LLM Picks Aggregation Rule in Federated Learning

🔥 What happened Researchers let Claude Opus 5.5 write the aggregation rule for federated learning every ten rounds. The LLM inspects weights in a sandbox and submits a short Python program that the server runs until the next call. 💡 Why it matters Across 12 label-skew runs, FedReason beat FedAvg by 1.26 percentage points – at an average API cost of $0.79 per 60 rounds. Without proxy feedback, accuracy collapsed from 73.8% to 38.8%: the LLM clearly needs concrete signals to choose well. ⚡ Our take A clever proof of concept, but the reliance on pricey API calls and proxy data makes it fragile for production. Betting on LLMs here just buys you new single points of failure.