step-back · reasoning · palm-2l
recommended confidence: moderate no check recorded
Substantial accuracy gains for PaLM-2L on MMLU STEM, TimeQA, MuSiQue, and StrategyQA reasoning benchmarks
- On MMLU high-school Physics and Chemistry, Step-Back Prompting improved PaLM-2L accuracy to 73.2% and 81.8% from baselines of 66.4% and 70.9%, surpassing GPT-4's 70.3% and 79.9%. arXiv:2310.06117
- On TimeQA, Step-Back Prompting with RAG lifted PaLM-2L accuracy from 41.5% to 68.7%, versus 57.4% for plain RAG. arXiv:2310.06117
- On MuSiQue, Step-Back Prompting with RAG reached 42.8% versus the PaLM-2L baseline of 35.5%. arXiv:2310.06117
- On StrategyQA, Step-Back Prompting alone essentially matched the baseline (82.7% vs 82.8%) while Step-Back + RAG reached 86.4%. arXiv:2310.06117
Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models · arXiv preprint, 2023 not read in full