0
votes
0
resp
24
views
Verbatim system prompt from the Anthropic Research multi agent pattern. The 'research lead' classifies a user query as depth first, breadth first, or straightforward, then plans an
Verbatim system prompt from the Anthropic Research multi agent pattern. The 'research lead' classifies a user query as depth first, breadth first, or straightforward, then plans an
Dave Hulbert's plain English distillation of Tree of Thoughts. Works on any chat model no search algorithm required and dramatically improves multi step reasoning. Source: https://
Canonical example from Wei et al. 2022 showing how interleaving worked reasoning into few shot exemplars dramatically improves arithmetic and commonsense tasks. Source: https://git
Worked HotpotQA example from Yao et al. 2022 showing the ReAct pattern: interleave reasoning traces with tool actions to ground answers in retrieved evidence. Source: https://githu
Classic Wei et al. few shot CoT exemplars. Each example shows the reasoning trace explicitly, which conditions the model to do the same on the final query. Source: https://www.prom
Wang et al. self consistency uses the same CoT shape as Wei et al., then samples multiple reasoning paths and majority votes. Use these exemplars verbatim and sample at temperature
Classic 1 shot example from Brown et al. 2020 GPT 3 paper showing in context learning: define a new word, give one usage example, and the model generalizes. Source: https://github.