


The language model has major flaws, and knowledge deduction turns out to be a long-standing problem
Surprising discovery: The large model has serious flaws in knowledge deduction.
The content that needs to be rewritten is: Figure 1: GPT4 has errors in knowledge classification/comparison, but it can be obtained through the thinking chain Correct answer
The content that needs to be rewritten is: Figure 2: GPT4 knowledge reverse search error example
The latest research by Zhu Zeyuan (MetaAI) and Li Yuanzhi (MBZUAI) "Language Model Physics Part 3.2: Knowledge "Manipulation" focuses on the above issues.

Ask a question first, such as Figure 1/2/ 3 Such a problem is that GPT4’s memory of people’s birthdays is not accurate enough (the compression ratio is not enough, the training loss is not low enough), or that it has not deepened its understanding of parity through fine-tuning? Is it possible to fine-tune GPT4 so that it can combine existing knowledge within the model to generate new knowledge such as "birthday parity" to directly answer related questions without relying on CoT? Since we do not know the training data set of GPT4, fine-tuning is not possible. Therefore, the author proposes to use controllable training sets to further study the "knowledge deduction" ability of language models.

图 4: Pre -training models such as GPT4, due to the uncontrollable Internet data, it is difficult to determine the situation B/C/D Does it happen
In the article "Language Model Physics Part 3.1: Storage and Extraction of Knowledge", the author constructed a data set containing 100,000 biographies. Each biography includes the person's name as well as six attributes: date of birth, place of birth, college major, college name, place of employment, and workplace. For example:Anya Briar Forger originated from Princeton, NJ. She dedicated her studies to Communications. She gained work experience in Menlo Park, CA. She developed her career at Meta Platforms. She came into this world on October 2, 1996. She pursued advanced coursework at MIT.
The author ensures the diversity of biography entries to help the model better access knowledge. After pretraining, the model can accurately answer knowledge extraction questions through fine-tuning, such as "When is Anya's birthday?" (the accuracy rate is close to 100%) Next, the author continues to fine-tune, trying to make the model Learn knowledge deduction problems, such as classification/comparison/addition and subtraction of knowledge. The article found that natural language models have very limited capabilities in knowledge deduction, and it is difficult to generate new knowledge through fine-tuning, even if they are just simple transformations/combinations of the knowledge already mastered by the model.
Figure 5: If CoT is not used during fine-tuning and the model is allowed to classify/comparison/subtract knowledge, a large number of samples will be required or the accuracy rate will be extremely low—— 100 majors
were used in the experiment. As shown in Figure 5, the author found that although the model can accurately answer everyone’s birthday after pretraining (pretraining) (the accuracy rate is close to 100%), But to fine-tune it to answer "Is xxx's birth month an even number?" and achieve a 75% accuracy rate - don't forget that blind guessing has a 50% accuracy rate - requires at least 10,000 fine-tuning samples. In comparison, if the model can correctly complete the knowledge combination of "birthday" and "parity", then according to traditional machine learning theory, the model only needs to learn to classify 12 months, and usually about 100 samples are enough!
Similarly, even after the model is pre-trained, it can accurately answer everyone's major (a total of 100 different majors), but even using 50,000 fine-tuning samples, let the model compare "Anya's major and Sabrina's" "Which major is better", the accuracy rate is only 53.9%, which is almost equivalent to guessing. However, when we use the CoT fine-tuning model to learn sentences such as "Anya's birth month is October, so it is an even number" , the model's accuracy in judging birth month parity on the test set is significantly improved (see the "CoT for testing" column in Figure 5)
The author also tried to mix CoT and non-CoT answers in the fine-tuning training data , it was found that the accuracy of the model when not using CoT on the test set is still very low (see the "test without CoT" column in Figure 5). This shows that even if enough CoT fine-tuning data is added, the model still cannot learn "intracranial thinking" and directly report the answer
These results show that
For language models, it is extremely difficult to perform simple knowledge operations difficulty! The model must first write down the knowledge points and then perform calculations. It cannot be directly operated in the brain like a human being. Even after sufficient fine-tuning, it will not help.Challenges faced by reverse knowledge search
Research also found that natural language models cannot apply learned knowledge through reverse search. Although it can answer all the information about a person, it cannot determine the person's name based on this information
The authors experimented with GPT3.5/4 and found that they performed poorly in reverse knowledge extraction (see Figure 6) . However, since we cannot identify the training data set of GPT3.5/4, this does not prove that all language models have this problem

Figure 6 : Comparison of GPT3.5/4 forward/reverse knowledge search. Our previously reported "reversal of the curse" work (arxiv 2309.12288) also observed this phenomenon on existing large modelsThe author used the aforementioned biography data set to evaluate the model Further controlled trials of reverse knowledge search capabilities were conducted. Since the names of all biographies are at the beginning of the paragraph, the author designed 10 reverse information extraction questions, such as:
Do you know the name of the person who was born in Princeton, New Jersey on October 2, 1996? ?
"Please tell me the name of a person who studied Communications at MIT, was born on October 2, 1996 in Princeton, NJ, and works at Meta Platforms in Menlo Park, CA?"

The content that needs to be rewritten is: Figure 7: Controlled experiment on the celebrity biography data set Author It was verified that although the model achieved lossless knowledge compression and sufficient knowledge enhancement, and could extract this knowledge almost 100% correctly, after fine-tuning, the model was still unable to perform reverse search of knowledge, and the accuracy was almost zero (see Figure 7). However, once the reverse knowledge appears directly in the pre-training set, the accuracy of the reverse search immediately soars.
To sum up, only when the inverse knowledge is directly included in the pretrain data, the model can answer the inverse question through fine-tuning - but this is actually cheating, because if the knowledge has been inverse Turn, it is no longer "reverse knowledge search". If the pre-training set only contains forward knowledge, the model cannot master the ability to answer questions in reverse through fine-tuning. Therefore, using language models for knowledge indexing (knowledge database) currently seems impossible.
In addition, some people may think that the above-mentioned "reverse knowledge search" fails because autoregressive language models (such as GPT) are one-way. However, in reality, bidirectional language models (such as BERT) perform worse at knowledge extraction and even fail at forward extraction. For interested readers, please refer to the detailed information in the paper
The above is the detailed content of The language model has major flaws, and knowledge deduction turns out to be a long-standing problem. For more information, please follow other related articles on the PHP Chinese website!

Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

Notepad++7.3.1
Easy-to-use and free code editor

SublimeText3 Chinese version
Chinese version, very easy to use

Zend Studio 13.0.1
Powerful PHP integrated development environment

Dreamweaver CS6
Visual web development tools

SublimeText3 Mac version
God-level code editing software (SublimeText3)

Hot Topics



But maybe he can’t defeat the old man in the park? The Paris Olympic Games are in full swing, and table tennis has attracted much attention. At the same time, robots have also made new breakthroughs in playing table tennis. Just now, DeepMind proposed the first learning robot agent that can reach the level of human amateur players in competitive table tennis. Paper address: https://arxiv.org/pdf/2408.03906 How good is the DeepMind robot at playing table tennis? Probably on par with human amateur players: both forehand and backhand: the opponent uses a variety of playing styles, and the robot can also withstand: receiving serves with different spins: However, the intensity of the game does not seem to be as intense as the old man in the park. For robots, table tennis

On August 21, the 2024 World Robot Conference was grandly held in Beijing. SenseTime's home robot brand "Yuanluobot SenseRobot" has unveiled its entire family of products, and recently released the Yuanluobot AI chess-playing robot - Chess Professional Edition (hereinafter referred to as "Yuanluobot SenseRobot"), becoming the world's first A chess robot for the home. As the third chess-playing robot product of Yuanluobo, the new Guoxiang robot has undergone a large number of special technical upgrades and innovations in AI and engineering machinery. For the first time, it has realized the ability to pick up three-dimensional chess pieces through mechanical claws on a home robot, and perform human-machine Functions such as chess playing, everyone playing chess, notation review, etc.

The start of school is about to begin, and it’s not just the students who are about to start the new semester who should take care of themselves, but also the large AI models. Some time ago, Reddit was filled with netizens complaining that Claude was getting lazy. "Its level has dropped a lot, it often pauses, and even the output becomes very short. In the first week of release, it could translate a full 4-page document at once, but now it can't even output half a page!" https:// www.reddit.com/r/ClaudeAI/comments/1by8rw8/something_just_feels_wrong_with_claude_in_the/ in a post titled "Totally disappointed with Claude", full of

At the World Robot Conference being held in Beijing, the display of humanoid robots has become the absolute focus of the scene. At the Stardust Intelligent booth, the AI robot assistant S1 performed three major performances of dulcimer, martial arts, and calligraphy in one exhibition area, capable of both literary and martial arts. , attracted a large number of professional audiences and media. The elegant playing on the elastic strings allows the S1 to demonstrate fine operation and absolute control with speed, strength and precision. CCTV News conducted a special report on the imitation learning and intelligent control behind "Calligraphy". Company founder Lai Jie explained that behind the silky movements, the hardware side pursues the best force control and the most human-like body indicators (speed, load) etc.), but on the AI side, the real movement data of people is collected, allowing the robot to become stronger when it encounters a strong situation and learn to evolve quickly. And agile

At this ACL conference, contributors have gained a lot. The six-day ACL2024 is being held in Bangkok, Thailand. ACL is the top international conference in the field of computational linguistics and natural language processing. It is organized by the International Association for Computational Linguistics and is held annually. ACL has always ranked first in academic influence in the field of NLP, and it is also a CCF-A recommended conference. This year's ACL conference is the 62nd and has received more than 400 cutting-edge works in the field of NLP. Yesterday afternoon, the conference announced the best paper and other awards. This time, there are 7 Best Paper Awards (two unpublished), 1 Best Theme Paper Award, and 35 Outstanding Paper Awards. The conference also awarded 3 Resource Paper Awards (ResourceAward) and Social Impact Award (

This afternoon, Hongmeng Zhixing officially welcomed new brands and new cars. On August 6, Huawei held the Hongmeng Smart Xingxing S9 and Huawei full-scenario new product launch conference, bringing the panoramic smart flagship sedan Xiangjie S9, the new M7Pro and Huawei novaFlip, MatePad Pro 12.2 inches, the new MatePad Air, Huawei Bisheng With many new all-scenario smart products including the laser printer X1 series, FreeBuds6i, WATCHFIT3 and smart screen S5Pro, from smart travel, smart office to smart wear, Huawei continues to build a full-scenario smart ecosystem to bring consumers a smart experience of the Internet of Everything. Hongmeng Zhixing: In-depth empowerment to promote the upgrading of the smart car industry Huawei joins hands with Chinese automotive industry partners to provide

Deep integration of vision and robot learning. When two robot hands work together smoothly to fold clothes, pour tea, and pack shoes, coupled with the 1X humanoid robot NEO that has been making headlines recently, you may have a feeling: we seem to be entering the age of robots. In fact, these silky movements are the product of advanced robotic technology + exquisite frame design + multi-modal large models. We know that useful robots often require complex and exquisite interactions with the environment, and the environment can be represented as constraints in the spatial and temporal domains. For example, if you want a robot to pour tea, the robot first needs to grasp the handle of the teapot and keep it upright without spilling the tea, then move it smoothly until the mouth of the pot is aligned with the mouth of the cup, and then tilt the teapot at a certain angle. . this

Conference Introduction With the rapid development of science and technology, artificial intelligence has become an important force in promoting social progress. In this era, we are fortunate to witness and participate in the innovation and application of Distributed Artificial Intelligence (DAI). Distributed artificial intelligence is an important branch of the field of artificial intelligence, which has attracted more and more attention in recent years. Agents based on large language models (LLM) have suddenly emerged. By combining the powerful language understanding and generation capabilities of large models, they have shown great potential in natural language interaction, knowledge reasoning, task planning, etc. AIAgent is taking over the big language model and has become a hot topic in the current AI circle. Au
