> ## Content Index
> Fetch the complete content index at: https://soundbytes.news/llms.txt
> Use this file to discover other available public pages before exploring further.

# A new test checks whether chatbots spot the real problem behind a question
- URL: https://soundbytes.news/a-new-test-checks-whether-chatbots-spot-the-real-problem-behind-a-question/
- Published: 2026-11-13T17:22:06.000Z
- Updated: 2026-11-18T08:49:50.000Z
- Description: Ai2 has published XYBench, a test of 8,115 real-world questions that checks whether chatbots spot a mistaken assumption. In the lab's own tests, five chatbots often stayed with the mistaken approach.
- Author: Echo
- Tags: Tested and measured

Someone asks how to thicken frosting that keeps sliding off cupcakes. In Ai2's example, the real problem is that the cupcakes are still too warm. Ai2's new test, [XYBench](https://allenai.org/blog/xybench?ref=soundbytes.news), checks whether chatbots notice such a mistake, and Ai2 describes it in a post dated 13 November 2026.

The mismatch is called the XY problem. A person wants to achieve X, assumes approach Y will work, and asks for help with Y. Ai2 developed the test with researchers at Carnegie Mellon, NVIDIA and Johns Hopkins.

XYBench has 8,115 real-world questions about programming and everyday activities such as cooking and home repair. They come from Stack Overflow, a programming question-and-answer site, and from Stack Exchange. Others are hand-picked examples, and the rest were generated by an AI model from WikiHow how-to guides.

Each answer is scored on three things: whether it includes a fix for the stated request and for the real goal, which of the two it focuses on, and whether it steers the user toward the real goal and away from the mistaken approach.

Another AI model, Claude Sonnet 4.5, marks each answer. Two human reviewers checked 500 responses. They agreed with the AI judge least often on which request an answer focused on: 74.2% to 79.8% of the time.

Ai2 tested five chatbots (AI models): Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro, with one response per question. In the lab's own tests, the chatbots often suggested a fix but kept their answers focused on the original, mistaken approach.

On Stack Overflow questions, a fix for the stated request appeared in 75% to 86% of responses. A fix for the real goal appeared in only 59% to 71%.

Across the benchmark, [the paper's abstract](https://huggingface.co/papers/2609.06842?ref=soundbytes.news) says that at identifying misconceptions, the chatbots scored at most 63%, against 79% to 90% for humans.

Telling Sonnet 4.6 the real goal outright helped only in part. Sonnet 4.6 then included a fix for the real goal more often (up from 59% to 67%) but steered people away from the mistaken approach less often (down from 47% to 34%).

Shown two answers, one literal and one goal-focused, six chatbots (the five above plus GPT-4.1) picked the goal-focused one in 71% to 84% of Stack Overflow cases. Ai2 says this exposes a gap between recognising useful advice and producing it, which a multiple-choice test alone would miss.

Ai2 says XYBench currently covers English-language questions and mainly step-by-step problems. Extending it to other languages, cultures and conversations remains an open direction, it says.

[The paper](https://aclanthology.org/2026.emnlp-main.2646/?ref=soundbytes.news) won Best Resource Paper at the 2026 Conference on Empirical Methods in Natural Language Processing, and Ai2 has published it with the dataset and code. Ai2 says researchers can use XYBench to test whether changes to prompting or training help models identify misconceptions.

## Sources

- [XYBench: Helping with the problem behind the question](https://allenai.org/blog/xybench?ref=soundbytes.news) · allenai.org
- [Paper page - XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions?](https://huggingface.co/papers/2609.06842?ref=soundbytes.news) · huggingface.co
- [XYBench: Can LLMs Respond Pragmatically to Queries with Misconceptions? - ACL Anthology](https://aclanthology.org/2026.emnlp-main.2646/?ref=soundbytes.news) · aclanthology.org