Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Social Learning: Towards Collaborative Learning with Large Language Models

Amirkeivan Mohtashami, Florian Hartmann, Sian Gooding, Lukas Zilka, Matt Sharifi, Blaise Aguera y Arcas

TL;DR

This work formalizes privacy-aware social learning for LLM-driven agents, enabling knowledge transfer between models without sharing private data. It proposes two mechanisms—Verbal Social Learning (instruction sharing) and Live models (synthetic example sharing)—and evaluates them across diverse tasks, using memorization-based privacy metrics to quantify data leakage. The results show that both approaches can achieve competitive performance compared with using original labels and prompts, with the Secret Sharer adaptation indicating only modest memorization. The paper also explores extensions (teaching to larger models, voting aggregators) and outlines future directions for improving teaching strategies, generalizing to other modalities, and strengthening privacy guarantees.

Abstract

We introduce the framework of "social learning" in the context of large language models (LLMs), whereby models share knowledge with each other in a privacy-aware manner using natural language. We present and evaluate two approaches for knowledge transfer between LLMs. In the first scenario, we allow the model to generate abstract prompts aiming to teach the task. In our second approach, models transfer knowledge by generating synthetic examples. We evaluate these methods across diverse datasets and quantify memorization as a proxy for privacy loss. These techniques inspired by social learning yield promising results with low memorization of the original data. In particular, we show that performance using these methods is comparable to results with the use of original labels and prompts. Our work demonstrates the viability of social learning for LLMs, establishes baseline approaches and highlights several unexplored areas for future work.

Social Learning: Towards Collaborative Learning with Large Language Models

TL;DR

This work formalizes privacy-aware social learning for LLM-driven agents, enabling knowledge transfer between models without sharing private data. It proposes two mechanisms—Verbal Social Learning (instruction sharing) and Live models (synthetic example sharing)—and evaluates them across diverse tasks, using memorization-based privacy metrics to quantify data leakage. The results show that both approaches can achieve competitive performance compared with using original labels and prompts, with the Secret Sharer adaptation indicating only modest memorization. The paper also explores extensions (teaching to larger models, voting aggregators) and outlines future directions for improving teaching strategies, generalizing to other modalities, and strengthening privacy guarantees.

Abstract

We introduce the framework of "social learning" in the context of large language models (LLMs), whereby models share knowledge with each other in a privacy-aware manner using natural language. We present and evaluate two approaches for knowledge transfer between LLMs. In the first scenario, we allow the model to generate abstract prompts aiming to teach the task. In our second approach, models transfer knowledge by generating synthetic examples. We evaluate these methods across diverse datasets and quantify memorization as a proxy for privacy loss. These techniques inspired by social learning yield promising results with low memorization of the original data. In particular, we show that performance using these methods is comparable to results with the use of original labels and prompts. Our work demonstrates the viability of social learning for LLMs, establishes baseline approaches and highlights several unexplored areas for future work.
Paper Structure (35 sections, 4 figures, 14 tables)

This paper contains 35 sections, 4 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Problem Setting & Methods
  3. Social Learning Protocol
  4. Methods
  5. Verbal Social Learning: Sharing Instructions
  6. Live models: Sharing Examples
  7. Related Work
  8. LLMs and agents
  9. Federated Learning
  10. Experiments
  11. Live models: Sharing Examples
  12. Extensions to sharing examples
  13. Teaching to a larger student
  14. Voting aggregator
  15. Verbal Social Learning: Sharing Instructions
  16. ...and 20 more sections

Figures (4)

  • Figure 1: An illustration of our social learning framework. Teachers have access to private data that they cannot directly share. The student does not have access to such data. Instead it relies on the teachers to create instructions or non-private examples to teach it the task. After receiving these instructions, the student aggregates them into a single prompt. This prompt is used by the student at inference time to respond to a user's queries.
  • Figure 2: The prompt used to generate instructions for a task.
  • Figure 3: Example reconstruction likelihood is the score the model assigns to a generated example (in blue) which follows the original examples. The score is only computed on the generated example.
  • Figure 4: Manually added prefix instruction to specify GSM8K format. No instruction to perform CoT is given.