ARC-C

text
+
+
+
+
About

ARC-C (ARC Challenge) is an advanced reasoning benchmark featuring complex multiple-choice science questions that require deep understanding and integration of knowledge across various domains. More challenging than ARC-E, it tests sophisticated reasoning, synthesis of information, and domain-specific knowledge in AI systems. The benchmark is designed to challenge even state-of-the-art AI models with questions requiring advanced logical inference and scientific comprehension.

+
+
+
+
Evaluation Stats
Total Models31
Organizations11
Verified Results0
Self-Reported31
+
+
+
+
Benchmark Details
Max Score1
Language
en
+
+
+
+
Performance Overview
Score distribution and top performers

Score Distribution

31 models
Top Score
96.9%
Average Score
77.6%
High Performers (80%+)
15

Top Organizations

#1Anthropic
3 models
92.9%
#2Amazon
3 models
92.5%
#3AI21 Labs
2 models
89.3%
#4Meta
4 models
88.4%
#5Microsoft
3 models
86.4%
+
+
+
+
Leaderboard
31 models ranked by performance on ARC-C
LicenseLinks
Jul 23, 2024
Llama 3.1 Community License
96.9%
Feb 29, 2024
Proprietary
96.4%
Jul 23, 2024
Llama 3.1 Community License
94.8%
Nov 20, 2024
Proprietary
94.8%
Feb 29, 2024
Proprietary
93.2%
Aug 22, 2024
Jamba Open Model License
93.0%
Nov 20, 2024
Proprietary
92.4%
Jan 30, 2025
Apache 2.0
91.3%
Aug 23, 2024
MIT
91.0%
Nov 20, 2024
Proprietary
90.2%
Showing 1 to 10 of 31 models
+
+
+
+
Resources