ArcticQA Benchmark Evaluates LLM Abstention in Scientific Contexts

The ArcticQA and ArcticAbstain datasets test the ability of Gemini, Claude, and GPT models to abstain from answering scientific questions when no valid option exists. Results show significant variability in abstention rates, ranging from 0.0% to 63.0% in answer-present conditions.