88Z AI, 21 July — Meituan's LongCat team released the LoHoSearch benchmark, generating difficult questions using Wikipedia knowledge graphs. Tests showed GPT-5.5 achieved only 34.74% accuracy, exposing bottlenecks in search agents for complex reasoning tasks.
At a glance
- LoHoSearch generates difficult questions from Wikipedia knowledge graphs, increasing difficulty via dual dimensions.
- GPT-5.5 achieved only 34.74% accuracy, revealing bottlenecks in search agents for long-chain complex reasoning.
The update
Meituan's LongCat team has officially launched LoHoSearch, a next-generation evaluation benchmark for search agents. Built on a knowledge graph of 7.62 million Wikipedia entities, it automatically generates 544 high-difficulty test questions. By controlling both search space and structural complexity, LoHoSearch significantly raises the evaluation bar. Results show that the current leading model, GPT-5.5, achieved an accuracy of only 34.74%, far below its performance on BrowseComp. This gap reveals significant technical bottlenecks in current search agents when handling long-chain complex reasoning tasks, making it difficult to cope with challenges involving high structural complexity.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 龙猫LongCatSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.