FrozenLake

FrozenLake is a benchmark environment used to evaluate reinforcement learning algorithms, consisting of a grid where agents navigate to reach a goal while avoiding holes, often characterized by stochastic elements and sparse rewards.

Papers using FrozenLake (2)

Process Reward Informed Tree Rollout for Effective Multi-Turn RL2026TSR: Trajectory-Search Rollouts for Multi-Turn RL of LLM Agents2026