{"paper":{"title":"Eureka: Human-Level Reward Design via Coding Large Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Large language models can design reward functions for robot tasks that outperform those created by human experts.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.RO","authors_text":"Anima Anandkumar, De-An Huang, Dinesh Jayaraman, Guanzhi Wang, Linxi Fan, Osbert Bastani, William Liang, Yecheng Jason Ma, Yuke Zhu","submitted_at":"2023-10-19T17:31:01Z","abstract_excerpt":"Large Language Models (LLMs) have excelled as high-level semantic planners for sequential decision-making tasks. However, harnessing them to learn complex low-level manipulation tasks, such as dexterous pen spinning, remains an open problem. We bridge this fundamental gap and present Eureka, a human-level reward design algorithm powered by LLMs. Eureka exploits the remarkable zero-shot generation, code-writing, and in-context improvement capabilities of state-of-the-art LLMs, such as GPT-4, to perform evolutionary optimization over reward code. The resulting rewards can then be used to acquire"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Without any task-specific prompting or pre-defined reward templates, Eureka generates reward functions that outperform expert human-engineered rewards. In a diverse suite of 29 open-source RL environments that include 10 distinct robot morphologies, Eureka outperforms human experts on 83% of the tasks, leading to an average normalized improvement of 52%.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That LLM-generated reward code will continue to produce stable, non-degenerate policies when transferred to new tasks or real hardware, rather than exploiting simulator quirks that do not appear in the reported 29 environments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Eureka uses LLMs for evolutionary optimization of reward code to outperform human experts on 83% of 29 RL tasks with 52% average improvement and enables gradient-free RLHF.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large language models can design reward functions for robot tasks that outperform those created by human experts.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"400320205c025bb9e78d1d801867ebf301ca5aefe735e9adec57308f006a3542"},"source":{"id":"2310.12931","kind":"arxiv","version":2},"verdict":{"id":"521bdc28-8cd4-4e95-be37-5921a4011a8c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:10:45.963363Z","strongest_claim":"Without any task-specific prompting or pre-defined reward templates, Eureka generates reward functions that outperform expert human-engineered rewards. In a diverse suite of 29 open-source RL environments that include 10 distinct robot morphologies, Eureka outperforms human experts on 83% of the tasks, leading to an average normalized improvement of 52%.","one_line_summary":"Eureka uses LLMs for evolutionary optimization of reward code to outperform human experts on 83% of 29 RL tasks with 52% average improvement and enables gradient-free RLHF.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That LLM-generated reward code will continue to produce stable, non-degenerate policies when transferred to new tasks or real hardware, rather than exploiting simulator quirks that do not appear in the reported 29 environments.","pith_extraction_headline":"Large language models can design reward functions for robot tasks that outperform those created by human experts."},"references":{"count":14,"sample":[{"doi":"","year":null,"title":"If you see phrases like [NUM: default_value], replace the entire phrase with a numerical value","work_id":"e1404de5-7f99-4b50-b730-b8dd755a9245","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"If you see phrases like {CHOICE: choice1, choice2, ...}, it means you should replace the entire phrase with one of the choices listed","work_id":"08d69275-c02c-44ab-ac69-f6ba9199ff18","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"If you see [optional], it means you only add that line if necessary for the task, otherwise remove that line","work_id":"cd6a8140-9399-4f50-b78d-ffda67ed8f68","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Do not invent new objects not listed here","work_id":"9aa3520b-1406-4dc3-b23d-d4453816fbc5","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Always start the description with [start of plan] and end it with [end of plan]","work_id":"76f847c9-fc23-4dda-927a-262b3d2c5909","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":14,"snapshot_sha256":"e0f3eb979b31a037bc6b19dae15be2ca86d18273a46c2bba621f617a7a72af06","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"f69cc1d530b44f7709403e4ba609fffa2a572d54bca81a04ba0d6f8219c438f8"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}