XYBench: Strongest LLMs Answer Literally and Detect Under 63% of User Misconceptions
MaartenSap · x · 2026-10-25
Researchers including CMU's Maarten Sap introduce XYBench, a benchmark of 8,115 queries testing whether LLMs handle XY-problem-style questions containing misconceptions (e.g., parsing XML with regex). Evaluation uses three criteria from cooperative response theory: presence of a pragmatic solution, emphasis on it, and identification of the misconception.
Even the strongest models mostly answer the literal request (0.75–0.92) and rarely the intended one (0.33–0.71), and lag well behind humans at spotting misconceptions (≤63% vs. 79–90%). Models overwhelmingly prefer pragmatic answers in multiple choice yet fail to produce them freely. Oracle ablations show explicit user intent helps, but a large gap remains—suggesting pragmatic redirection is a fundamentally underdeveloped capability. Poster at ICLR on Oct 27.
More from Research
- Safety researcher: verification should go all the way down to Maxwell's equations, not stop at assembly — JacquesThibs · 2026-10-25
- Local 35B authored 800 math problems under cloud-agent supervision — authoring erred 5x more than cold-solving — Individual_Page9676 · 2026-10-25
- SA-LapNet fixes spin exactly in neural-network quantum Monte Carlo, breaking the purity-energy trade-off — bravo_abad · 2026-10-25
- Multi-domain training mix boosts supervised hallucination detectors 7.25% on out-of-distribution QA — MaartenSap · 2026-10-25
- EVALUESTEER: best reward models fail to align with user values over 25% of the time — MaartenSap · 2026-10-25
- Ouvia study: only about half of real-world speech translation interactions rated usable — MaartenSap · 2026-10-25