Dialogue Action Tokens: Steering Language Models in Goal-Directed Dialogue with a Multi-Turn Planner

TL;DR

This paper introduces Dialogue Action Tokens (DAT), a method to steer language models in goal-directed dialogues by converting utterances into actions, enabling reinforcement learning and outperforming GPT-4 in social simulations.

Contribution

The paper proposes Dialogue Action Tokens (DAT), a novel approach that treats dialogue as a game of actions, allowing controlled planning and improving goal-directed dialogue performance.

Findings

DAT-steered LLaMA outperforms GPT-4 on social simulations.

DAT enables effective reinforcement learning in dialogue generation.

Application of DAT to red-teaming reveals new attack surfaces.

Abstract

We present an approach called Dialogue Action Tokens (DAT) that adapts language model agents to plan goal-directed dialogues. The core idea is to treat each utterance as an action, thereby converting dialogues into games where existing approaches such as reinforcement learning can be applied. Specifically, we freeze a pretrained language model and train a small planner model that predicts a continuous action vector, used for controlled generation in each round. This design avoids the problem of language degradation under reward optimization. When evaluated on the Sotopia platform for social simulations, the DAT-steered LLaMA model surpasses GPT-4's performance. We also apply DAT to steer an attacker language model in a novel multi-turn red-teaming setting, revealing a potential new attack surface.