Eureka: Human-Level Reward Design via Coding Large Language Models

By Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, D. Jayaraman, Yuke Zhu, Linxi Fan, Anima Anandkumar. International Conference on Learning Representations, 2023

673 citations recorded by the source corpus.

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 complex skills via reinforcement learning. 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%. The generality of Eureka also enables a new gradient-free in-context learning approach to reinforcement learning from human feedback (RLHF), readily incorporating human inputs to improve the quality and the safety of the generated rewards without model updating. Finally, using Eureka rewards in a curriculum learning setting, we demonstrate for the first time, a simulated Shadow Hand capable of performing pen spinning tricks, adeptly manipulating a pen in circles at rapid speed.

Subfields: Artificial Intelligence, Information Systems, Materials Chemistry

View Paper: https://www.semanticscholar.org/paper/6ca16c1c2c60ceda87242c8f8e522d12cc4a13bc

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Zipeng Fu
2. Hao Peng
3. Heng Ji
4. Moo Jin Kim
5. Qingqing Zhao
6. Gordon Wetzstein
7. Xingyao Wang
8. Yangyi Chen

In Information Systems

1. Xingyao Wang
2. Heng Ji
3. Hao Peng
4. Chelsea Finn

In Materials Chemistry

1. Heng Ji
2. Chelsea Finn

Citations in the corpus

• Evolution of Heuristics: Towards Efficient Automatic Algorithm Design Using Large Language Model (2024, neutral): “Evolutionary search has been adopted in both code generation (Liventsev et al., 2023; Ma et al., 2023; Lehman et al., 2024; Hemberg et al., 2024) and text generation (Guo et al., 2023b; Fernando et al., 2023; Xu et al., 2023a).”
• ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution (2024, neutral): “…powerful methods for prompt optimization (Zhou et al., 2022; Wang et al., 2023d; Guo et al., 2023), reinforcement learning (RL) reward design (Ma et al., 2023b), algorithmic (self-)improvement (Zelikman et al., 2023; Liu et al., 2023c;a), neural architecture search (Chen et al., 2023a), and…”
• Executable Code Actions Elicit Better LLM Agents (2024, background): “…LLMs to acquire capabilities such as tool invocation and memory management (Mialon et al., 2023; Schick et al., 2023) and venture into real-world tasks such as controlling robots (Ahn et al., 2022; Huang et al., 2023; Ma et al., 2023) and performing scientific experiments (Bran et al., 2023).”
• CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (2025, positive): “…works utilize VLMs as intermediate components for perception and control, leveraging their strong semantic understanding and reasoning capabilities to decompose complex tasks [18, 23, 31, 50], detect objects [16, 22], or generate dense rewards [11, 37, 68] or goals [2, 9, 13, 41, 42, 49, 65, 74].”
• Large Language Model (LLM) for Telecommunications: A Comprehensive Survey on Principles, Key Techniques, and Opportunities (2024, methodology): “[36] proposed a solution by generating multiple solutions in each iteration and then applying evolutionary strategies to select better strategies.”
• A Survey on Vision--Language--Action Models for Embodied AI (2024, neutral): “Additionally, VLMs can function as high-level task planners if they possess sufficient reasoning capabilities.”