Allen AI's XYBench finds chatbots answer the question asked, not the problem behind it
AI2 and collaborators released XYBench, an 8,115-question benchmark testing whether AI assistants spot the misconception hidden in a user's request and redirect them to what they actually need. Five leading models trailed human answers.
In plain words
People often ask AI for help with a specific method when their real goal needs a different approach—like asking how to fix sliding frosting when the warm cupcakes are the actual problem. XYBench tests whether assistants catch this mismatch. In AI2's evaluation, five major models usually answered the literal question but rarely steered users toward the better solution, unlike human reference answers.
What was announced
The Allen Institute for AI, working with researchers at Carnegie Mellon, NVIDIA, and Johns Hopkins, built XYBench from 8,115 real-world questions: 1,557 from Stack Overflow, 184 from Stack Exchange, 102 curated examples, and 6,272 generated from WikiHow guides. Each question pairs a stated request with an underlying goal inferable from context. The benchmark won the Best Resource Paper Award at EMNLP 2026 and measures three things per response: whether solutions address the stated request, the underlying goal, or both; where the answer's emphasis falls; and whether the model identifies the misconception.
Evaluated in single-turn settings, Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro all included solutions to the stated request more often than to the underlying goal. On Stack Overflow, stated-request solutions appeared in 75–86% of responses versus 59–71% for goal solutions, and models focused primarily on the underlying goal in only 35–41% of answers. An oracle experiment supplying the user's intent explicitly helped somewhat but did not close the gap with humans on misconception identification. When judging between two answers rather than writing one, the models chose the goal-focused response 71–84% of the time, suggesting they can recognize good advice better than they produce it. Claude Sonnet 4.5 served as the automated judge, with human agreement ranging from 74.2–79.8% on emphasis to 97.5–100% on solution detection.
What it does not show
- All findings come from the benchmark's authors; no independent evaluation is cited.
- XYBench covers only English-language queries and mainly procedural problems, so results may not generalize to other languages, cultures, or conversational settings.
- The automated judge is an AI model (Claude Sonnet 4.5); human agreement on pragmatic emphasis was 74.2–79.8%, the lowest of the three metrics.
- Models were tested in single-turn settings with one response per question, which may not reflect multi-turn assistant behavior.
For security and IT teams
- Scope
- XYBench is a research benchmark of 8,115 English-language, mostly procedural questions with published data and code, intended for measuring model behavior on misconception-based requests rather than a deployed product feature.
Sources
Source-support review completed 13 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Updates and corrections
No corrections so far.
First published 13 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.