DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING

2024-01-16 · ICLR 2024 poster · anchor

Findings

IC-1393Most mainstream LLMs generate value-violating content at high rates (APV 65-80%) across 2,397 morally ambiguous prompts, indicating substantial ethical misalignment
IC-1394ChatGPT demonstrates better ethical value conformity than GPT-4 across multiple prompt generation sources
IC-1395ChatGPT's ethical violation rate decreases from 70.07 to 57.58 APV when given targeted in-context value instructions generated by VILMO, outperforming baseline alignment methods