Can LLMs spend a shared reasoning budget wisely?</p>\n<ul>\n<li><strong>R³-Bench</strong>: A benchmark for resource-rational reasoning under shared computational budgets.</li>\n<li><strong>Six-problem contests</strong>: Math, competitive programming, and abstract reasoning tasks compete for one shared token or action budget.</li>\n<li><strong>Competence vs. allocation</strong>: A same-model empirical oracle compares contest performance with each model's demonstrated single-problem capability.</li>\n<li><strong>Unrealized headroom in 71/72 cells</strong>: The oracle strictly outperforms the model's own contest allocation in 71 main-result cells and ties in one.</li>\n<li><strong>Tools are not enough</strong>: Online strategy updates remain limited; lightweight schedulers help in 6/9 strong-pressure model-domain cases, but no scheduler works universally.</li>\n</ul>\n<p>Paper: <a href=\"https://arxiv.org/abs/2608.16033\" rel=\"nofollow\">https://arxiv.org/abs/2608.16033</a><br>Code: <a href=\"https://github.com/NineAbyss/R-3-Bench\" rel=\"nofollow\">https://github.com/NineAbyss/R-3-Bench</a><br>Dataset: <a href=\"https://huggingface.co/datasets/R-3-Bench/R-3-Bench\">https://huggingface.co/datasets/R-3-Bench/R-3-Bench</a></p>\n","updatedAt":"2026-08-18T02:40:01.262Z","author":{"_id":"64be4408c05a0df0d2b6012e","avatarUrl":"/avatars/09d8427505a418090391dc5a3f8bfef2.svg","fullname":"PSWang","name":"CedarWang","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7570841312408447},"editors":["CedarWang"],"editorAvatarUrls":["/avatars/09d8427505a418090391dc5a3f8bfef2.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.16033","authors":[{"_id":"6a83c409675db694db8cd481","name":"Peisong Wang","hidden":false},{"_id":"6a83c409675db694db8cd482","user":{"_id":"68410d555a3b810d2ad6125d","avatarUrl":"/avatars/7aa0cc17a5ceda176e494cc64a35284c.svg","isPro":false,"fullname":"Zhiwei Ma","user":"libliot","type":"user","name":"libliot"},"name":"Zhiwei Ma","status":"claimed_verified","statusLastChangedAt":"2026-08-18T08:45:05.114Z","hidden":false},{"_id":"6a83c409675db694db8cd483","name":"Bowen Liu","hidden":false},{"_id":"6a83c409675db694db8cd484","name":"Feixue Liu","hidden":false},{"_id":"6a83c409675db694db8cd485","name":"Aochuan Chen","hidden":false},{"_id":"6a83c409675db694db8cd486","name":"Chenyi Zi","hidden":false},{"_id":"6a83c409675db694db8cd487","name":"Hongchuan Zeng","hidden":false},{"_id":"6a83c409675db694db8cd488","name":"Yuhan Li","hidden":false},{"_id":"6a83c409675db694db8cd489","name":"Jia Li","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/64be4408c05a0df0d2b6012e/Iv5USxvk51osCZEBrrt_B.png","https://cdn-uploads.huggingface.co/production/uploads/64be4408c05a0df0d2b6012e/iNs-kfhmApn_6-CQjSW7U.jpeg"],"publishedAt":"2026-08-17T00:00:00.000Z","submittedOnDailyAt":"2026-08-18T00:00:00.000Z","title":"R^3-Bench: LLMs Struggle with Resource-Rational Reasoning under Shared Budgets","submittedOnDailyBy":{"_id":"64be4408c05a0df0d2b6012e","avatarUrl":"/avatars/09d8427505a418090391dc5a3f8bfef2.svg","isPro":false,"fullname":"PSWang","user":"CedarWang","type":"user","name":"CedarWang"},"summary":"In cognitive science, resource rationality asks how an agent should allocate limited computation to maximize expected value. Most reasoning and agent benchmarks use independent per-task budgets; existing shared-budget studies do not calibrate suite performance against the same model's demonstrated single-problem competence. We introduce R^3-Bench, which evaluates six-problem suites under shared budgets across mathematics, competitive programming, and abstract reasoning in tool-free and agentic settings. Matched single-problem response curves define an offline empirical oracle over observed successes. Across 72 main-table cells for six models, the oracle mean matches or exceeds the contest mean in all cells and is strictly higher in 71. Under moderate tool-free pressure, equal-allocation replay also exceeds contest performance for four of six models. Trajectory diagnostics reveal limited strategy updating and pressure-dependent failure patterns. In a three-model diagnostic under strong agentic pressure, at least one fixed scheduler exceeds the contest mean in six of nine cells, but no policy dominates across domains. These results expose a persistent gap between demonstrated competence and shared-budget realization.","upvotes":11,"discussionId":"6a83c409675db694db8cd48a","projectPage":"https://github.com/NineAbyss/R-3-Bench","githubRepo":"https://github.com/NineAbyss/R-3-Bench","githubRepoAddedBy":"user","ai_summary":"R³-Bench reveals that shared computation budgets cause reasoning agents to underperform relative to their single-problem capabilities across math, coding, and abstract reasoning tasks.","ai_keywords":["resource rationality","R³-Bench","shared budgets","response curves","empirical oracle","strategy updating","fixed scheduler"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":2,"organization":{"_id":"68909d0ec75ddecb2867fcba","name":"HKUST-DSAIL","fullname":"Data Science & Artificial Intelligence Lab at HKUST","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/657dc533aed79486aeffbf55/biw8iB5cUADFtw0_a6LCM.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6a5f28c9c3daab50c7ff9f76","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a5f28c9c3daab50c7ff9f76/M-c-NBoMcVsxFJcok0LGF.png","isPro":false,"fullname":"R^3Bench","user":"R-3-Bench","type":"user"},{"_id":"68410d555a3b810d2ad6125d","avatarUrl":"/avatars/7aa0cc17a5ceda176e494cc64a35284c.svg","isPro":false,"fullname":"Zhiwei Ma","user":"libliot","type":"user"},{"_id":"67e8f2845c0900f17bfbb5fd","avatarUrl":"/avatars/9c70427411a5bf43c4a68e11664c3f80.svg","isPro":false,"fullname":"Felicia","user":"ffeeiik","type":"user"},{"_id":"68676ddbb1bd4bd32016d63b","avatarUrl":"/avatars/f9648f8390dc73292ced1a60d5f6d100.svg","isPro":false,"fullname":"RLVER","user":"RLVER","type":"user"},{"_id":"67b80da55e1b74491fc08eb9","avatarUrl":"/avatars/991baa07ce8792b2a04f30e27aabf786.svg","isPro":false,"fullname":"S2R","user":"S2R-data","type":"user"},{"_id":"68d6591428e169473e936a31","avatarUrl":"/avatars/e926e835b6c6f3989c81faa233da55fc.svg","isPro":false,"fullname":"NSPO","user":"ICLR2026NSPO","type":"user"},{"_id":"6a83c7344b4c3155571e91d0","avatarUrl":"/avatars/9a5bca3e984a95e62a13eae3eb36b3a5.svg","isPro":false,"fullname":"ANG","user":"CdricW","type":"user"},{"_id":"64e408524b78ab05968a8493","avatarUrl":"/avatars/6d987ca56eb9c2aa23bf2216f198b991.svg","isPro":false,"fullname":"chenyi","user":"BarristanZi","type":"user"},{"_id":"66bac1248ba60a7a91ba0836","avatarUrl":"/avatars/a92e6fd8b3104842f34f7ceaf5ceaa5f.svg","isPro":false,"fullname":"Wang Yuyao","user":"littlewyy","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":"5f32b2367e583543386214d9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1635314457124-5f32b2367e583543386214d9.jpeg","isPro":true,"fullname":"Sergei Averkiev","user":"averoo","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"68909d0ec75ddecb2867fcba","name":"HKUST-DSAIL","fullname":"Data Science & Artificial Intelligence Lab at HKUST","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/657dc533aed79486aeffbf55/biw8iB5cUADFtw0_a6LCM.png"},"query":{}}">

Cite arxiv.org/abs/2608.16033 in a model README.md to link it from this page.

Cite arxiv.org/abs/2608.16033 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.