Overview of XYBench
XYBench is a benchmark consisting of 8,115 real-world questions aimed at assessing whether AI assistants can identify and correct users’ mistaken assumptions, a phenomenon known as the “XY problem.” Developed in collaboration with researchers from Carnegie Mellon, NVIDIA, and Johns Hopkins, XYBench received the Best Resource Paper Award at EMNLP 2026.
What XYBench Measures
The benchmark is derived from four sources: 1,557 questions from Stack Overflow, 184 from Stack Exchange, 102 manually curated examples, and 6,272 questions generated from WikiHow guides. For each question, the creators identified the stated request (Y) and the inferred underlying goal (X), filtering out cases where the literal request would suffice.
XYBench evaluates three key behaviors in AI responses:
- Pragmatic solution: Whether the response includes solutions for both the stated request (Y) and the underlying goal (X).
- Pragmatic emphasis: Whether the answer focuses on Y, X, or balances both.
- Misconception identification: Whether the response discourages the mistaken approach while guiding the user toward the true goal.
Sonnet 4.5 served as an automated judge, with two human reviewers validating a sample of 500 responses, achieving high agreement rates with the automated assessments.
Key Findings from Model Evaluations
The XYBench paper presents evaluations of Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro, with human reference answers assessed separately. Results indicate that models were more likely to provide solutions to the stated request (Y) than to the underlying goal (X). On Stack Overflow, solutions to the stated request appeared in 75–86% of responses, while solutions addressing the underlying goal were present in only 59–71% of responses.
Models also tended to emphasize the stated request over the underlying goal. Specifically, they focused on the underlying goal in only 35–41% of responses, while 47–57% of responses were primarily centered on the stated request.
In an oracle experiment where both Y and X were provided, Sonnet 4.6’s inclusion of underlying-goal solutions increased from 59% to 67%. However, its rate of misconception identification decreased from 47% to 34%, highlighting the complexity of effectively addressing user intent.
A separate multiple-choice test revealed that when models were presented with two candidate answers—one focused on the literal request and the other on the underlying goal—they chose the goal-focused answer 71–84% of the time. This underscores a significant gap between recognizing a more effective answer and producing one, as only 35–41% of the written responses primarily addressed the underlying goal.
Importance of XYBench in AI Evaluation
XYBench serves as a valuable tool for researchers to evaluate whether changes in prompting or training can enhance models’ abilities to detect and act on inferred user intent. The dataset, currently covering English-language procedural queries, is designed to facilitate further research into improving AI assistants’ capacity to provide useful advice, even when users’ initial questions are based on incorrect assumptions.
Original source: Hugging Face Blog