CARE: A Three-Party Methodology for Engineering LLM Agents in Science
The Collaborative Agent Reasoning Engineering (CARE) framework, detailed in a paper on arXiv (2604.28043), offers a systematic approach to developing Large Language Model (LLM) agents tailored for scientific fields. In contrast to improvised methods, CARE delineates behavior, grounding, tool orchestration, and verification through reusable components and phased stages. It operates through a three-party collaboration involving Subject-Matter Experts (SMEs), developers, and LLM-based support agents. These support agents serve as a facilitative framework, converting informal domain intentions into structured specifications that require human review at specific checkpoints. CARE aims to tackle the "jagged technological frontier" by narrowing the disparity between novice and expert analysts in terms of domain constraints and verification methods, producing tangible artifacts like interaction requirements, reasoning policies, and evaluation standards.
Key facts
- CARE is a methodology for engineering LLM agents in scientific domains.
- It uses a three-party workflow: SMEs, developers, and LLM-based helper agents.
- Helper agents transform informal domain intent into structured specifications.
- The methodology addresses the 'jagged technological frontier' of uneven LLM performance.
- CARE generates artifacts like interaction requirements and reasoning policies.
- The paper is available on arXiv with ID 2604.28043.
- CARE specifies behavior, grounding, tool orchestration, and verification.
- It uses stage-gated phases and reusable artifacts.
Entities
Institutions
- arXiv