Core Concepts
Polaris is a safety-focused Large Language Model (LLM) constellation designed for real-time patient-AI healthcare conversations, showcasing advanced medical reasoning and rapport-building capabilities.
Abstract
1. Introduction:
Polaris 2 developed as the first safety-focused LLM constellation for long multi-turn voice conversations in healthcare.
Trained on proprietary data, clinical care plans, and medical reasoning documents.
2. System Overview:
Objectives include use cases, constellation architecture, orchestration, and LLM details.
Multi-agent system with primary and specialist support agents optimized for real-time healthcare conversation.
3. Conversational Alignment:
Data sources include foundation model training data and simulated conversations between nurses and patient actors.
Training stages involve general instruction tuning, conversation tuning, and agent tuning.
4. Specialist Support Agents:
Privacy & Compliance Specialist, Checklist Specialist, Medication Specialist, Labs & Vitals Specialist, Nutrition Specialist, Hospital & Payor Policy Specialist.
5. Evaluation:
Extensive clinician evaluation shows Polaris performs on par with human nurses across various dimensions.
6. Related Work:
Discusses challenges in current healthcare AI applications and the impact of LLMs in healthcare.
7. Future Work:
Focuses on developing non-diagnostic technology for healthcare workforce augmentation through generative AI agents.
Stats
Polarisは、医療分野でのリアルタイム患者-AIヘルスケア会話に特化した安全性重視の大規模言語モデル(LLM)コンステレーションシステムを開発しました。