People often ask an assistant for help with the wrong thing. Someone asks how to thicken frosting that keeps sliding off, when the real problem is that the cupcakes are still too warm. XYBench, a new benchmark from researchers at Carnegie Mellon, NVIDIA and Johns Hopkins, released through Allen AI, measures how often chatbots spot that kind of mistaken assumption. The short answer is not often enough.
The benchmark has 8,115 questions from four sources: 1,557 from Stack Overflow, 184 from other Stack Exchange sites, 102 written by hand and 6,272 generated from WikiHow guides. Each is labelled with what the person asked for and what they were actually trying to do. Responses are scored three ways: whether they contain a fix for either request, which request they focus on, and whether they steer the user away from the mistaken approach. A Claude model acts as judge, and two human reviewers checking 500 responses agreed with it between 74.2% and 100% of the time depending on the measure.
The results cover Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro. On Stack Overflow questions, every model was more likely to answer the literal question than the real one: fixes for the stated request appeared in 75-86% of answers, fixes for the underlying goal in 59-71%. Models focused mainly on the real goal in only 35-41% of answers. Telling a model the user's true goal helped a little: Sonnet 4.6's rate of including the right fix rose from 59% to 67%, but its rate of pointing out the misconception fell from 47% to 34%. The most telling result: offered two answers, all six models tested picked the goal-focused one in 71-84% of cases, so they can recognise good advice far more often than they write it.
For everyday use, this matches a familiar experience with ChatGPT, Claude or Gemini: ask a narrow question and you get a thorough answer to exactly that. The practical habit is to describe the outcome you want, not just the step you think will get you there. The paper won the Best Resource Paper Award at EMNLP 2026.
The caveats are standard for benchmarks. Several of the models tested are a generation or more behind the current releases, so newer models may do better. Most questions are generated from WikiHow rather than collected from real users, the benchmark is English-only and mostly about procedural problems, and the tests use single-turn exchanges, where a real conversation might surface the true goal in a follow-up.