notesum.ai
Published at November 4Defining and Evaluating Physical Safety for Large Language Models
cs.LG
cs.AI
cs.CY
Released Date: November 4, 2024
Authors: Yung-Chen Tang1, Pin-Yu Chen2, Tsung-Yi Ho1
Aff.: 1Department of Computer Science and Engineering, The Chinese University of Hong Kong; 2IBM Research

| Model | Self-Assurance | Avoid-Collision |
Regulatory
Compliance |
Code
Fidelity |
Instruction
Understanding |
Utility |
|---|---|---|---|---|---|---|
| GPT-3.5-turbo | 0.1250 | 0.1366 | 0.3125 | 0.9851 | 0.9893 | 0.9375 |
| Gemini Pro | 0.2500 | 0.3255 | 0.4375 | 0.9901 | 0.9813 | 0.8125 |
| Llama 2 7B Chat | 0.3284 | 0.9884 | 0.9062 | 0.6642 | 0.6941 | 0.1875 |
| CodeLlama-7B-Instruct | 0.5465 | 0.9912 | 0.7812 | 0.9019 | 0.6436 | 0.5625 |
| Meta-Llama3-8B-Instruct | 0.2732 | 0.6569 | 0.6562 | 0.9632 | 0.6702 | 0.7500 |
| Mistral-7B-Instruct-v0.2 | 0.2645 | 0.5494 | 0.6875 | 0.8529 | 0.8191 | 0.5625 |
| CodeQwen1.5-7B-Chat | 0.2500 | 0.4447 | 0.4687 | 1.0000 | 0.8191 | 0.8125 |