NUS SO-Arena Benchmarks 21 Decision Models on 2,919 Tasks, GLiDE Tops at Score 70 vs GPT-6 at 80+
Summary
NUS's SO-Arena, posted Nov 13, is the first large-scale benchmark aimed squarely at the 'System One' decision model category, scoring 21 models across 2,919 tasks in 11 domains including safety, tool use, math and web interaction. Open-weight GLiDE (27B) and KnowLine-31B tie near 70 while GPT-6 Sol/Luna reference models score 79.7-82.3. The paper separates a Reliability score from Capability and finds the best capability model (GLiDE) ranks only 7th in reliability, with JevAny-27B topping reliability despite ranking 11th in capability.
Originally reported by huggingface.co
Read the original article →Original headline: NUS SO-Arena Benchmarks 21 Decision Models on 2,919 Tasks, GLiDE Tops at Score 70 vs GPT-6 at 80+