Inverse-RLignment: Large Language Model Alignment from Demonstrations through Inverse Reinforcement Learning

TL;DR

This paper introduces a novel demonstration-based alignment method for large language models using inverse reinforcement learning, addressing issues of noisy labels and high annotation costs, and demonstrating strong empirical results.

Contribution

It formalizes Alignment from Demonstrations within a reinforcement learning framework, proposing divergence minimization objectives and an efficient algorithm for improved LLM alignment.

Findings

Effective alignment on Harmless and Helpful tasks

Addresses noisy labels and privacy concerns in LLM alignment

Demonstrates strong empirical performance with a simple approach

Abstract

Aligning Large Language Models (LLMs) is crucial for enhancing their safety and utility. However, existing methods, primarily based on preference datasets, face challenges such as noisy labels, high annotation costs, and privacy concerns. In this work, we introduce Alignment from Demonstrations (AfD), a novel approach leveraging high-quality demonstration data to overcome these challenges. We formalize AfD within a sequential decision-making framework, highlighting its unique challenge of missing reward signals. Drawing insights from forward and inverse reinforcement learning, we introduce divergence minimization objectives for AfD. Analytically, we elucidate the mass-covering and mode-seeking behaviors of various approaches, explaining when and why certain methods are superior. Practically, we propose a computationally efficient algorithm that extrapolates over a tailored reward model for AfD. We validate our key insights through experiments on the Harmless and Helpful tasks, demonstrating their strong empirical performance while maintaining simplicity.