Social learning: Collaborative learning with large language models

Framework enables LLM-to-LLM knowledge transfer using natural language instead of gradient-based federated learning

Tested on spam detection (SMS), grade school math, and text-based QA tasks

8 synthetic examples generated by teacher model performed comparably to original data on multiple tasks

16 synthetic examples further reduced performance gap relative to original data

Privacy measurement: Student model showed minimal memorization of private canary data points compared to direct data sharing

Uses PaLM 2-S models in experiments

Addresses use cases where data cannot be directly shared due to privacy concerns

Published by Google Research (Mohtashami & Hartmann), March 2024

Framework enables privacy-aware knowledge sharing between LLMs using natural language only

Tested on spam detection, grade school math, and text-based QA tasks

8 synthetic examples generated by teacher model perform comparably to original data on multiple tasks

16 examples further reduce performance gap versus original training data

Privacy measurement using adapted 'Secret Sharer' method shows student models only slightly memorize teacher's private examples

Approach contrasts with federated learning methods that rely on gradient sharing

Uses PaLM 2-S models for teacher-student evaluation

Task-dependent effectiveness: synthetic examples outperform instructions on some tasks, instructions outperform examples on others