Ai2's XYBench finds AI assistants keep answering the question asked, not the problem behind it
The Allen Institute for AI (Ai2) has released XYBench, a benchmark of 8,115 questions that tests whether AI assistants notice when a user is asking about the wrong thing. The announcement on the Hugging Face blog, dated 13 November, says the paper won the Best Resource Paper Award at EMNLP 2026. It was developed with researchers at Carnegie Mellon, NVIDIA and Johns Hopkins.
The benchmark is named after the XY problem: someone wants to achieve X, assumes approach Y will work, and asks only about Y. Ai2's example is a person asking how to thicken frosting that keeps sliding off cupcakes, when the real problem is that the cupcakes are still too warm.
The questions come from four sources: 1,557 from Stack Overflow, 184 from other Stack Exchange sites, 102 curated by hand and 6,272 generated from WikiHow guides with Claude Sonnet 4.5 and then refined. For each, the team recorded the stated request and the underlying goal, using top-voted and accepted human answers where available, and removed cases where answering the literal question would solve the problem. Responses are scored on three measures: whether they contain a solution to the stated request and to the real goal, which of the two they emphasise, and whether they steer the user away from the mistaken approach. Sonnet 4.5 acts as the automated judge; two human reviewers checking 500 responses agreed with it 97.5-100% of the time on whether a solution appeared, but only 74.2-79.8% on emphasis.
The five evaluated models were Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro. All were more likely to answer the stated request than the real goal. On Stack Overflow, solutions to the stated request appeared in 75-86% of responses and to the underlying goal in 59-71%, and models made the real goal the main focus in only 35-41% of answers. Every model trailed human reference answers on identifying the misconception.
Two further experiments sharpen the result. Telling Sonnet 4.6 the user's real goal raised its rate of solving it on Stack Overflow from 59% to 67% but lowered misconception identification from 47% to 34%. And when the six models tested, including GPT-4.1, were shown two answers, one literal and one goal-focused, they chose the goal-focused one 71-84% of the time.

Why it matters
The gap between choosing the better answer and writing it is the useful finding. A multiple-choice test would rate these models as good at spotting XY problems; generating answers shows they mostly do not act on it. For coding assistants, where a wrong approach can cost hours, that is the behaviour worth measuring, and the authors note the benchmark is English-only and mostly procedural.