Eigen RadarAI
Analysis

XYBench tests whether AI answers address the user’s actual goal

Ai2 has released XYBench, an evaluation that separates a user’s requested method from their intended goal. Across practical questions, models often included a useful alternative without making it the focus of the response. The English-language benchmark measures solution presence, emphasis and recognition of mistaken assumptions separately, using forum questions and generated examples.

Artificial Intelligence··Night
Unfrosted cupcakes cooling on wire racks on a bright kitchen counter, with an untouched bowl of thick frosting waiting behind them, echoing XYBench's example: cooling the cake, not thicker frosting, solves the problem.

A requested method can miss the goal

Ai2, an artificial intelligence research institute, announced XYBench on November 13 to examine questions built on mistaken assumptions. The evaluation separates the method a person requests from the goal they want to achieve. Its cupcake example asks for thicker frosting when cooling the cake addresses the underlying problem.[1], [2]

Ai2 describes 8,115 questions, including forum material from Stack Overflow and Stack Exchange, manually curated examples and questions generated from WikiHow guides. Highly rated and accepted forum answers helped identify intended goals. The research paper predates the announcement: it was published as a preprint on September 6.[1]

A useful answer does not always lead

Five models generated answers in the main evaluation. On Stack Overflow questions, solutions to the underlying goal appeared in 59–71 per cent of responses, but only 35–41 per cent focused primarily on that goal. Solution presence, emphasis and recognition of the mistaken assumption are scored separately. A separate multiple-choice exercise used six models, which selected goal-focused answers in 71–84 per cent of cases.[1], [2]

Human reviewers checked 500 responses

Two human reviewers assessed 500 responses alongside the automated Sonnet 4.5 judge. Reported agreement was highest for whether a solution was present and lower for answer emphasis. These agreement figures describe the grading process. The dataset covers English, mainly practical tasks, and a single response to each question; the WikiHow questions are synthetic. Long conversations and performance in other languages were outside this evaluation.[1]

References

  1. News sourceAi2 / Hugging FaceAi2 releases XYBench to test responses to mistaken assumptions↩1↩2↩3↩4
  2. News sourceCASRAIXYBench evaluates how assistants handle mistaken premises↩1↩2