Concordia: Self-Improving Synthetic Tables for Federated LLMs

2026-05-10

AI Summary

This work addresses the challenge of adapting large language models in federated learning under strict data isolation and non-IID conditions, particularly when only synthetic tabular data is available and neither raw nor validation data can be shared. To this end, the authors propose Concordia, a framework that enables the first end-to-end co-optimization of synthetic data generation and federated learning objectives. Clients fine-tune models via LoRA on synthetic data and train lightweight utility scorers using private validation feedback to reweight samples. Simultaneously, a shared set of heterogeneous scorers across clients drives self-improvement of local synthetic data generators through Group Relative Policy Optimization (GRPO). Evaluated on privacy-sensitive tabular tasks in finance and healthcare, Concordia consistently outperforms static or decoupled baselines, achieving notable gains in federated performance, cross-client stability, and robustness to distribution shifts.

Abstract

Federated learning (FL) enables training large language models (LLMs) without sharing raw data, but adapting LLMs under strict data isolation and non-IID client distributions remains challenging in practice. Synthetic data offers a natural privacy-preserving surrogate for local training, yet existing federated pipelines typically treat synthetic generation as static or loosely coupled with downstream optimization, leading to rapidly diminishing utility under heterogeneous clients. We study federated adaptation of LLMs on tabular tasks where raw records and validation data cannot be shared, and local training must rely entirely on synthetic tables. We propose Concordia, a tri-level optimization framework that aligns synthetic data generation with federated validation utility despite these constraints. At the client level, models are adapted via parameter-efficient LoRA training on synthetic tables. Clients additionally learn lightweight utility scorers from private validation feedback to reweight synthetic samples during local training. At the outer level, each client refines its own synthetic table generator using group-relative policy optimization (GRPO), guided by an ensemble of heterogeneous scorers shared across clients, without aggregating generator parameters or exposing validation data. Experiments on privacy-sensitive tabular benchmarks from finance and healthcare demonstrate that Concordia consistently improves federated performance, cross-client stability, and robustness to distribution shift compared to static and decoupled synthetic-data baselines.

Problem

• federated learning
• large language models
• synthetic data
• data isolation
• non-IID

Innovation

• federated learning
• synthetic data generation
• tri-level optimization
• group-relative policy optimization
• parameter-efficient fine-tuning