SearchGymBench - Complex QA: leaderboard
Tool-based search over a synthetic world from SearchGymBench (Zhang, He, Zhu, Wu, Yu, Chu, Zhang, Tan and Jia, 2026): questions about fictional entities whose answers must be found with search and page-access tools in a generated Wikipedia-style corpus, so memorised knowledge cannot help. Accuracy in percent on the complex questions (180 in the released test set) that join two parallel lookups or nest one question inside another, judged by Qwen2.5-72B-Instruct with up to 64 agent turns; higher is better.
Metric: Accuracy (%; parallel and nested multi-question tasks over a synthetic Wikipedia-style corpus answered with search and page-access tools, Qwen2.5-72B-Instruct judge; up to 64 turns). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 47.6 | #403 |
| 2 | Kimi K2 | 24.4 | #435 |
| 3 | Qwen 3 8B | 3.3 | #1838 |
| 4 | Qwen 3 4B | 3.3 | #2297 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=searchgymbench-complex-qa · How It Works · Data refreshed daily, snapshot 2026-11-12.