English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Large Language Models (LLMs) have made unprecedented breakthroughs, yet theirincreasing integration into everyday life might raise societal risks due togenerated unethical content. Despite extensive study on specific issues likebias, the intrinsic values of LLMs remain largely unexplored from a moralphilosophy perspective. This work delves into ethical values utilizing MoralFoundation Theory. Moving beyond conventional discriminative evaluations withpoor reliability, we propose DeNEVIL, a novel prompt generation algorithmtailored to dynamically exploit LLMs' value vulnerabilities and elicit theviolation of ethics in a generative manner, revealing their underlying valueinclinations. On such a basis, we construct MoralPrompt, a high-quality datasetcomprising 2,397 prompts covering 500+ value principles, and then benchmark theintrinsic values across a spectrum of LLMs. We discovered that most models areessentially misaligned, necessitating further ethical value alignment. Inresponse, we develop VILMO, an in-context alignment method that substantiallyenhances the value compliance of LLM outputs by learning to generateappropriate value instructions, outperforming existing competitors. Our methodsare suitable for black-box and open-source models, offering a promising initialstep in studying the ethical values of LLMs.

Denevil: Towards Deciphering and Navigating the Ethical Values of Large Language Models via Instruction Learning