Introducing the Open Chain of Thought Leaderboard
Hugging Face Blog
A new leaderboard launched to measure how effectively language models generate step-by-step reasoning traces, rather than scoring absolute accuracy on multiple-choice reasoning tasks. The evaluation uses four benchmarks from AGIEval (LogiQA and LSAT subsets) and measures accuracy gain as the difference between performance with and without chain-of-thought prompting across six generation regimes combining two prompting strategies and three decoding methods. Early results with 30 models show that smaller 7-billion-parameter models sometimes outperform larger ones at chain-of-thought reasoning, instruction-tuning improves both baseline and marginal gains, and no single prompting approach works best across all models and tasks.