Reproduce “ELIGN: Expectation Alignment as a Multi-Agent Intrinsic Reward” 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
Zixian Ma, Rose Wang, Li Fei-Fei, Michael Bernstein, Ranjay Krishna
ELIGN introduces an intrinsic reward for multi-agent systems, fostering coordination and competitive behaviors under decentralized training and partial observability by incentivizing agents to align with or deviate from neighbors' expectations. This approach consistently outperforms curiosity-based baselines across various tasks, demonstrating improved scalability and zero-shot generalization.
Key insights
- Complex collective intelligence and synchronized behaviors in multi-agent systems can emerge from simple, local interactions and predictions among neighbors, drawing inspiration from biological self-organization principles.
- Shifting intrinsic motivation from novelty-seeking (curiosity) to promoting predictability or unpredictability among agents can effectively drive coordination and competitive strategies.
- The accuracy of the learned dynamics model is critical for the effectiveness of expectation-alignment-based intrinsic rewards, as model degradation directly impacts task performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.