The Art of Saying No: Contextual Noncompliance in Language Models

TL;DR

This paper develops a taxonomy and evaluation suite for assessing when language models should not comply with user requests, revealing high compliance in some categories and proposing training strategies to improve noncompliance handling.

Contribution

It introduces a comprehensive taxonomy of contextual noncompliance, a new evaluation suite, and explores training methods like low rank adapters to improve models' noncompliance capabilities.

Findings

Models show high compliance in understudied categories.

GPT-4 incorrectly complies with up to 30% of requests.

Parameter efficient training improves noncompliance without harming capabilities.

Abstract

Chat-based language models are designed to be helpful, yet they should not comply with every user request. While most existing work primarily focuses on refusal of "unsafe" queries, we posit that the scope of noncompliance should be broadened. We introduce a comprehensive taxonomy of contextual noncompliance describing when and how models should not comply with user requests. Our taxonomy spans a wide range of categories including incomplete, unsupported, indeterminate, and humanizing requests (in addition to unsafe requests). To test noncompliance capabilities of language models, we use this taxonomy to develop a new evaluation suite of 1000 noncompliance prompts. We find that most existing models show significantly high compliance rates in certain previously understudied categories with models like GPT-4 incorrectly complying with as many as 30% of requests. To address these gaps, we explore different training strategies using a synthetically-generated training set of requests and expected noncompliant responses. Our experiments demonstrate that while direct finetuning of instruction-tuned models can lead to both over-refusal and a decline in general capabilities, using parameter efficient methods like low rank adapters helps to strike a good balance between appropriate noncompliance and other capabilities.