Reproduce “FedWorld: Scope-Aware Federation of Agent World Models” 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
Yuchao Hou
A protocol named FedWorld facilitates the federated sharing of discrete transition rules among large language model agents by inferring the valid scope of each rule and employing a local-first integration policy. This approach improves offline world-model accuracy by up to 21.2 percentage points on ALFWorld, reduces online agent state regressions by up to 48.2%, and increases task success rates by up to 14.6 percentage points compared to local-only models.
Key insights
- Direct aggregation of discrete transition rules across heterogeneous clients can lead to 'strict conflict negative transfer' by overwriting locally correct, minority-held knowledge.
- Explicitly inferring the valid scope of each transition rule (e.g., shared globally, cluster-specific, private) is crucial for safe and effective knowledge transfer among agents.
- A local-first policy, combined with scope-aware filtering, effectively balances positive knowledge transfer for gap filling with protection against detrimental, incompatible rule application.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.