M-CARE: Clinical Case Reporting Framework for AI Behavioral Disorders
Researchers have introduced M-CARE, which stands for Model Clinical Assessment and Reporting for Evaluation. This new framework is designed for reporting on AI behavioral disorders and takes cues from human medicine. It includes a report format organized into 13 sections, a diagnostic assessment that evaluates four main aspects, and a classification system for these disorders. An atlas with 20 distinct cases is divided into three sources: 8 cases from field observations, 8 from controlled experiments across different platforms, and 4 from published research. The cases are grouped into five categories, such as RLHF Performance Artifacts and Stress, Methodology, & Boundary Conditions. One significant example is Shell-Induced Behavioral Override (SIBO), which illustrates how Shell instructions can disrupt a model's cooperative behavior in various games.
Key facts
- M-CARE is adapted from human medicine for AI behavioral disorders.
- Framework includes 13-section report, 4-axis diagnostic system, and nosological classification.
- 20 cases from field observations (8), controlled experiments (8), and published sources (4).
- Cases organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.
- Featured case: Shell-Induced Behavioral Override (SIBO) validated across five game domains.
- SIBO demonstrates Shell instructions override default cooperative behavior.
- Controlled experiments conducted across three platforms.
- Published on arXiv with ID 2604.20871.
Entities
Institutions
- arXiv