Update README.md
Browse files
README.md
CHANGED
|
@@ -2,4 +2,37 @@
|
|
| 2 |
license: mit
|
| 3 |
language:
|
| 4 |
- zh
|
| 5 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: mit
|
| 3 |
language:
|
| 4 |
- zh
|
| 5 |
+
---
|
| 6 |
+
# xbench-evals
|
| 7 |
+
|
| 8 |
+
🌐 [Website](https://xbench.org) | 📄 [Paper](https://xbench.org/files/xbench_profession_v2.4.pdf) | 🤗 [Dataset](https://huggingface.co/datasets/xbench)
|
| 9 |
+
|
| 10 |
+
Evergreen, contamination-free, real-world, domain-specific AI evaluation framework
|
| 11 |
+
|
| 12 |
+
xbench is more than just a scoreboard — it's a new evaluation framework with two complementary tracks, designed to measure both the intelligence frontier and real-world utility of AI systems:
|
| 13 |
+
- AGI Tracking: Measures core model capabilities like reasoning, tool-use, and memory
|
| 14 |
+
- Profession Aligned: A new class of evals grounded in workflows, environments, and business KPIs, co-designed with domain experts
|
| 15 |
+
|
| 16 |
+
We open source the dataset and evaluation code for two of our AGI Tracking benchmarks: ScienceQA and DeepSearch.
|
| 17 |
+
|
| 18 |
+
## xbench-ScienceQA
|
| 19 |
+
ScienceQA is part of xbench's AGI Tracking series, focused on evaluating fundamental knowledge capabilities across scientific domains. For detailed evaluation procedures and further information, please refer to the [website](https://xbench.org/#/agi/scienceqa) and Eval Card [xbench-ScienceQA.pdf](https://xbench.org/files/Eval%20Card%20xbench-ScienceQA.pdf) (Chinese version)
|
| 20 |
+
|
| 21 |
+
| Rank | Model | Company | Score | BoN (N=5) | Time cost (s) |
|
| 22 |
+
|------|:-----------------------------------:|:------------:|:-----:|:---------:|:-------------:|
|
| 23 |
+
| 1 | o3-high | OpenAI | 60.8 | 78.0 | 87.7 |
|
| 24 |
+
| 2 | Gemini 2.5 Pro | Google | 57.2 | 74.0 | 63.7 |
|
| 25 |
+
| 3 | Doubao-1.5-thinking-pro | ByteDance | 53.6 | 69.0 | 116.9 |
|
| 26 |
+
| 4 | DeepSeek-R1 | DeepSeek | 50.4 | 71.0 | 161.6 |
|
| 27 |
+
| 5 | o4-mini-high | OpenAI | 50.4 | 67.0 | 48.2 |
|
| 28 |
+
| 6 | Claude Opus 4 - Extended Thinking | Anthropic | 46.6 | 69.0 | 30.8 |
|
| 29 |
+
| 7 | Gemini 2.5 flash | Google | 46.2 | 70.0 | 24.1 |
|
| 30 |
+
| 8 | Qwen3 - Think | Alibaba | 45.4 | 66.0 | 105.9 |
|
| 31 |
+
| 9 | Grok 3 Mini (with Thinking) | xAI | 42.6 | 53.0 | 193.1 |
|
| 32 |
+
| 10 | Claude Sonnet 4 - Extended Thinking | Anthropic | 39.4 | 61.0 | 28.3 |
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
## Submit your agent
|
| 36 |
+
If you are developing an AI agent and would like to evaluate it using the latest version of xbench, we welcome you to contact us. Please submit a public access link of your agent, and we will complete the evaluation within an agreed timeframe and share the results with you promptly.
|
| 37 |
+
|
| 38 |
+
Contact: [team@xbench.org]()
|