In recent times, large language models (LLMs) have evolved and transformed natural language processing with their few-shot prompt techniques. These models have extended their usability in almost all domains, ranging from machine translation, natural language understanding, text completion, sentiment analysis, speech recognition, etc. With the few shots hint approach, LLMs are given a few examples of a particular task, along with some natural language instructions and usage; they are able to adapt and learn to perform the task correctly. Tasks that require iterative steps and constraint propagation come with many limitations when using these request techniques, to overcome which a new approach has been introduced.
A team of researchers from Microsoft Research, Redmond, USA, recently introduced a new method called Reprompting, which addresses all the limitations that come with prompting techniques. This approach automatically looks for some useful and effective chain of thought (CoT) indicators. Thought chain prompts help improve the reasoning ability of large language models and help them perform complex reasoning tasks. For this, some thought demonstration chains are provided as examples during the prompts. Reprompting finds CoT prompts very efficiently without any human involvement.
Researchers have used an iterative sampling approach known as Gibbs sampling in the Reprompting algorithm. Frame the problem as sampling from a joint distribution of CoT recipes. Since the distribution is difficult to characterize directly, Gibbs sampling has been used as an approximation method. This sampling method helps determine the best instructions by trying different ones and deciding which one works best.
The reprogramming algorithm starts with a sample of initial CoT recipes with the help of zero shot indications, where no indication information is provided. Zero trigger prompts allow an LLM to generate task responses without prior training. The algorithm then iteratively samples new recipes using previously sampled solutions as top prompts, and these new recipes are used to solve other training problems, with the goal of finding a set of prompts that share similar CoT prompts.
The algorithm has been tested on all five Big-Bench Hard (BBH) tasks that require multi-step reasoning. BBH focuses on tasks that are believed to be beyond the capabilities and potential of current language models. ChatGPT and InstructGPT have been used as LLM for algorithm evaluation. Upon evaluation, reprompting has been shown to perform better than human-written, zero-shot, and few-shot QoT prompting techniques.
Iteration also showed significant potential in model blending by using different LLMs to initialize and test new recipes. It can help in the transfer of knowledge from a stronger model to a weaker model, resulting in noticeably better performance than the weaker model shows. Reprompting outperformed human-written CoT on BBH tasks by up to 17 points. Researchers have mentioned that CoT recipes that work well in one model may not work well in another, highlighting the need to optimize CoT so that each model has fairer comparisons.
In summary, the Reprompting algorithm is an excellent automated approach to find effective CoT prompts for LLM without human intervention. It is a valuable approach to address the limitations of existing methods and achieve superior performance on tasks that require multi-step reasoning.
review the Paper. Don’t forget to join our 21k+ ML SubReddit, discord channel, and electronic newsletter, where we share the latest AI research news, exciting AI projects, and more. If you have any questions about the article above or if we missed anything, feel free to email us at [email protected]
🚀 Check out 100 AI tools at AI Tools Club
Tanya Malhotra is a final year student at the University of Petroleum and Power Studies, Dehradun, studying BTech in Computer Engineering with a specialization in Artificial Intelligence and Machine Learning.
She is a data science enthusiast with good analytical and critical thinking, along with a keen interest in acquiring new skills, leading groups, and managing work in an organized manner.
