Thanos: Enhancing Conversational Agents with Skill-of-Mind-Infused Large Language Model (2411.04496v1)

Abstract: To increase social bonding with interlocutors, humans naturally acquire the ability to respond appropriately in a given situation by considering which conversational skill is most suitable for the response - a process we call skill-of-mind. For LLM-based conversational agents, planning appropriate conversational skills, as humans do, is challenging due to the complexity of social dialogue, especially in interactive scenarios. To address this, we propose a skill-of-mind-annotated conversation dataset, named Multifaceted Skill-of-Mind, which includes multi-turn and multifaceted conversational skills across various interactive scenarios (e.g., long-term, counseling, task-oriented), grounded in diverse social contexts (e.g., demographics, persona, rules of thumb). This dataset consists of roughly 100K conversations. Using this dataset, we introduce a new family of skill-of-mind-infused LLMs, named Thanos, with model sizes of 1B, 3B, and 8B parameters. With extensive experiments, these models successfully demonstrate the skill-of-mind process and exhibit strong generalizability in inferring multifaceted skills across a variety of domains. Moreover, we show that Thanos significantly enhances the quality of responses generated by LLM-based conversational agents and promotes prosocial behavior in human evaluations.

Enhancing Conversational Agents with Skill-of-Mind-Infused LLM

The concept of infusing conversational agents with what the authors term "Skill-of-Mind" represents a novel approach to addressing the challenges faced by LLM-based models in social dialogue. The paper suggests that human-like conversational adaptability can be integrated into LLMs to improve their interactive and social capabilities. The significant contribution of this research is introducing the Multifaceted Skill-of-Mind dataset, which serves as a foundation for developing skill-of-mind-infused LLMs.

Multifaceted Skill-of-Mind Dataset and Annotation Process

The dataset underpins the research and is drawn from twelve diverse dialogue datasets. It encompasses approximately 100,000 conversations that span multiple interactive scenarios such as task-oriented dialogues, long-term interactions, and counseling sessions. The key value of this dataset lies in its granularity and diversity, annotated with multifaceted conversational skills and explanations derived through a method termed perspective-taking. The dataset emphasizes the importance of social dynamics, including demographics and memory-based content, primarily designed to aid LLMs in better contextualizing conversations.

Development of Skill-of-Mind-Infused LLMs

The research introduces a novel family of models, instantiated in sizes of 1B, 3B, and 8B parameters, specifically trained on the Multifaceted Skill-of-Mind dataset. These models, through extensive experimentation, demonstrate the ability to reason about and infer appropriate conversational skills, aligning responses with social contexts more effectively than traditional LLMs. This process emulates the human ability to reflect and contextualize dialogue intricacies to enhance conversation quality.

Strong Numerical Results and Generalizability

The models exhibit robust performance not only within the curated dataset but also when tested on out-of-domain scenarios. They show higher skill classification accuracy and better alignment with conversational needs compared to existing LLMs, as evidenced by evaluations on datasets like BlendedSkillTalk and ProsocialDialogue. These models notably excel in prosocial behavior detection, indicating safety improvements and ethical interaction alignment.

Theoretical and Practical Implications

This work's implications extend across multiple domains. Theoretically, it advances the concept of embedding social reasoning abilities into LLMs, suggesting paths for integrating cognitive models in AI development. Practically, it paves the way for more sophisticated conversational agents capable of nuanced, context-aware interactions. Such systems are particularly useful in applications needing high engagement and personalized interaction strategies, including virtual assistants and mental health support bots.

Future Prospects

While the research establishes a promising framework, future work could focus on embedding the Skill-of-Mind capability intrinsically within conversational agents rather than relying on input prompts. Expanding the generalizability across more varied scenarios and honing the balance between complex reasoning and social interaction in dialogue agents could enhance effectiveness further.

In summary, the paper by Lee et al. represents a significant step towards creating more adept and socially aware conversational agents that emulate human-like cognitive planning and responsiveness. As these frameworks gain broader application, they have the potential to significantly enhance AI-driven communication across sectors.