Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Plan-Grounded Large Language Models for Dual Goal Conversational Settings

Diogo Glória-Silva, Rafael Ferreira, Diogo Tavares, David Semedo, João Magalhães

TL;DR

The paper tackles dual-goal mixed-initiative conversations where an LLM must ground dialogue in a procedural plan while also handling user instructions. It introduces PlanLLM, a plan-grounded model that can navigate procedural steps, answer plan-grounded questions, handle open requests, and enforce safety norms, trained via multi-objective supervision and Direct Preference Optimization (DPO). A large-scale, real-world augmented dialogue dataset is created to fuel training, combining plan-focused data with context and tone conditioning. Experiments demonstrate that PlanLLM (notably the Vicuna-DPO variant) achieves a 2.1x improvement over strong baselines and generalizes well to unseen domains, such as DIY, while maintaining safety and user satisfaction. The work offers a practical pathway to robust, goal-directed conversational systems that can guide users through complex tasks with guardrails and adaptable behavior.

Abstract

Training Large Language Models (LLMs) to follow user instructions has been shown to supply the LLM with ample capacity to converse fluently while being aligned with humans. Yet, it is not completely clear how an LLM can lead a plan-grounded conversation in mixed-initiative settings where instructions flow in both directions of the conversation, i.e. both the LLM and the user provide instructions to one another. In this paper, we tackle a dual goal mixed-initiative conversational setting where the LLM not only grounds the conversation on an arbitrary plan but also seeks to satisfy both a procedural plan and user instructions. The LLM is then responsible for guiding the user through the plan and, at the same time, adapting to new circumstances, answering questions, and activating safety guardrails when needed. We propose a novel LLM that grounds the dialogue on a procedural plan, can take the dialogue initiative, and enforces guardrails on the system's behavior, while also improving the LLM's responses to unexpected user behavior. Experiments in controlled settings and with real users show that the best-performing model, which we call PlanLLM, achieves a 2.1x improvement over a strong baseline. Moreover, experiments also show good generalization to unseen domains.

Plan-Grounded Large Language Models for Dual Goal Conversational Settings

TL;DR

The paper tackles dual-goal mixed-initiative conversations where an LLM must ground dialogue in a procedural plan while also handling user instructions. It introduces PlanLLM, a plan-grounded model that can navigate procedural steps, answer plan-grounded questions, handle open requests, and enforce safety norms, trained via multi-objective supervision and Direct Preference Optimization (DPO). A large-scale, real-world augmented dialogue dataset is created to fuel training, combining plan-focused data with context and tone conditioning. Experiments demonstrate that PlanLLM (notably the Vicuna-DPO variant) achieves a 2.1x improvement over strong baselines and generalizes well to unseen domains, such as DIY, while maintaining safety and user satisfaction. The work offers a practical pathway to robust, goal-directed conversational systems that can guide users through complex tasks with guardrails and adaptable behavior.

Abstract

Training Large Language Models (LLMs) to follow user instructions has been shown to supply the LLM with ample capacity to converse fluently while being aligned with humans. Yet, it is not completely clear how an LLM can lead a plan-grounded conversation in mixed-initiative settings where instructions flow in both directions of the conversation, i.e. both the LLM and the user provide instructions to one another. In this paper, we tackle a dual goal mixed-initiative conversational setting where the LLM not only grounds the conversation on an arbitrary plan but also seeks to satisfy both a procedural plan and user instructions. The LLM is then responsible for guiding the user through the plan and, at the same time, adapting to new circumstances, answering questions, and activating safety guardrails when needed. We propose a novel LLM that grounds the dialogue on a procedural plan, can take the dialogue initiative, and enforces guardrails on the system's behavior, while also improving the LLM's responses to unexpected user behavior. Experiments in controlled settings and with real users show that the best-performing model, which we call PlanLLM, achieves a 2.1x improvement over a strong baseline. Moreover, experiments also show good generalization to unseen domains.
Paper Structure (46 sections, 4 equations, 5 figures, 23 tables)

This paper contains 46 sections, 4 equations, 5 figures, 23 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Learning to Reason about Procedural Plans
  4. Procedural Plan
  5. Model Grounding and Dialogue
  6. Multi-objective Plan-Grounded Dialogue
  7. Plan Supervised Fine-Tuning
  8. Plan Preference Optimization
  9. Generating Plan-Grounded Dialogues
  10. Real-World Augmented Dialogue Data
  11. Contextual Dialogue Generation and Preference Data
  12. Evaluation and Discussion
  13. Experimental Setup
  14. Models
  15. Procedural Plans and Dialogues
  16. ...and 31 more sections

Figures (5)

  • Figure 1: An example of a dual goal conversational setting where the user is executing a manual task with the guidance of an LLM assistant.
  • Figure 2: Plan-grounded large language models can dialogue, navigate, and reason about procedural plans. Please refer to the annex for more user-LLM dialogues.
  • Figure 3: Win rate of all trained models against the ground-truth dialogues.
  • Figure 4: The form that the participants of the user study had to fill out at the end of each interaction.
  • Figure 5: The form used for the second user study, pertaining to the overall dialogue quality.