PhantomBench: Benchmarking the Non-existential Threat of Language Models

In PhantomBench, Haeji Jung and Hila Gonen of the University of British Columbia and Amii introduce a benchmark for testing whether language models recognize when they have no knowledge. The benchmark contains 62,411 linguistically plausible but non-existent terms and entities generated by blending words, recombining entity n-grams, and filtering exact matches against Dolma v1.7 with Infini-gram. The authors evaluate 21 models, including Llama, Gemma, Qwen, Mistral, Google’s Gemini 2.5, OpenAI’s GPT-OSS-20B, and DeepSeek-R1, using prompts about existence, meaning, date, and place. Hallucination rates reach 86.7%, and presupposing that a concept exists is especially damaging: models hallucinate far more when asked what an imaginary concept means than whether it exists. On the full benchmark, Gemma 3 12B has a 73.3% hallucination rate, while Llama 3.1 8B performs better at 7.3%, showing that model scale and capability do not guarantee calibrated abstention. Reasoning models and the largest variants in some families can hallucinate more, while domain-specialized models such as MedGemma and BioMistral are inconsistent. Most importantly, abstention on non-existent concepts correlates strongly with behavior on rare real concepts, with Pearson correlation 0.755, suggesting PhantomBench is a practical proxy for long-tail knowledge reliability in high-stakes applications.