Home Page >  News List >> Tech >> Tech

Passing the MIT undergraduate math exam with a GPT-4 full score! This set of prompts has become popular

Tech 2023-06-17 14:38:46 Source: Network
AD

Hengyu originates from Aofei TempleQuantum bit | official account QbitAISurprisingly, the MIT math exam was defeated by GPT-4?!Suddenly, someone loudly announced in their latest paper work:Mathematics and EECS of GPT-4 at MIT(Department of Electrical Engineering and Computer Science)In the undergraduate degree examination,The demonstrated abilities fully meet the graduation requirements.And properlyGet a perfect score!MIT

Hengyu originates from Aofei Temple

Quantum bit | official account QbitAI

Surprisingly, the MIT math exam was defeated by GPT-4?!

Suddenly, someone loudly announced in their latest paper work:

Mathematics and EECS of GPT-4 at MIT(Department of Electrical Engineering and Computer Science)In the undergraduate degree examination,The demonstrated abilities fully meet the graduation requirements.

And properlyGet a perfect score!

MIT.

GPT-3.5.

.

GPT-4.

GPT-3.5!

Is it possible that in the future, a stronger model than GPT-4 is no longer needed to solve academic problems?

Another netizen showed his "cutting edge" of online surfing, playing a stick that YannLeCun roast about "GPT-4 is not as intelligent as a dog" these two days:

GPT-4 Open MIT Exam

Specifically, GPT-4 participated in such a test this time:

4550.

These 4550 problems and solutions, from MIT Mathematics Department and EECS students.

Including:

6-1: Electrical Science and Engineering;
6-2: Electrical Engineering and Computer Science;
6-3: Computer Science and Engineering;
6-4: Artificial intelligence and decision-making;
18-1: General Mathematics;
18-2: Applied mathematics;
18-3: Pure Mathematics;
18-C.

Detailed classification summary for each major

The questions are all from the MIT dataset, randomly generating 228 questions,Issues that do not involve images or existing solutions.

.

.

This time, those who participated in the examNot only GPT-4 and GPT-3.5, but also StableVicuna-13B, LLaMA-30B, and LLaMA-60B.

4.

GPT-4100%LLaMA-30B30%.

It is worth noting that,The original version of GPT-4 is ready to use out of the box without any tuning, and achieved a 90% score in this MIT exam.

Few-Shot+CoT+Self-critique+Experts.

GPT-4.

In addition, the research team also carried out engineering optimization in the Tooltip,Specific 'spells'As follows:

Wait, is the rater GPT-4 himself?

LLM.

2AI.

similarXiao Ming planted 5 lemon trees and received 6 lemons from each tree every year. How many lemons did he receive in total over the past 10 years.

At the beginning of last year, the joint research of MIT+Harvard+Columbia University+University of Waterloo indicated that by transforming mathematical problems into equivalent programming problems, OpenAI's Codex, GPT-3's fellow teacher, could grasp the advanced knowledge and reachMIT undergraduate level.

Learned 6 randomly selected example questions from MIT undergraduate basic mathematics courses, with 25 randomly selected questions for each of the 6 courses, plus an ACT level(American College Entrance Examination)60.

210AI.

AIMIT undergraduate levelCodex

Codex.

So, this time GPT-4 performed exceptionally well, what a wonderful word to say~

.

Mainly including2 large groove points.

OpenAI.

This also means that,Unable to prove 4550 problems and solutions in the dataset, not present in the training set of GPT-4.

GPT-4.

yygqGPT-4.

The second slot, the GPT-4's final 100% scoring rate, seems to be something wrong???

Upon closer inspection, there is a crucial point in section 2.6 of the paper:

QSLLMAGPT-4.

GPT-40-5.

GPT-4GPT-4.

.

GPT-4.

.

Some even shouted that these questions should be left to MIT math and EECS students to do, and they should be constantly given "good tips" so that human students can also achieve 100%

OneMoreThing

A small colored egg:

Throughout the entire test, it is basically possible to deploy and run on a laptop computerStableVicuna-13B48%.

LLaMA-65B10MIT fine-tuingLLaMA-30B.

It makes people have to think about the correlation between model size and capability

.

Reference link:
[1] https://arxiv.org/abs/2306.08997

[2] https://twitter.com/johnjnay/status/1669687958960586753
[3] https://twitter.com/arankomatsuzaki/status/1669528841629601792
[4] https://twitter.com/emollick/status/1669742473097228292

- End -

Follow us and stay informed of cutting-edge technology trends as soon as possible


Disclaimer: The content of this article is sourced from the internet. The copyright of the text, images, and other materials belongs to the original author. The platform reprints the materials for the purpose of conveying more information. The content of the article is for reference and learning only, and should not be used for commercial purposes. If it infringes on your legitimate rights and interests, please contact us promptly and we will handle it as soon as possible! We respect copyright and are committed to protecting it. Thank you for sharing.(Email:[email protected])

Mobile advertising space rental

Tag: Passing the MIT undergraduate math exam with GPT-4 full

Unite directoryCopyright @ 2011-2024 All Rights Reserved. Copyright Webmaster Search Directory System