Shulou(Shulou.com)11/24 Report--

CTOnews.com, June 7, the biggest weakness of the current GPT-4 model is the arithmetic ability. Because the logical reasoning ability of the model has yet to be improved, even if many people think that the calculation problem is relatively simple, GPT-4 can not get the correct results.

Recently, researchers at the National University of Singapore launched the Goat model, saying it is "specially designed for arithmetic problems". "after fine-tuning the LLaMA model, Goat achieved higher accuracy and better performance than GPT-4 in arithmetic," the researchers said.

▲ source Arxiv researchers have proposed a new method to classify tasks according to the learnable type of arithmetic, and then use the basic arithmetic principle to decompose the non-learnable tasks into a series of learnable tasks (CTOnews.com Note: disassemble the complex computing process into simple steps) and import them into the AI model.

This new method can make the model learn the answer pattern and generalize the process into invisible data, rather than relying solely on "weight memory calculation", so it can effectively improve the arithmetic performance. in zero-sample learning, the answer can be generated by adding and subtracting large numbers with "near-perfect accuracy".

▲ image source Arxiv researchers train on GPU with 24 GB memory, and test the resulting model using BIG-bench arithmetic sub-task. The accuracy results are outstanding, leading the industry models such as Bloom, GPT-NeoX, OPT and so on. Among them, the accuracy of Goat-7B with zero samples is even higher than that of PaLM-540 model with few samples, and it is much better than GPT-4 in the calculation of large numbers.

CTOnews.com 's friends can find links to the paper here.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.