Researchers have developed a new way to use large language models (LLMs) to improve federated learning, a method that allows many devices to collaboratively train an AI model without sharing their private data. This approach, called FedReason, uses an LLM to intelligently decide how to combine the different models trained on each device. The innovation matters because federated learning often struggles when devices have very different data or hardware, and picking the right way to merge their models is tricky. FedReason shows a promising step toward smarter, more flexible AI training that respects privacy and adapts to diverse conditions.
Key Takeaways
- FedReason leverages a large language model to dynamically select how to aggregate client models during federated learning, rather than relying on a fixed rule.
- It operates by analyzing model weights and testing candidate aggregation methods on a small proxy dataset, then generating a custom Python program for combining models.
- In experiments on image recognition tasks with challenging data and device differences, FedReason outperformed standard aggregation methods like FedAvg by over 1 percentage point in accuracy.
- FedReason also demonstrated resilience against malicious clients trying to disrupt training, excluding harmful updates early without needing extra trusted models.
Federated learning is a way to train AI models across many devices—like smartphones or sensors—without moving their private data to a central server. Each device trains a local model on its own data, then sends updates to a server that combines them into a global model. However, devices often have very different amounts or types of data, or even different model sizes, making it hard to merge their updates effectively. Traditionally, servers use a fixed “aggregation rule” such as averaging the weights (FedAvg), but these rules are designed for specific scenarios and don’t adapt well to diverse conditions.
The researchers behind FedReason asked: can a large language model, like those used in natural language processing, help the server choose the best way to combine client models on the fly? Unlike typical training, the LLM doesn’t learn from data but instead “reasons” about the model weights by running code in a secure environment on the server. Every ten training rounds, FedReason calls the LLM, which examines summarized information about the client models, tests a few candidate aggregation strategies on a small proxy dataset, and then outputs a short Python program that defines how to combine the models for the next rounds.
This process keeps the full model weights secure on the server and only shares limited information with the LLM. If the LLM’s program is invalid or fails, the system defaults back to the standard FedAvg method to ensure stability. The researchers tested FedReason using the Claude Opus 5.5 language model on standard image classification benchmarks (CIFAR-10 and CIFAR-100) with ResNet-18 neural networks. They introduced challenging conditions such as uneven label distributions (label skew), varying model sizes, and malicious clients sending harmful updates.
The results were encouraging. FedReason consistently outperformed FedAvg and other baseline methods, improving accuracy by about 1.26 percentage points on average in scenarios with uneven data distributions. It was also the strongest server-side aggregation approach on CIFAR-100, though it did not surpass methods that change how clients train their models (like SCAFFOLD). Notably, FedReason quickly identified and excluded malicious clients attempting to poison the training without needing a trusted “root” model, maintaining robust performance. The approach relies on a small proxy dataset to evaluate aggregation strategies, and without this feedback, accuracy dropped significantly, highlighting the importance of this testing step.
While FedReason incurs some cost—about $0.79 per 60 training rounds in API fees for the language model—the flexibility and improved accuracy it offers could be valuable for real-world federated learning applications. This includes scenarios like mobile device personalization, healthcare data analysis, or any setting where privacy and diverse data are key concerns. Future work may explore using other language models, reducing reliance on proxy data, or integrating this approach with client-side training improvements. By combining AI reasoning with privacy-preserving training, FedReason points toward more adaptive and reliable federated learning systems.
Based on research published on arXiv by Ali Mohammadi Ruzbahani, Abbas Yazdinejad, Hadis Karimipour.