The Pulse

Ai2’s XYBench tests whether AI answers address users’ real goals

Ai2 released a benchmark of 8,115 questions designed to test whether AI assistants recognize misconceptions in users’ requests. In its evaluation, five models were less likely than human reference answers to steer users away from mistaken approaches.

AI.info Team · · fact-checked · against 2 sources ↓ · Published 13 Nov 2026

Ai2 has released XYBench, a test of whether AI assistants can recognize when a user’s question rests on a mistaken assumption and respond to the underlying problem instead. In tests of five models, the research organization found that all were less likely than human reference answers to steer users away from the mistaken approach.

The benchmark contains 8,115 questions spanning programming and everyday tasks. It examines the “XY problem”: a person wants to achieve a goal, assumes a particular method will work and asks for help with that method rather than the goal itself.

Questions behind the questions

One example concerns frosting that slides off cupcakes. A user might ask how to thicken the frosting, when the issue is that the cupcakes are still too warm; a useful response would address the temperature rather than concentrate on changing the frosting.

The benchmark was developed with researchers at Carnegie Mellon, NVIDIA and Johns Hopkins. Its paper received the Best Resource Paper Award at EMNLP 2026.

The collection includes 1,557 questions from Stack Overflow, 184 from Stack Exchange and 102 manually curated examples. The remaining 6,272 questions were generated from WikiHow guides.

For forum questions, the researchers used the three highest-scored human answers and, where available, the accepted answer to establish the intended request and possible solutions. For WikiHow examples, they used Sonnet 4.5 to generate questions around plausible misconceptions and refined them to give enough context for an expert to infer the goal.

What the benchmark measures

XYBench separates three aspects of a response: whether it contains solutions to the stated request and the intended goal; which of those the answer emphasizes; and whether it directs the user toward the goal while discouraging or avoiding the mistaken approach. The distinction matters because an answer can mention a useful alternative without making it the focus.

Claude Sonnet 4.5 acts as the automated judge, classifying suggestions and assessing emphasis. Two human reviewers independently assessed 500 model responses to check its judgments. Average agreement with the judge ranged from 97.5–100% on whether a solution appeared, 84.6–93.3% on misconception identification and 74.2–79.8% on emphasis, according to Ai2.

Models and human answers

The main evaluation tested Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro in a single-turn setting, with one response per question. Human reference answers were evaluated separately using the same measures.

On Stack Overflow questions, the models included solutions to the stated request in 75–86% of responses, compared with 59–71% for the underlying goal, Ai2 reported. The models focused primarily on the underlying goal in 35–41% of responses and on the stated request in 47–57%.

The top human answer focused on the underlying goal 55% of the time, compared with 35–41% for the models. For steering users toward the goal rather than the mistaken approach, the human figure was 79%, while model results ranged from 47% to 60%.

Knowing intent was not enough

In another experiment, researchers supplied the intended goal alongside the question. For Sonnet 4.6 on Stack Overflow, the share of answers including an underlying-goal solution rose from 59% to 67%, but the misconception-identification rate fell from 47% to 34%, according to Ai2.

A separate test asked the five models and GPT-4.1 to choose between an answer focused on the literal request and one addressing the misconception and underlying goal. When shown two possible answers for Stack Overflow questions, all six selected the goal-focused response in 71–84% of cases.

The benchmark currently covers English-language queries and mainly procedural problems. Extending it to other languages, cultures and conversational settings is an open direction for future work.