specific reward functions for reinforcement learning is hard and needs labeled data or domain expertise","maker",[48],{"id":49},"post",{"value":51,"source":52,"evidence":53},"Teams adapting reinforcement learning to agents for new tasks","inferred",[54],{"id":49},null,{"value":57,"source":46,"evidence":58},"Make reinforcement learning easier to adopt outside large labs",[59],{"id":49},[61,65],{"value":62,"source":46,"evidence":63},"GRPO",[64],{"id":49},{"value":66,"source":46,"evidence":67},"Qwen 2.5 models",[68],{"id":49},"unknown",{"praise":71,"objections":90,"requests":109,"questions":116,"alternatives":121},[72,83],{"value":73,"source":74,"evidence":75},"Interesting, more general than prior RLPR-style approaches","commenter",[76,80],{"id":77,"by":78,"text":79},"44536791","sadiq","Excellent, look forward to giving this a go. I was looking at: https:\u002F\u002Farxiv.org\u002Fabs\u002F2506.18254 but your approach is even more general.",{"id":81,"by":11,"text":82},"44537930","I really like RLPR for when you have a known-good answer to compare to as well!",{"value":84,"source":74,"evidence":85},"Writeups are appreciated",[86],{"id":87,"by":88,"text":89},"44536454","someoneontenet","Love these write ups!",[91,102],{"value":92,"source":74,"evidence":93},"Name collides with NVIDIA's RULER long-context benchmark",[94,98],{"id":95,"by":96,"text":97},"44537103","spmurrayzzz","Might end up being some confusion with the RULER benchmark from NVIDIA given the (somewhat shared) domain: https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FRULER EDIT: by shared I only mean the…",{"id":99,"by":100,"text":101},"44539398","swyx","yeah unforutnately for you this is one of the well known long context benchmarks. too late tho, soldier on.",{"value":103,"source":74,"evidence":104},"Possible ordering bias in LLM judge ranking",[105],{"id":106,"by":107,"text":108},"44537726","maxrmk","Very cool. Do you do anything to mitigate ordering bias in the evaluation function, or do you just expect it to average out over time?",[110],{"value":111,"source":74,"evidence":112},"Explain why o3 underperforms Qwen on the customer support agent task",[113],{"id":114,"by":100,"text":115},"44539411","how does o3 on the customer support agent task so dreadfully underperform qwen?",[117],{"value":118,"source":74,"evidence":119},"Whether ordering bias is mitigated in the judge",[120],{"id":106,"by":107,"text":108},[122],{"value":123,"source":74,"evidence":124},"RLPR",[125,126],{"id":77,"by":78,"text":79},{"id":81,"by":11,"text":82},{"comments":128,"answered":36,"key":129},4,[130,131,134],{"id":81,"text":82},{"id":132,"text":133},"44537934","Dang, hadn't seen that. Namespace collision strikes again.",{"id":135,"text":136},"44537925","No, we don't do anything. Theoretically we could judge several times with different ordering. We could measure order bias really easily though; we just need to look at the…",{"commenters":138,"max_depth":31,"maker_replies":128,"first_comment_mins":139,"lifespan_hours":15},7,155,[141,143,145,147,149,150,153,155,157,160,162],[77,32,41,142,78,32],"project\u002Fpositive\u002Fidea",[81,32,32,144,11,36],"commenter\u002Fneutral\u002Fnone",[95,36,31,146,96,32],"project\u002Fneutral\u002Fdocs",[132,36,32,148,11,36],"maker\u002Fneutral\u002Fnone",[99,36,32,144,100,32],[106,151,31,152,107,32],2,"project\u002Fneutral\u002Fdesign",[135,151,32,154,11,36],"maker\u002Fneutral\u002Fdesign",[87,31,41,156,88,32],"maker\u002Fpositive\u002Fdocs",[158,31,32,159,11,36],"44536788","maker\u002Fpositive\u002Fnone",[114,128,31,161,100,32],"project\u002Fneutral\u002Fperformance",[163,41,41,164,165,32],"44537230","project\u002Fpositive\u002Fnone","ndgold",[],[168,173,178,183,188,192],{"id":169,"name":170,"points":171,"category":172},"44522683","nodaysoff.run",958,"creative",{"id":174,"name":175,"points":176,"category":177},"44572377","Shoggoth Mini",610,"hardware",{"id":179,"name":180,"points":181,"category":182},"44526015","Pangolin",500,"infrastructure",{"id":184,"name":185,"points":186,"category":187},"44556684","Refine",408,"productivity",{"id":189,"name":190,"points":191,"category":187},"44523409","BrowserOS",291,{"id":193,"name":194,"points":195,"category":20},"44524544","Cactus",231,false,"v2",{"id":199,"name":5,"line":6,"ai":196,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":200,"top":18,"kw":201,"maker":11,"domain":10,"tier":19,"attempts":32,"strip":202,"url":9,"mix":203,"praise":73,"objection":92,"sub":21},44535078,0.33,[21,24,25,26,27],"50300305035",[31,31,32,32],1794036678320]