Teaching AI to pick social skills before responding makes conversations more natural.
Thanos model family (1B, 3B, 8B parameters) helps chatbots choose the right conversational skills like humans do
https://arxiv.org/abs/2411.04496
🤖 Original Problem:
Current LLMs struggle with social commonsense reasoning and strategic communication in interactive scenarios. They face challenges in personalizing responses and handling the one-to-many problem where multiple plausible responses exist.
-----
🛠️ Solution in this Paper:
→ Introduces Multifaceted Skill-of-Mind Dataset with 100K conversations annotated with explanations and conversational skills
→ Creates hierarchical taxonomy of conversational skills in 5 categories: Interpersonal, Memory & Knowledge, Cognitive & Problem-Solving, Communication & Listening, and Task-Oriented
→ Develops Thanos model family (1B, 3B, 8B parameters) trained on this dataset to predict appropriate conversational skills and generate contextual explanations
→ Uses skill-of-mind as guidance to narrow down response options and improve inference speed
-----
🎯 Key Insights:
→ Humans naturally choose conversational skills based on social context - this process can be modeled
→ Providing skill guidance helps reduce the one-to-many problem in dialogue generation
→ Single skill-expert agent is more efficient than multiple specialized agents
→ Skill-of-mind promotes prosocial behavior through ethical skill choices
-----
📊 Results:
→ Achieved high human evaluation scores: 3.72 for relevance, 3.75 for plausibility, 3.74 for understanding
→ Outperformed baseline models in skill classification and response generation tasks
→ Demonstrated strong generalizability across various dialogue scenarios
→ Showed significant improvements in response quality when used as augmented input prompts
Share this post