Amazon has released a new language model that outperforms GPT-3.5 on ScienceQA.
Just over two months ago, OpenAI launched ChatGPT to the public, catapulting the AI-powered chatbot into mainstream discourse, sparking debates about how it might change business, education, and more.
Google and Baidu, based in China, launched chatbots to demonstrate their so-called “generative AI” (technology that can make conversational text, graphics, and more).
As of now, Amazon’s new language models outperform GPT-3.5 by 16 percentage points (75.17%) and even outperform many humans on the ScienceQA benchmark.
The ScienceQA benchmark is a large set of multimodal science questions with annotated answers. It has over 21,000 multimodal multiple-choice questions (MCQs).
As a result of recent technological advances, large language models (LLMs) can perform well on tasks requiring complex reasoning. It is done through chain-of-thought prompting, which involves developing intermediate steps of sense to illustrate how something is done.
Researchers often use the Multimodal-CoT paradigm to investigate CoT reasoning in multimodality, but most current work on CoT only examines the language modality. Multimodality relies on multiple inputs like vision and language.
According to Amazon researchers, using visual features helps develop more effective rationales, which result in more accurate answer inferences.
They show that 1B-models outperform GPT-3.5 on the ScienceQA benchmark by 16% using Multimodal-CoT. Their error analysis implies that future research may benefit from utilizing more efficient visual characteristics, incorporating common sense information, and utilizing filtering techniques to enhance CoT reasoning.
Industry behemoths are already working to develop a standard for chatbot development. Amazon has recently joined the fight. Other businesses must take action because these rivalries will surely pave the way for the greatest solution and item. Let’s wait and see.