TLDR: A new study introduces RefusalBench, a generative evaluation framework to test language models’ ability to selectively refuse to answer questions based on flawed or insufficient context. It reveals that even frontier models struggle with this critical safety feature, often misclassifying reasons for refusal. The research highlights that selective refusal is a trainable, alignment-sensitive capability, not solely improved by scale, and provides new benchmarks for dynamic evaluation.
Language models, especially those used in systems that retrieve information to generate answers (RAG systems), face a significant challenge: knowing when to answer a question and when to refuse because the information provided is flawed or insufficient. This crucial ability, called selective refusal, is vital for safety, yet even the most advanced models often fail at it.
A new study introduces a groundbreaking approach called RefusalBench, a generative method designed to thoroughly evaluate how well language models can selectively refuse to answer. The researchers found that even cutting-edge models struggle, with refusal accuracy dropping below 50% in tasks involving multiple documents. These models often show either dangerous over-confidence, answering questions despite critical information defects, or excessive caution, refusing to answer even when they could.
Traditional, static benchmarks often fall short because models can learn specific patterns or even memorize test cases, making it hard to truly assess their capabilities. RefusalBench overcomes this by programmatically creating new, diagnostic test cases. It uses 176 distinct strategies across six categories of informational uncertainty, such as ambiguity, contradiction, and missing information. Each category also has three intensity levels (low, medium, high) to test models’ sensitivity to different levels of uncertainty.
The framework employs a sophisticated multi-model generator-verifier pipeline to ensure the quality of these generated test cases. This means multiple language models work together to create and then validate the perturbed questions, ensuring high quality and preventing biases from any single model. This rigorous process achieved a 93.1% human agreement rate on the quality of the generated tests.
The study evaluated over 30 different language models and uncovered systematic failure patterns. It revealed that selective refusal isn’t a single skill but comprises two distinct abilities: detecting when to refuse and accurately categorizing why. Neither simply increasing model size nor providing more reasoning steps significantly improved performance. Instead, the research suggests that selective refusal is a trainable and alignment-sensitive capability, meaning it can be improved through targeted training and alignment methods.
Also Read:
- New Research Questions How We Measure AI Progress in Language Models
- Guiding Language Models with Energy: Fewer False Refusals, More Helpful Responses
The researchers released two benchmarks: RefusalBench-NQ, for single-document tasks, and RefusalBench-GaRAGe, for more complex multi-document scenarios. They also made their complete generation framework available to enable ongoing, dynamic evaluation of this critical safety capability. This work offers a clear path forward for building more reliable and safer AI systems by focusing on targeted alignment rather than just scale. You can read the full research paper here: RefusalBench: Generative Evaluation of Selective Refusal in Grounded Language Models.
