I’m currently the Chief Evangelist @ HumanFirst. I explore & write about all things on the intersection of AI and language. Including NLU design, evaluation & optimisation. Data-centric prompt tuning & LLM observability, evaluation & fine-tuning.

Intuitively we as humans break a bigger task or problem into sub-tasks, after which we chain these sub-tasks together. Using the output of 1 sub-task because the input for the subsequent sub-test.

By utilizing chain-of-thought prompting inside the OpenAI Playground, a way wherein specific examples of chain of thought are provided as guidance, it is feasible to showcase how large language models can develop sophisticated reasoning capabilities.

Research has shown that sufficiently large language models can enable the emergence of reasoning abilities when prompted in this manner.

A chain of thought is a series of intermediate natural language reasoning steps that result in the ultimate output, and we check with this approach as chain-of-thought prompting. ~ Source

The utilisation of chain-of-thought reasoning is especially evident when examining LLM-based agents.

LLM based Agents have the flexibility to decompose a matter right into a chain-of-thought and answer the query in a piecemeal fashion. Considering each of the steps.

Inside a playground setting, the principle behind chain-of-thought prompting might be illustrated by following a few-shot learning approach.

A LLM is given a number of examples on decompose a fancy and ambiguous query. And by doing so, establish a chain-of-thought process.

As an example, the query Who's thought to be the daddy of the iPhone and what's the square root of his 12 months of birth? can only be appropriately answered by decomposing the query in sequential chain-of-thought steps.

Below is the decomposed chain-of-thought reasoning of the Agent based on the aforementioned query. The decomposition was performed completely autonomously by the Agent.

As seen within the image above, Chain-of-thought prompting enables large language models to deal with difficult arithmetic, common sense, and symbolic reasoning challenges.

Reasonably than adjusting a definite language model checkpoint for every recent task, one can simply provide the model with several input-output examples to exhibit the duty.

Consider the image below, with two OpenAI playground examples. The highest example is a few-shot training prompt, providing text_davinci-003 with one query and answer pair, followed by a matter to be answered by the LLM.

The highest example doesn’t contain a chain-of-thought reasoning example, hence the LLM only provides the reply.

The underside playground example has a straightforward chain-of-thought reasoning prompt, giving the LLM a reference as answer the query.

You possibly can see the reply given by the LLM is expanded, showing the chain-of-thought reasoning utilized by the LLM.

Moreover, advantages of this approach is improved accuracy with the model illustrating the way it reach a conclusion or answer. The augmented answer makes for a more conversational and informative response.

I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things on the intersection of AI and language; starting from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.