Embedding-Aligned Language Models

TL;DR

This paper introduces EAGLE, a reinforcement learning-based method for aligning large language models with objectives in a latent embedding space, enhancing controlled and domain-specific text generation.

Contribution

It presents a novel RL approach that guides LLMs in adhering to embedding space objectives, improving content relevance and domain alignment.

Findings

EAGLE effectively surfaces content gaps in user demand datasets.

Optimal design of state-dependent actions improves efficiency.

Demonstrates controlled text generation aligned with domain knowledge.

Abstract

We propose a novel approach for training large language models (LLMs) to adhere to objectives defined within a latent embedding space. Our method leverages reinforcement learning (RL), treating a pre-trained LLM as an environment. Our embedding-aligned guided language (EAGLE) agent is trained to iteratively steer the LLM's generation towards optimal regions of the latent embedding space, w.r.t. some predefined criterion. We demonstrate the effectiveness of the EAGLE agent using the MovieLens 25M and Amazon Review datasets to surface content gaps that satisfy latent user demand. We also demonstrate the benefit of using an optimal design of a state-dependent action set to improve EAGLE's efficiency. Our work paves the way for controlled and grounded text generation using LLMs, ensuring consistency with domain-specific knowledge and data representations.