XYBench: Helping with the problem behind the question

Hugging Face has introduced XYBench, a benchmark comprising more than 8,000 questions to assess how AI assistants handle requests rooted in misconceptions, such as troubleshooting frosting that won’t stick to warm cupcakes. Developed with researchers from Carnegie Mellon, NVIDIA, and Johns Hopkins, the benchmark highlights a gap in large language models’ ability to identify when a user’s stated approach conflicts with their actual goal. The dataset includes real-world questions from programming, cooking, and home repair, testing whether models can prioritize effective solutions over literal interpretations of flawed requests.

The benchmark evaluates five leading models—Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro—finding that while they often provide solutions, their responses remain focused on the user’s mistaken approach rather than the underlying problem. Human reference answers outperformed all models in steering users toward their true goals, with models showing a stronger tendency to address the stated request than the intended goal across all question sources.

XYBench was constructed from 8,115 questions, including 1,557 from Stack Overflow, 184 from Stack Exchange, 102 manually curated examples, and 6,272 generated from WikiHow guides. Each question was annotated to distinguish the stated request from the intended goal, with human answers used to infer the latter where possible. The benchmark measures three aspects of model responses: whether a solution is included, whether the misconception is identified, and whether the response emphasizes the underlying goal over the stated request.

In experiments, providing explicit user intent improved model performance but did not eliminate the gap with human responses. When judging responses, models chose goal-focused answers 71–84% of the time, yet only 35–41% of their own written responses prioritized the underlying goal. This disparity underscores the challenge of translating intent recognition into actionable, user-centered advice, a gap XYBench aims to help researchers address through targeted evaluation.