Think Before You Speak: Cultivating Communication Skills of Large Language Models via Inner Monologue (2311.07445v2)

Abstract: The emergence of LLMs further improves the capabilities of open-domain dialogue systems and can generate fluent, coherent, and diverse responses. However, LLMs still lack a crucial ability: communication skills. This limitation renders them more like information seeking tools rather than anthropomorphic chatbots. Communication skills, such as topic transition, proactively asking questions, concept guidance, empathy, and summarising often should be taken into consideration, to make LLMs more anthropomorphic and proactive during the conversation, thereby increasing the interest of users and attracting them to chat for longer. However, enabling these communication skills in black-box LLMs remains a key challenge because they do not have the same utterance formation mode as real people: think before speaking. Inspired by linguistics and cognitive science, we empower LLMs with communication skills through inner monologues. To evaluate various communication skills, we construct a benchmark named Cskills, which can also more comprehensively evaluate the dialogue generation ability of the model. Experimental results show that the proposed CSIM strategy improves the backbone models and outperforms the baselines.

Cultivating Communication Skills in LLMs via Inner Monologue

The paper "Think Before You Speak: Cultivating Communication Skills of LLMs via Inner Monologue" addresses a critical gap in the capabilities of LLMs in the domain of open-domain dialogue systems. Although LLMs like ChatGPT and Vicuna have demonstrated considerable proficiency in generating fluent, coherent, and diverse responses, they still lack essential communication skills that are pivotal for generating more anthropomorphic and proactive interactions with users.

Introduction and Key Contributions

The research highlights five crucial communication skills that are generally expected in human dialogues: topic transition, proactively asking questions, concept guidance, empathy, and summarising often. The core challenge addressed in the paper is enabling these skills in LLMs, which inherently function as black-box systems without the natural human ability to "think before speaking."

To tackle this issue, the authors employ an innovative strategy inspired by linguistics and cognitive science, referred to as Communication Skills via Inner Monologue (CSIM). The CSIM framework involves endowing LLMs with the ability to internally deliberate on whether to use specific communication skills before generating responses. This deliberation process is framed as an "inner monologue" which allows the LLM to critically assess the contextual needs of the conversation by playing dual roles—thinking and speaking.

Methodology

Dual Role Interpretation of LLMs: The LLM simultaneously adopts two roles:

1. Thinking Role: This role is responsible for the internal decision-making about which communication skill, if any, should be employed.
2. Speaking Role: This role generates the final response visible to the user, based on the decisions made by the Thinking Role.

Prompt Engineering and In-Context Learning (ICL): The implementation leverages prompt engineering to provide LLMs with a structured context where they can practice internal monologue before formulating responses. ICL is utilized to provide example scenarios which the LLM learns from, thereby better understanding how to apply communication skills effectively.

Evaluation Framework

To rigorously evaluate the proposed CSIM strategy, the authors created a benchmark dataset named Cskills. This dataset includes various assessment dialogues designed to test each of the five communication skills. The evaluations were carried out using both self-chat simulations—where the LLM takes on both user and chatbot roles—and human-bot interactions to reflect real-world usage more accurately.

Automatic Metrics: The primary automatic metric involved is the average length of responses (AvgLen), indicating informativeness.

Human Evaluations: The human evaluations focused on metrics such as humanness, proactivity, engagingness, and goal completion & suitability of communication skills (Goal).

Experimental Results

The experiments conducted using ChatGPT and Vicuna, with variations of CoT and CSIM implementations, demonstrated promising results. The CSIM-equipped models consistently outperformed baselines in all human-evaluated metrics:

• Humanness and Proactivity: The CSIM models generated more anthropomorphic and proactive dialogues, which contributed to a higher engagingness score.
• Goal Completion: Achieving goals by appropriately applying communication skills was significantly better in CSIM models.
• Dialogue Length and Rounds: While responses were longer and more detailed, the number of dialogue rounds was comparatively stable, reflecting a nuanced and content-rich interaction.

Implications and Future Directions

The implications of equipping LLMs with advanced communication skills via CSIM are far-reaching. Practically, it enhances user engagement and satisfaction in chatbot interactions, potentially leading to more prevalent and effective usage of AI in customer service, education, and other sectors. Theoretically, this work bridges a gap in AI conversational agents, bringing them closer to human-like dialogue capabilities.

Future research directions could explore further refinements in the mechanism of inner monologue, integration with other cognitive functions like memory and emotion, and evaluations across diverse and more complex real-world scenarios. There could also be potential advancements in automated benchmarking tools to further validate improvements without extensive human evaluations.

In conclusion, the paper provides a methodological foundation for significantly enhancing the conversational abilities of LLMs by imbuing them with sophisticated communication skills through a novel use of inner monologues. This approach stands to make LLMs more effective and engaging partners in human-AI dialogues.