{"slug": "xybench-ai2-tests-whether-ai-assistants-spot-the-problem-behind-your-question", "title": "XYBench: Ai2 Tests Whether AI Assistants Spot the Problem Behind Your Question", "summary": "Ai2 published XYBench on November 13, 2026, an 8,115-question benchmark testing whether AI assistants detect a misconception behind a user's request and steer toward the real goal; the paper won the Best Resource Paper Award at EMNLP 2026. Testing Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro, Ai2 found the models focused mainly on the real goal in only 35-41% of Stack Overflow answers, yet chose the goal-focused answer 71-84% of the time when judging. Ai2 built XYBench with Carnegie Mellon, NVIDIA and Johns Hopkins researchers from 1,557 Stack Overflow, 184 Stack Exchange, 102 curated and 6,272 WikiHow-derived questions.", "body_md": "On November 13, 2026, Ai2 published [XYBench](https://huggingface.co/blog/allenai/xybench?ref=explainx), a benchmark that asks a simple question: when you ask an AI assistant for help with the wrong approach, does it notice? Ai2 says five frontier models often answered the stated request and kept their focus on the mistaken approach. Human reference answers did better.\n\nThe paper won the Best Resource Paper Award at EMNLP 2026, according to Ai2. This post explains the benchmark, the numbers, and what you can do with them if you build or buy an assistant.\n\n## TL;DR\n\n| Question | Answer | \n|---|---|\n| What does XYBench test? | Whether a model spots a misconception in a request and steers toward the real goal. | \n| Size | 8,115 questions: 1,557 Stack Overflow, 184 Stack Exchange, 102 curated, 6,272 WikiHow-derived. | \n| Models tested | Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, DeepSeek-V4 Pro. | \n| Main finding | On Stack Overflow, models focus mainly on the real goal in only 35-41% of answers. | \n| Surprise | The same models choose the goal-focused answer 71-84% of the time when judging. | \n| Who built it? | Ai2 with Carnegie Mellon, NVIDIA and Johns Hopkins researchers. | \n\n## What the XY problem is\n\nThe XY problem is an old idea from forums and help desks. A person wants to accomplish X, assumes approach Y will work, and asks for help with Y. The [XY problem](https://en.wikipedia.org/wiki/XY_problem?ref=explainx) is well known to programmers who answer questions online.\n\nAi2's example is baking. Someone asks how to thicken frosting that keeps sliding off cupcakes. The real problem is that the cupcakes are still warm. A helpful answer says so. A literal answer recommends more powder sugar.\n\nAi2's argument is that useful assistants must help people who ask misguided questions, not only those who ask the right question in the right detail. Users have different knowledge levels and cannot always know what to ask.\n\n## How XYBench was built\n\nFor every question the team labeled the stated request (Y) and the intended request (X), the underlying goal that can be inferred from context.\n\n- **Public forums.** 1,557 Stack Overflow and 184 Stack Exchange questions. The team used the three highest-scored answers and the accepted answer to establish X and possible solutions. It filtered out cases where answering the literal request was enough.\n- **Manual set.** 102 curated examples.\n- **WikiHow.** 6,272 questions. Ai2 used Claude Sonnet 4.5 to write questions around plausible misconceptions in procedural guides, then refined them so an expert could infer the goal without it being stated. It removed items where answering the literal request would mostly solve the problem.\n\nThat mix covers programming and everyday tasks such as cooking and home repair. Ai2 notes the benchmark is English-only and mostly procedural.\n\n## Three things it measures\n\nAi2 borrows from pragmatics, the study of how context and intent shape meaning. Each answer is scored on three aspects:\n\n1. **Pragmatic solution.** Does the response include a solution to the stated request (Y), the intended request (X), or both? Each is checked separately.\n2. **Pragmatic emphasis.** Does the answer focus on Y, on X, or balance the two? Ai2's example: a model spends paragraphs recommending a stronger vacuum cleaner and mentions a clogged hose once. It contains the right solution but the wrong emphasis.\n3. **Misconception identification.** Does the answer steer the user toward X while discouraging the mistaken approach Y?\n\nSplitting the three lets researchers see whether progress on one also moves the others.\n\n## What the experiments show\n\nAi2 ran each model in a single-turn setting with one response per question, and scored human reference answers with the same metrics.\n\n*Image: [Ai2](https://huggingface.co/blog/allenai/xybench?ref=explainx)*\n\n- Across all four sources, every model was more likely to include a solution to Y than to X.\n- On Stack Overflow, stated-request solutions appeared in 75-86% of responses, versus 59-71% for underlying-goal solutions.\n- On emphasis, models focused mainly on the underlying goal in only 35-41% of Stack Overflow responses, and on the stated request in 47-57%. The rest balanced the two.\n- Compared with human reference answers, all five models were less likely to steer users away from the mistaken approach.\n\n## Does telling the model the goal fix it?\n\nNo, not fully. In an oracle experiment Ai2 supplied both X and Y alongside each question. On Stack Overflow, Sonnet 4.6's rate of including a solution to the underlying goal rose from 59% to 67%. Its misconception-identification rate fell from 47% to 34%. Every model in this test still trailed humans on misconception identification.\n\nAi2's reading: detecting intent is only part of the challenge. Models also need to act on intent when it differs from what was literally asked. That has a practical meaning for prompt design. Adding \"my real goal is...\" to a prompt helps get a fix, but it may make the model less likely to tell you your approach is wrong.\n\n## Picking versus writing\n\nThe most striking result is a gap between judging and producing.\n\n*Image: [Ai2](https://huggingface.co/blog/allenai/xybench?ref=explainx)*\n\nAi2 gave the five models, plus GPT-4.1, each question with two candidate answers: one for the literal request, one explaining the misconception and addressing the goal. On Stack Overflow, all six picked the goal-focused answer in 71-84% of cases. Yet only 35-41% of answers the models wrote themselves focused mainly on the goal.\n\nSo the knowledge is there. It does not reliably show up in the free-form answer. A multiple-choice test would have missed that.\n\n## How trustworthy is the judge?\n\nXYBench uses Claude Sonnet 4.5 as the automated judge to classify the suggestions in each response and rate emphasis. Two human reviewers independently assessed 500 responses. Averaged across the reviewers, agreement with the judge was:\n\n| Aspect | Agreement with humans | \n|---|---|\n| Whether a solution appeared | 97.5-100% | \n| Misconception identification | 84.6-93.3% | \n| Pragmatic emphasis | 74.2-79.8% | \n\nEmphasis is the hardest to judge and the least reliable. Treat small differences in emphasis scores with care. This is the kind of grader check recommended in [Anthropic's guide to building strong safety evaluations](https://explainx.ai/blog/anthropic-strong-safety-evals-guide), which also advises using a different model as grader when possible. Note that a Claude model judged an evaluation that included a Claude model among five subjects. Ai2 did validate against humans, which is the right step, but the point is worth keeping in mind.\n\n## What this means for builders\n\n- **Add a goal-check step.** Before answering troubleshooting questions, have the assistant state what it thinks the user's goal is and whether the requested approach serves it. The picking result suggests models can recognize the better answer when asked to compare.\n- **Test your own assistant.** The data and code are linked from the Ai2 post. Run a few hundred Stack Overflow items and look at emphasis, not only correctness.\n- **Do not rely on stating the goal.** The oracle result shows that supplying X can reduce misconception identification.\n- **Watch for sycophancy overlap.** An assistant that answers the literal question and never pushes back is easier to like and less useful. The same tension appears in how labs write behavior rules, as in[OpenAI's updated Model Spec on human intent](https://explainx.ai/blog/openai-updated-model-spec-five-values-faithfulness-human-intent-chain-of-command-2026) .\n- **Pick models by job.** Our guide to[choosing among GPT-6 models](https://explainx.ai/blog/openai-model-names-luna-terra-sol-astra-ted-sanders-how-to-pick-gpt-6-model-hn-2026) applies: test on your own questions.\n\n## A worked example you can run\n\nTake a programming question in the style of the Stack Overflow set: \"How do I make my regex match across all 40,000 lines faster?\" The stated request (Y) is a faster regex. The likely goal (X) is faster log processing, where reading the file line by line with a plain string check may beat any regex tuning.\n\nScore an assistant's answer on the three XYBench axes. Does it contain a solution to Y (a regex rewrite)? To X (a different parsing approach)? Which one gets most of the words? And does it tell the user that tuning the regex may not be the right path? An answer that offers the rewrite in three paragraphs and mentions the alternative in one sentence has the right solution but the wrong emphasis, exactly the pattern Ai2 reports.\n\nTo run this yourself, write 20 questions from your own support logs where the real cause differed from the stated ask, collect answers from the assistants you use, and have two people label emphasis independently. Compare their agreement before trusting any automated judge, since emphasis was the weakest metric in Ai2's check.\n\n## Limits and open questions\n\n- English only, and mostly procedural problems.\n- A single-turn setting. In a real conversation a model could ask a clarifying question, which XYBench does not measure.\n- The WikiHow questions are model-generated, so they reflect what a model considers a plausible misconception.\n- The model list is Ai2's choice and includes older GPT-5 and Sonnet 4.6 versions, not the newest flagships. Our own coverage of newer models, such as [Fable 5.2 and its uneven abilities](https://explainx.ai/blog/claude-fable-5-2-jagged-agi-mollick-expert-judgment-2026) , suggests rerunning before drawing conclusions about today's top models.\n- The 2026 findings on how [experts follow agent judgment calls](https://explainx.ai/blog/anthropic-agents-judgment-calls-experts-followed-hidden-lean-86-percent-study-2026) show a related risk: people defer to confident answers, so an assistant that stays on the wrong path can pull a user with it.\n\nAi2 says future work should test which training methods help models act on inferred intent, and extend XYBench to other languages, cultures and conversational settings.\n\n## Related reading\n\n*Details are accurate as of November 13, 2026. Figures are as reported by Ai2.*", "url": "https://wpnews.pro/news/xybench-ai2-tests-whether-ai-assistants-spot-the-problem-behind-your-question", "canonical_source": "https://explainx.ai/blog/ai2-xybench-xy-problem-benchmark-8115-questions-models-miss-real-goal-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog", "published_at": "2026-11-13 00:00:00+00:00", "updated_at": "2026-11-13 19:24:17.379244+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-safety", "ai-agents"], "entities": ["Ai2", "XYBench", "Carnegie Mellon", "NVIDIA", "Johns Hopkins", "Claude Sonnet 4.6", "GPT-5", "Gemini 3.1 Pro"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/xybench-ai2-tests-whether-ai-assistants-spot-the-problem-behind-your-question", "markdown": "https://wpnews.pro/news/xybench-ai2-tests-whether-ai-assistants-spot-the-problem-behind-your-question.md", "text": "https://wpnews.pro/news/xybench-ai2-tests-whether-ai-assistants-spot-the-problem-behind-your-question.txt", "jsonld": "https://wpnews.pro/news/xybench-ai2-tests-whether-ai-assistants-spot-the-problem-behind-your-question.jsonld"}}