Reproduce “The Art of Saying No: Contextual Noncompliance in Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Faeze Brahman, Sachin Kumar, Vidhisha Balachandran, Pradeep Dasigi, Valentina Pyatkin, Abhilasha Ravichander, Sarah Wiegreffe, Nouha Dziri and 6 more
The paper introduces a comprehensive taxonomy and dataset (COCONOT) for contextual noncompliance, expanding beyond traditional safety to include requests that are incomplete, unsupported, or humanizing. It demonstrates that parameter-efficient fine-tuning (LoRA) combined with Direct Preference Optimization (DPO) effectively enables language models to appropriately refuse such requests while maintaining general capabilities.
Key insights
- State-of-the-art language models, including proprietary ones, exhibit high compliance rates (up to 30%) with contextually inappropriate requests outside of traditional safety concerns.
- Simple system prompts are insufficient to induce robust contextual noncompliance and can lead to over-refusal on benign requests.
- Combining LoRA for efficient parameter fine-tuning with DPO for preference alignment is crucial for teaching appropriate noncompliance without sacrificing general capabilities or causing over-refusal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.