Reproduce “Frog Soup: Zero-Shot, In-Context, and Sample-Efficient Frogger Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiang Li, Yiyang Hao, Doug Fulop
Stanford University researchers demonstrated that reasoning-focused large language models (LLMs) can play the Atari game Frogger without explicit game-specific training. They showed that LLMs achieve scores comparable to trained reinforcement learning agents in a zero-shot setting and can improve the sample efficiency of Deep Q-Networks by up to 40% through generated demonstrations.
Key insights
- Reasoning LLMs can transfer general reasoning and planning abilities to play novel game environments like Frogger without explicit game-specific training.
- LLMs can utilize in-context learning from past game states and rewards to adapt and refine their decision-making strategies.
- LLM-generated demonstrations can significantly improve the sample efficiency and convergence of traditional reinforcement learning algorithms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.