XYBench WikiHow - Fully Pragmatic Response: leaderboard

Each model answers 6,272 synthetic how-to questions generated from WikiHow articles in which the user asks about a mistaken approach to an underlying goal. The score is the percentage of answers that, by a Claude Sonnet 4.5 judge, point out the misconception, include a solution to the underlying goal and focus on it; the Stack Overflow solution, emphasis and misconception boards from Ai2's blog score those parts separately.

Metric: Rate (%). Source: arxiv.org. Saturation forecast: Around 2030. 5 models tracked. Updated 16 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)18#93
2DeepSeek V4 Pro17#191
3GPT-516#167
4Claude Sonnet 4.615#165
5Qwen 3.5 9B15#814

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=xybench-wikihow-fully-pragmatic-response · How It Works · Data refreshed daily, snapshot 2026-11-16.