Cooperative multi‑agent reinforcement learning (MARL) enables teams of robots, autonomous vehicles and other agents to acquire coordinated behaviors, but the learned policies are heavily shaped by the reward function. Because task success alone offers sparse guidance, intermediate rewards are usually handcrafted, and any change in environment or team forces a redesign.
Large language models (LLMs) can generate reward functions as code, yet these rewards are typically validated in a single setting and lack reusability. To address this, we introduce GARDEN (Generalist Autonomous Reward Design through Evidence Navigation), which builds a single reward program for a family of related scenarios and transfers it unchanged to unseen environments.
During development, an LLM acts as a reward engineer: it inspects environment code, training logs, and agent behaviors, proposes draft rewards, and evaluates them in trial training runs. The learning algorithm and the rule for selecting the final reward remain fixed, while the evidence‑navigation loop iteratively refines the reward based on observed evidence.
We evaluate GARDEN on unseen scenarios of the Overcooked and SMAX benchmarks. On two kitchen layouts, task returns increase by 58.2% and 52.5%, respectively; in two combat scenarios, the mean win rate rises from 51.0% to 89.9%. With matched candidate and probe budgets, GARDEN outperforms fixed diagnostic feedback (84.2%) and Eureka‑inspired reward reflection (79.2%).
These results demonstrate that a single autonomously engineered reward, refined through evidence navigation, can be effectively reused across related cooperative tasks, dramatically reducing the manual effort traditionally required for reward design.
Review: GARDEN showcases the power of LLM‑driven reward synthesis combined with systematic evidence navigation, offering a promising path toward scalable, reusable reward engineering in MARL.