metadata
license: mit
task_categories:
- question-answering
language:
- en
- zh
ORION Evaluation Results
ORION is a multilingual benchmark designed to evaluate open-domain reasoning across diverse web-related domains. Each example in the dataset requires multi-step logical composition grounded in verifiable sources, challenging advanced AI assistants and retrieval-augmented models. The dataset consists of 310 questions (170 in Chinese and 140 in English), each accompanied by verified answers, diverse acceptable variants (e.g., aliases or synonymous expressions, separated by |), and evidence URLs to ensure fair and flexible evaluation.The table below reports the accuracy of three AI systems on ORION.
| AI System | Chinese (%) | English (%) | Overall (%) |
|---|---|---|---|
| Kimi Exploration Edition | 14.7 | 20.0 | 17.1 |
| Doubao Search | 23.5 | 30.7 | 26.8 |
| Qwen2.5-Max Search | 20.0 | 20.7 | 20.3 |