COT: Generative Methods for Goal-Directed Planning in Narrow Domains

Hyper-personalization for #wellness management requires multimodal support, including goal-directed planning for behavioral changes. The Uplift Array AI and product teams on a data model and AI architectures that support generative ML for planning. Generative methods for constructing action plans for various domains of wellness require an approach that differs from querying large language models for two significant reasons. First, the expertise and symbolic reasoning necessary to build personalized wellness programs must be prioritized. We have experts and local knowledge bases with specialized information. 

Second, frozen LLMs are currently not optimized for generating goal-directed plans, especially for narrow domains. This paper's chain-of-thought (#COT) prompt methods support mathematical and common sense reasoning. Adding a local knowledge base tokenized and vectorized produces finetuned prompts and promising results in narrow domains and is a viable option for generative pipelines.   

#LLM #largelanguagemodels #prompt #promptengineering #symbolicreasoning #chainofthought #AI #ML #generativeML #GPT #LaMDA #PaLM

https://arxiv.org/pdf/2201.11903.pdf