Abstract
In the rapidly evolving field of natural language processing, dialogue systems primarily employ a single-step dialogue paradigm. Although this paradigm is efficient, it lacks the depth and fluidity of human interactions and does not appear natural. We introduce a novel Step-by-Step Dialogue Paradigm (Stephanie), designed to mimic the ongoing dynamic nature of human conversations. By employing a dual learning strategy and a further-split post-editing method, we generated and utilized a high-quality step-by-step dialogue dataset to fine-tune existing large language models, enabling them to perform step-by-step dialogues. We thoroughly present Stephanie. Tailored automatic and human evaluations are conducted to assess its effectiveness compared to the traditional single-step dialogue paradigm. We will release code, Stephanie datasets, and Stephanie LLMs to facilitate the future of chatbot eras.
Introduction
Current Large Language Models (LLMs) typically operate under a "Single-Step Dialogue Paradigm." When a user sends a message, the AI responds with a single, often lengthy, and perfectly structured paragraph. While this is effective for information retrieval, it feels artificial in social contexts. Human conversation, particularly in instant messaging, is characterized by "burstiness"—sending multiple shorter messages or fragments to convey thoughts, greetings, and follow-up questions.
The lack of fluidity in standard AI responses leads to lower engagement in social or emotional support scenarios. Stephanie addresses this by training models to break their responses into multiple coherent "bubbles," mimicking the natural flow of human interaction.
Methodology: The Stephanie Paradigm
To transition LLMs from monolithic responses to a step-by-step flow, the researchers implemented three core strategies:
- Dual Learning Strategy: This approach uses contrastive learning by providing the model with both "Positive Learning Objectives" (natural step-by-step examples) and "Negative Learning Objectives" (standard single-step examples). This helps the model distinguish between a "wall of text" and a fluid conversation.
- Further-Split Post-Editing: Initial model outputs were often still too dense. A "further-split" method was developed to manually restructure these outputs, ensuring they follow a natural emotional and logical progression.
- Plug-and-Play Finetuning: Special delimiters were introduced into the training data to represent message breaks. This allows models like Llama-3 and Phi-3 to output multiple messages within a single turn without requiring architectural changes to messaging UIs.
Dataset and Evaluation
The Stephanie dataset was built upon PERSONACHAT, a collection of multi-turn dialogues grounded in specific character traits. The researchers curated 5,457 high-quality dialogue segments, transforming them into the step-by-step format using their dual-learning and further-split methods.
Evaluation focused on several key metrics, including Average Consecutive Message Counts (ACMC). While standard AI has an ACMC near 1.0, Stephanie models were trained to increase this "burstiness" to better reflect human behavior.
Results and Discussion
The Stephanie paradigm was tested across several state-of-the-art models, including GPT-4 and Llama3-70b. Key findings include:
- Enhanced Engagement: Stephanie-Llama3-70b achieved an engagement score of 95.86 (on a 100-point scale), significantly outperforming the standard single-step baseline of 78.74.
- Naturalness: Human evaluators consistently preferred Stephanie's output, citing higher naturalness and interestingness compared to traditional paradigms.
- Conversational Dynamics: Stephanie successfully reduced the average words per message from approximately 11.77 to 5.87, while more than doubling the ACMC from 1.07 to 2.51.
- Lexical Diversity: Analysis using the "Distinct-N" metric confirmed that breaking responses into smaller steps does not sacrifice linguistic variety; in many cases, it maintains or improves it.
Conclusion
Stephanie represents a significant shift in how we design social AI. By moving away from the "assistant" bias of single-step responses and embracing the fragmented, dynamic nature of human chat, Stephanie creates a more engaging and natural user experience. The release of the Stephanie datasets and finetuned models aims to pave the way for a new generation of chatbots that feel less like search engines and more like conversational partners.
Summary updated