# FedReason: Large Language Models as Weight-Space Meta-Aggregators for Heterogeneous Federated Learning

DevFeed: [FedReason: Large Language Models as Weight-Space Meta-Aggregators for Heterogeneous Federated Learning](<https://devfeed.tech/articles/fedreason-large-language-models-as-weight-space-meta-aggregators-for-heterogeneous-federated-learning-119932.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.08383>)

Author: Ali Mohammadi Ruzbahani, Abbas Yazdinejad, Hadis Karimipour

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Federated Learning](<https://devfeed.tech/topics/federated-learning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [AI Sandbox](<https://devfeed.tech/topics/ai-sandbox.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Claude](<https://devfeed.tech/topics/claude.md>)

Tags: [aggregation](<https://devfeed.tech/tags/aggregation.md>), [cifar-10](<https://devfeed.tech/tags/cifar-10.md>), [claude-opus-5-5](<https://devfeed.tech/tags/claude-opus-5-5.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [federated-learning](<https://devfeed.tech/tags/federated-learning.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [proxy](<https://devfeed.tech/tags/proxy.md>), [python](<https://devfeed.tech/tags/python.md>), [sandbox](<https://devfeed.tech/tags/sandbox.md>), [server](<https://devfeed.tech/tags/server.md>), [training](<https://devfeed.tech/tags/training.md>)

## AI overview

FedReason uses an LLM to adapt federated-learning aggregation without server-side training: every ten rounds, it inspects printed results from sandboxed weight analysis, can score candidates on a 500-image proxy set, and submits a Python aggregation program. Full weight tensors stay on the server, and invalid programs trigger FedAvg fallback. In 168 runs on CIFAR-10/100 with ResNet-18, FedReason beat FedAvg by an average 1.26 points across 12 paired label-skew runs, and a same-budget standard-rule selection control by 1.08 points. It was the strongest server-side rule on CIFAR-100 but did not surpass SCAFFOLD, which changes client training. With three malicious clients, it excluded sign-flipping clients from round one and finished about one point below FLTrust without a root model. Removing proxy feedback sharply reduced accuracy in one seed; a 60-round run averaged $0.79 in API fees.

## Source excerpt

arXiv:2611.08383v1 Announce Type: new Abstract: Federated learning servers combine client models with a fixed aggregation rule, and each rule in common use is designed for one kind of heterogeneity. We study whether a large language model (LLM) that inspects the client models can choose the rule itself, without any training on the server. In FedReason, the LLM is called every ten rounds. It analyses the weights by running code in a restricted server-side sandbox, sees only the printed results, may score a few candidates on a 500-image proxy set, and submits a short Python aggregation program. The server applies the program until the next call and falls back to FedAvg whenever its output is invalid. Full weight tensors never leave the server. We evaluated FedReason with Claude Opus 5.5 on CIFAR-10 and CIFAR-100 with ResNet-18 under Dirichlet label skew, width-heterogeneous sub-models and Byzantine clients, in 168 runs against 14 baselines. Across 12 paired label-skew runs it beat FedAvg by 1.26 percentage points on average (10 wins), and a control that spends the same per-round proxy budget on choosing among standard rules by 1.08 points. It was the strongest server-side rule on CIFAR-100 but did not beat SCAFFOLD, which changes client training. With three of ten clients malicious, it excluded all sign-flipping clients from the first round and stayed about one point below FLTrust without training a root model. Without proxy feedback, accuracy fell from 73.8% to 38.8% in one seed. A 60-round run cost 0.79 USD in API fees on average.