The XY Problem, Now With a Benchmark
You want to frost cupcakes. The frosting keeps sliding off. You ask an AI how to thicken the frosting. The AI helpfully explains cornstarch ratios and refrigeration techniques.
The real problem? Your cupcakes are still warm.
This is the XY problem: you want X (frosted cupcakes), assume approach Y (thicker frosting) will work, and ask only about Y. AllenAI just released XYBench, an 8,000+ question benchmark testing whether LLMs can recognize when a user's literal request is based on a misconception—and redirect them to what they actually need. It won Best Resource Paper at EMNLP 2026.
The results are sobering. Every frontier model tested—Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro—showed the same pattern: they often mentioned better solutions but kept their answers focused on the user's mistaken approach. Human experts did the opposite.
What Makes This Hard
XYBench draws from 8,115 real-world questions spanning programming (Stack Overflow, Stack Exchange) and everyday domains (cooking, home repair). For each question, the benchmark identifies two things:
- The stated request (Y): what the user explicitly asks for
- The intended request (X): the underlying goal that can be inferred from context
The clever bit is the filtering. XYBench only includes cases where answering the literal question wouldn't solve the underlying problem. If thickening the frosting would actually work, it's out. The benchmark specifically tests whether models can detect when following the user's stated approach is a dead end.
For the WikiHow-derived examples, the team used Sonnet 4.5 to generate questions around plausible misconceptions in procedural guides, then refined them to include enough context for experts to infer intent without stating it outright. That's 6,272 synthetic examples grounded in real-world procedures.
Three Ways to Fail
XYBench measures three distinct behaviors, and models struggle with all three:
Pragmatic Solution
Does the response include solutions to Y (stated request) and X (intended request)? On Stack Overflow questions, models included solutions to the stated request in 75–86% of responses. Solutions to the underlying goal? Only 59–71%.
Models are more likely to answer what you asked than what you need.
Pragmatic Emphasis
Mentioning a solution doesn't mean emphasizing it. A model might suggest checking for a clogged vacuum hose but spend most of its answer recommending more powerful vacuum models. XYBench measures where the response focuses.
On Stack Overflow, models focused primarily on the underlying goal in just 35–41% of responses. They focused on the stated (mistaken) request in 47–57%. The remaining responses split emphasis roughly evenly.
Human reference answers? They prioritized the underlying goal far more consistently.
Misconception Identification
Does the response actively steer users away from the mistaken approach while redirecting toward the actual goal? This is the hardest test—it requires recognizing the misconception and explicitly addressing it.
Models lagged humans significantly on this metric across all question sources. Even when models mentioned better solutions, they rarely discouraged the user's flawed approach.
The Oracle Experiment
What if you just tell the model what the user really wants? XYBench includes an oracle experiment where both the stated request and intended request are supplied explicitly alongside each question.
On Stack Overflow, Sonnet 4.6's rate of including a solution to the underlying goal rose from 59% to 67%. Helpful! But its misconception-identification rate fell from 47% to 34%.
Making intent explicit helped models find relevant solutions but made them worse at steering users away from bad approaches. Across all sources with human reference answers, every model still trailed humans on misconception identification even with oracle information.
This suggests detecting intent is only part of the challenge. Models also need to act on it when it diverges from the stated request.
The Recognition Gap
Here's where it gets interesting. XYBench includes a multiple-choice experiment: given each question and two possible answers (one focused on the literal request, one explaining the misconception and addressing the underlying goal), which does the model prefer?
All six models tested—the original five plus GPT-4.1—chose the goal-focused answer in 71–84% of cases on Stack Overflow questions.
But when those same models wrote their own answers in the main evaluation? Only 35–41% primarily focused on the underlying goal.
Models can recognize useful advice. They're much worse at producing it. A multiple-choice test alone would completely miss this gap.
Why This Matters
The XY problem isn't edge-case user error. It's how people naturally interact with tools when they're uncertain. You don't know what you don't know, so you ask about the approach that makes sense to you—even when it's wrong.
If AI assistants only answer literal questions, they're useful mainly to experts who already know how to ask the right questions. For everyone else, they risk becoming expensive confirmation bias machines: they'll help you execute your bad plan really efficiently.
The benchmark methodology is solid. Claude Sonnet 4.5 serves as the automated judge, and human validation on 500 responses showed 97.5–100% agreement on whether solutions appeared, 84.6–93.3% on misconception identification, and 74.2–79.8% on pragmatic emphasis. That's strong enough to trust the automated evaluation at scale.
What XYBench Doesn't Cover
The benchmark currently covers English-language queries and mainly procedural problems. Extending it to other languages, cultures, and conversational settings remains open work.
The single-turn evaluation is also a limitation. Real conversations let users clarify intent across multiple exchanges. XYBench tests whether models get it right on the first response—a harder test, but not representative of how these tools are used in practice.
And the benchmark can't tell you why models fail. Is it training data? Reward modeling? Instruction-following bias? Fundamental limits of next-token prediction? XYBench surfaces the gap; fixing it requires different tools.
The Pragmatics Problem
What I find compelling about XYBench is that it's fundamentally testing pragmatics: the branch of linguistics concerned with how context and communicative intent shape meaning. This isn't a capabilities benchmark in the usual sense. Models demonstrably can generate better solutions—they just don't prioritize them.
That's a different kind of failure than "the model doesn't know the answer." It's more like "the model optimizes for the wrong thing." And it's not clear that throwing more parameters or training data at the problem necessarily helps.
The question XYBench raises: what training methods teach models to act on inferred intent rather than literal requests? RLHF on human preferences? Constitutional AI? Something else entirely?
We're building assistants that are supposed to help users who don't know what they need. XYBench is one of the first benchmarks that actually measures whether they can.