[ICML2026] Observation and action design in reinforcement learning using LLMs

LG AI Research presented a study at ICML 2026 in Seoul in which large language models directly design the observation and action spaces for reinforcement learning, work it says came from questions raised in real industrial settings.

• The observation space sets what information an agent receives and the action space sets the scope and units of its decisions, and the post says poorly matched designs can make an agent need more time and data to learn an effective policy.
• The method, called LOAM, generates observation and action functions from task, robot and state descriptions, and the post shows an abbreviated observation model the LLM produced for the HumanoidBench running task.
• A variant called LOAM-Race predicts the potential of candidates and allocates learning resources accordingly, with multiple design candidates competing during learning.
• Experiments are shown on the HumanoidBench stand, walk and run tasks and on Isaac Lab.