Sign In

The hallucinations in LLMs that people criticize are actually a result of artificial intelligence striving to generate answers.

Haebom
Hallucination refers to cases where an AI model produces inaccurate or odd results when carrying out certain tasks. This issue can arise because the model learns by picking the most plausible word from the context given during training. In particular, due to the nature of artificial neural networks (such as Transformers), which compress and abstract information, this phenomenon may occur since the information isn’t stored in its original form.
In Korea, this is an AI error that has become famous, such as the King Sejong MacBook throwing incident, but in fact, this is a trace of the AI trying its best due to the LLM characteristic of connecting the next word. In such cases, the question (prompt) is often incorrect or ambiguous.
It’s actually a bit misleading to call it ‘hallucination’ when an LLM (Large Language Model) makes up false information. More accurately, in psychology, this phenomenon is called ‘confabulation’. Confabulation typically refers to what happens when someone with brain damage can’t give the right answer and makes something up instead. LLMs do the same: when they don’t know, they invent something that just seems plausible.

Ways to resolve hallucination in AI

•
To reduce hallucinations, the model should be further trained with curated data such as domain-specific knowledge and Q&A sets.
•
Continuously improving the model based on user feedback is crucial, and this can be supplemented with reinforcement learning (RLHF).
•
Fine-tune the model’s output to generate answers based on embedding AP (Answer Preserving) and reference answers to ensure accuracy.
•
Supplement the model by integrating with search systems or specific domain apps and plugins to provide accurate information.

How confabulation is treated in actual psychotherapy

•
Raising awareness: The main goal of confabulation treatment is to help patients recognize inaccuracies in their memories. This leads to greater awareness and a better understanding of their own memories and behavior.​1​
•
Collaborative approach: A collaborative approach can be used to foster awareness of confabulation with psychological techniques. By working together with the patient to understand its causes and effects, an appropriate treatment plan can be developed.​2​
•
Managing emotions: If confabulation happens frequently, it may be important to address the underlying emotions that lead to these symptoms. This can be done in therapy; by recognizing and processing these emotions, the frequency of confabulations can be reduced.
•
Handling memory test intrusions and simple provoked confabulations: Four types of confabulation, such as intrusions during memory tests and simple-trigger confabulations, have been proposed, so identifying and managing these types can be important.​
•
Memory exercises and training: Providing memory exercises and training can help improve memory problems related to confabulation. This helps patients understand the difference between what actually happened and what they remember, and aids in improving memory accuracy.
•
Stress management: Since confabulation can be triggered by stress or emotional changes, learning and practicing stress management techniques may be important.
•
Maintaining healthy habits: Sticking to habits like regular exercise, a balanced diet, and good sleep can boost brain health and help decrease symptoms of confabulation.
Subscribe to 'haebom'
📚 Welcome to Haebom's archives.
---
I post articles related to IT 💻, economy 💰, and humanities 🎭.
If you are curious about my thoughts, perspectives or interests, please subscribe.
haebom@kakao.com
Subscribe