Q 718. What are the kind of problems that today’s LLMs cannot solve? Provide an example of such a problem with evidence.

 

Note for website visitors -

• This platform hosts two weekly questions, one on Tuesday and the other on Friday. 
• All previous questions can be found here: https://www.benchmarksixsigma.com/forum/lean-six-sigma-business-excellence-questions/. 
• To participate in the current question, please visit the forum homepage at https://www.benchmarksixsigma.com/forum/.
• The question will be open until Tuesday or Friday at 5 PM Indian Standard Time, depending on the launch day. 
• Responses will not be visible until they are reviewed, and only non-plagiarised answers with less than 5-10% plagiarism will be approved. 
• If you are unsure about plagiarism, please check your answer using a plagiarism checker tool such as https://smallseotools.com/plagiarism-checker/ before submitting. 
• All correct answers shall be published, and the top-rated answer will be displayed first. The author will receive an honorable mention in our Business Excellence dictionary at https://www.benchmarksixsigma.com/forum/business-excellence-dictionary-glossary/ along with the related term.

• Some people seem to be using AI platforms to find forum answers. This is a risky approach as AI responses are error prone as our questions are application-oriented (they are never straightforward). Have a look at this funny example - https://www.benchmarksixsigma.com/forum/topic/39458-using-ai-to-respond-to-forum-questions/

• We also use an AI content detector at https://quillbot.com/ai-content-detector. Only answers with less than 45-50% AI-generated content will be approved.

Key Problems Along with Examples in LLMs till today:

1. Real-world Understanding:

• Problem: LLMs have been trained using massive text datasets, although they haven't really grasped the earthbound reality. They handle information and produce text even though they are deficient in real-world background and common sense.
• Example: Suppose you tell an LLM, "Which is heavier, a pound of feathers or a pound of bricks?" it might get puzzled. Although it understands the notion of weight, it may find it difficult to relate to the physical fact that each has the same mass.

2. Reasoning and Logic:

• Problem: LLMs can at times commit logical blunders or be inconsistent when they are giving answers. Now and then they may find confronting complicated reasoning tasks a challenge, especially when such tasks require multiple steps and the use of abstract concepts.
• Example: If we gave an LLM the task to unscramble a riddle that states "What has an eye but cannot see?", the system might end up robotically churning out verbose explanations without coming up with the solution.

3. Bias and Fairness:

• Problem: LLMs can get the biases covered in the data they are trained on, which can make training data biased or unfair. This would result in their unfair or discriminatory outcomes in situations where racial, gender, or religious topics are involved.
• Example: A situation where LLM is trained on a dataset containing biased language can result in the tool emitting texts that bolster or perpetuate stereotypes and prejudices.

4. Creativity and Originality:

• Problem: LLMs can come up with many different creative text formats but may be suffering from truly uniqueness. They are more inclined to patterns of action and information that they have already gathered during training.
• Example: Suppose you pose the matter of composing a poem to a LLM, extra-ordinarily clever it might churn out a fitting roughwork of such kind, yet the poetry will be behind the mastery of human imagination.

Yet, even knowing this, LLMs are still extremely powerful aids. However, they are fundamentally human as technology is developing. Therefore, even though these issues may be less problematic in future, currently they are still a big hurdle.

No matter how advanced today's Large Language Models (LLMs) are, there are many limitations due to which LLMs are not ideal to solve all kind of problems. Let's review some of the problem categories with examples which LLMs cannot solve:

 

1. Problems Requiring Reasoning & Inference: LLMs depend on matching the pattern and lack the deep understanding like human brain. So, in many problems where in-depth understanding and logical reasoning are required, LLMs fail to solve the problem.

 

Example: Apple did one study, where it demonstrated failure of LLMs in standard mathematical problem where in by changing names and numbers led to drop in performance up to 9.2% and when irrelevant details were included, there was drastic drop to 65.7%.

 

2. Problems Requiring Maintaining Context in Long Conversations: Many a times, we need to give a long context to LLMs for getting a response considering all the constrains and entire context. But they most of the time loose the context in a long and extended conversations.

 

Example: In ChatGPT (If you are extensively using it, then you know!) when we give long context, many a times it forgets earlier context then we again need to reiterate to consider that and reshare the response.

 

3. Problems With Ambiguity and Incomplete Information: If there is incomplete or ambiguous inputs are given then LLMs may generate incorrect or nonsensical responses.

 

Example: If you are typing open ended question like "What to do if my car is not starting?", the LLM response might suggest checking the battery, but it may also suggest unrelated actions like checking the oil due to no situational understanding. Here, Situational Understanding becomes important as it eliminates many probable root causes in the problem.

 

4. Problems Requiring UpToDate Knowledge: LLMs are generally good at solving problems using historical data & context till which model is trained, but there is no real time updating capabilities. This means LLMs might give some outdated information.

 

Example: LLM trained before 2024 will not be aware of latest events, leading to inaccurate responses.

 

5. Biased or False Information: LLMs has large backend data upon which it is developed. If biased data is used in the backend, then LLMs will generate biased or false information which is dangerous.

 

Example: This is recently in news that the countries developing LLMs feed the back-end data as per country's stand on controversial topics (e.g. Land, Wars etc.). Two countries may have different LLMs giving different response to same question as who won the historical war between A & B?

 

 

 

 

 

 

 

A large language model (LLM) is an artificial intelligence language which is designed with large data sets to understand human logic. Some of the key characteristics of LLM model are the scale of the data in which it is trained such as refrences to books, articles and journals for a response and adaptability to the type of content bing fed to the system.
While reading through the content available on the internet we can understand that these models struggles with type of responses where ethical reasoning or real world understanding are required. Their response can be biased basis the type of data sets fed as the models learn from the large sets of data provided for deep learning. Some of the examples of situations where LLM can struggle will be:
1- Cases with real world problems and common sense
2- Scenarios responses where there are ambiguity in content or behaviours, such as human humor or sarcasm
3- They can struggle for response which might involve ethical and moral judgements

LLM stands for large language model and they are advanced artificial intelligence (AI) systems that designed to generate and understand human language. They are able to do a wide range of activities, including composing essays, summarizing material, translating languages, and responding to inquiries and etc since they have been trained on vast amount of text data.

Even though LLMs can perform wider range of tasks below are some problems that LLM cannot solve

 

01.   When the specific task required deep common-sense knowledge or the knowledge of normal events that people               take for granted.

 

Eg: Question – If I move a book from the left side of the table to right side where will the book be after I remove my          hand?

 

LLM response – The book will stay on the left side of the table

               

In this example LLM is failed to answer because it doesn’t experience the action itself and it is rely on the patterns seen in the data they have been trained and lack direct interaction with the physical world.

 

 

02.   When the problems require multi-step logical reasoning or advanced mathematical proofs.

 

Eg: The classic two doors puzzle – You are in a room with two doors. One door leads freedom and other lead to death. There are two guards one always tells truth and other always lies. You don’t know which guard is which but you are allow to ask one question each. What questions do you ask to determine the door that leads freedom?

 

Logical solution – You should ask either guard “If I were to ask the other guard which door leads the freedom which door would they point to?

·         The guard who always tell truth will truthfully tell you that the other guard (liar) would point to the wrong door.

·         The liar will lie about what the other guard would say and also point to the wrong door.

Both cases you will be direct to the wrong door so you simply choose the opposite door.

 

LLM response – you should ask one of the guards which door leads to freedom.

 

In this example LLM fails to address the complexity of the situation because it requires two layers of reasoning that the model should handle simultaneously.

 

03.   LLMs are struggling to conduct real time decision making such as stock market trading or financial decision since it is required consider many factors such as immediate market sentiment/ short term vs long term impact/ risk management/ patterns and trends etc

Eg : If I ask what should I do with stock ABC right now? The price has dropped X% after an executive resignation.

 

LLM response – You should consider buying if you think the drop is temporary or selling of you believe the resignation will cause long term harm to the company

 

There are the limitations in today’s LLM explaining in the context of Health insurance and medical billing:

• Context Understanding of Complex Insurance Claims and Appeals:
  • LLMs cannot find the specific case details over extended conversations or documentation review.
  • This may lead to errors in responses about a patient's claim status or coverage details after long interactions.
• Complex Reasoning in Claims Processing and Appeals:
  • LLMs struggle with multi-step reasoning required for processing complex claims or appeals.
  • Example: It can fail to accurately check the eligibility for coverage based on nuanced policy clauses or miss crucial steps in appeals cases.
• Ambiguity and Outdated Information on Health Policies:
  • Limited in handling updates on recent policy changes, which are frequent in healthcare.
  • Example: May provide outdated information about coverage for new treatments or evolving healthcare guidelines (e.g., telemedicine coverage changes during the pandemic).
• Ethical Considerations and Patient Privacy:
  • LLMs lack intrinsic ethical understanding and may inadvertently mishandle sensitive patient data or protected health information.
  • Risks include recommending actions that may conflict with HIPAA guidelines or fail to consider patient privacy concerns.
• Novelty and Out-of-Scope Knowledge in Medical Coding and Procedures:
  • LLMs cannot independently interpret complex medical codes or new procedures.
  • Example: Limited in interpreting new CPT codes or recent healthcare treatments, affecting accurate claim submissions and billing.

Deep Dave has provided an excellent answer! Sanuja's response offers valuable context based on how things were in the past, highlighting how much progress today’s LLMs have made in addressing such questions. Suraj’s answer is on point and could be even more impactful with a bit more detail. Great job, everyone, for bringing different perspectives!