Hugging Face Daily Papers · August 25, 2026 · 4 min read

ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction</p>\n","updatedAt":"2026-08-25T03:44:44.916Z","author":{"_id":"6789c65f2b4189d8cec17c98","avatarUrl":"/avatars/a9040d5b81748dd628ee2eb2a5a0d083.svg","fullname":"Mistletoe","name":"mis2","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7276180982589722},"editors":["mis2"],"editorAvatarUrls":["/avatars/a9040d5b81748dd628ee2eb2a5a0d083.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.13622","authors":[{"_id":"6a83fb48675db694db8cd648","user":{"_id":"64c1856bc3b5e0a9760e37f2","avatarUrl":"/avatars/a67c3f95f83b2be3a14f71c6faf3b995.svg","isPro":false,"fullname":"Yongqi Tong","user":"Yooki","type":"user","name":"Yooki"},"name":"Yongqi Tong","status":"claimed_verified","statusLastChangedAt":"2026-08-25T08:15:26.537Z","hidden":false},{"_id":"6a83fb48675db694db8cd649","name":"Tan Li Hui Faith","hidden":false},{"_id":"6a83fb48675db694db8cd64a","name":"Choy Zhen Wen Marcus","hidden":false},{"_id":"6a83fb48675db694db8cd64b","name":"Zhou Jin","hidden":false},{"_id":"6a83fb48675db694db8cd64c","name":"Kewei Fu","hidden":false},{"_id":"6a83fb48675db694db8cd64d","name":"Jiang-Ming Yang","hidden":false},{"_id":"6a83fb48675db694db8cd64e","name":"Jianshe Li","hidden":false},{"_id":"6a83fb48675db694db8cd64f","name":"Xin Zhang","hidden":false}],"publishedAt":"2026-08-13T00:00:00.000Z","submittedOnDailyAt":"2026-08-25T00:00:00.000Z","title":"ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction","submittedOnDailyBy":{"_id":"6789c65f2b4189d8cec17c98","avatarUrl":"/avatars/a9040d5b81748dd628ee2eb2a5a0d083.svg","isPro":true,"fullname":"Mistletoe","user":"mis2","type":"user","name":"mis2"},"summary":"Open-ended real-world interaction admits multiple valid behaviors: an agent may answer directly, ask for clarification, provide progress updates, or confirm before acting. This flexibility breaks a core assumption behind group-based RL: rollouts compared within a group are no longer guaranteed to be behaviorally comparable. As a result, reward-model preferences over interaction style can distort relative advantages and steer optimization toward reward-preferred behaviors rather than context-appropriate ones. We formalize this as a reward fairness problem and propose ARC (Advantage Regularization via Conditioning), a training recipe that restores fairer relative comparison through strategy-conditioned rollout grouping, together with hybrid rewards and entropy regularization. We study ARC in our proposed \\inter, a novel paradigm for responsive, steerable, and execution-aware user-agent interaction that decouples user-visible communication from latent reasoning and tool use. \\inter\\ also provides the annotation and distillation pipeline for constructing \\inter-86K, our strategy-annotated training corpus for supervised and RL training. Empirically, ARC substantially strengthens the core τ/τ^2 tool-use benchmarks, while \\inter\\ reduces time-to-first-token from 4.91s to 1.27s relative to a think-style baseline. Together, these results suggest that a central bottleneck in open-ended interactive learning is not only how agents are rewarded, but whether their behaviors are compared fairly in the first place. The ARC implementation and \\inter-86K training data will be released.","upvotes":12,"discussionId":"6a83fb49675db694db8cd650","ai_summary":"ARC improves fairness in group-based reinforcement learning for open-ended agents by conditioning rollout comparisons on strategy, enabling more context-appropriate behavior in responsive user-agent interaction.","ai_keywords":["group-based RL","reward fairness","ARC","Advantage Regularization via Conditioning","strategy-conditioned rollout grouping","hybrid rewards","entropy regularization","tool-use benchmarks","time-to-first-token"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"68d4ae9941614425f8aa7490","name":"ant-intl","fullname":"ant-international","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68e9c86d2551cb38000fa77c/qG2KcieHqYFuWXi2-weuP.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6789c65f2b4189d8cec17c98","avatarUrl":"/avatars/a9040d5b81748dd628ee2eb2a5a0d083.svg","isPro":true,"fullname":"Mistletoe","user":"mis2","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"65647e2b50a80d26dbfdf49c","avatarUrl":"/avatars/aff0de9f9e4ed322e05d7f832c3c060d.svg","isPro":false,"fullname":"Xu Zhihao","user":"naiweizi","type":"user"},{"_id":"64c1856bc3b5e0a9760e37f2","avatarUrl":"/avatars/a67c3f95f83b2be3a14f71c6faf3b995.svg","isPro":false,"fullname":"Yongqi Tong","user":"Yooki","type":"user"},{"_id":"65dd9e349f9df8e0620be4bf","avatarUrl":"/avatars/b8f6c2fb91635f9633325bf29492a166.svg","isPro":false,"fullname":"jianshe LI","user":"jsl1212","type":"user"},{"_id":"69fda02b5094601dfeebe3bc","avatarUrl":"/avatars/7a97c954fa1c95269efd5b99681991e2.svg","isPro":false,"fullname":"Pan Wang","user":"wangpan-ustc","type":"user"},{"_id":"661e7839252631c8ba54932e","avatarUrl":"/avatars/903eda3f9931b7f85a7312d618785521.svg","isPro":false,"fullname":"Marcus Choy","user":"koolkiz","type":"user"},{"_id":"647db0031a1fcad2fdbfc698","avatarUrl":"/avatars/58c5446ed3e58236bb2454854d852b7c.svg","isPro":false,"fullname":"Tan Li Hui, Faith","user":"Amesery","type":"user"},{"_id":"6590a65b89f1ff0463828e53","avatarUrl":"/avatars/4ab424eede2fe9c114252b1e5dd1ba25.svg","isPro":false,"fullname":"sizhe","user":"sizhe04","type":"user"},{"_id":"67f369add1dfddad5d2f19a5","avatarUrl":"/avatars/2bc629a4ac65273d726f918f107d0037.svg","isPro":false,"fullname":"Hang Wang","user":"Xiaobizaiz1","type":"user"},{"_id":"6474e1afb68461d5cf7c41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6474e1afb68461d5cf7c41cc/bcoiD_qPrjHUBlB259djg.png","isPro":false,"fullname":"Dawei Li","user":"wjldw","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"68d4ae9941614425f8aa7490","name":"ant-intl","fullname":"ant-international","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68e9c86d2551cb38000fa77c/qG2KcieHqYFuWXi2-weuP.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.13622.md","query":{}}">

Papers

arxiv:2608.13622

ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction

Published on Aug 13

· Submitted by

Mistletoe on Aug 25

ant-international

Upvote

Authors:

Yongqi Tong ,

Abstract

ARC improves fairness in group-based reinforcement learning for open-ended agents by conditioning rollout comparisons on strategy, enabling more context-appropriate behavior in responsive user-agent interaction.

Generated by thinkingmachines/Inkling-Small

Open-ended real-world interaction admits multiple valid behaviors: an agent may answer directly, ask for clarification, provide progress updates, or confirm before acting. This flexibility breaks a core assumption behind group-based RL: rollouts compared within a group are no longer guaranteed to be behaviorally comparable. As a result, reward-model preferences over interaction style can distort relative advantages and steer optimization toward reward-preferred behaviors rather than context-appropriate ones. We formalize this as a reward fairness problem and propose ARC (Advantage Regularization via Conditioning), a training recipe that restores fairer relative comparison through strategy-conditioned rollout grouping, together with hybrid rewards and entropy regularization. We study ARC in our proposed \inter, a novel paradigm for responsive, steerable, and execution-aware user-agent interaction that decouples user-visible communication from latent reasoning and tool use. \inter\ also provides the annotation and distillation pipeline for constructing \inter-86K, our strategy-annotated training corpus for supervised and RL training. Empirically, ARC substantially strengthens the core τ/τ^2 tool-use benchmarks, while \inter\ reduces time-to-first-token from 4.91s to 1.27s relative to a think-style baseline. Together, these results suggest that a central bottleneck in open-ended interactive learning is not only how agents are rewarded, but whether their behaviors are compared fairly in the first place. The ARC implementation and \inter-86K training data will be released.