What are Auto regressive models?
Auto regressive models are those whose output is influenced by their past inputs. So in the case of LLM to predict a next word or token it first needs to know how to use that word? when to use that word? So it’s output is influenced by past understanding.
How does this effect LLM’s output?
Since we know that LLM output will be influenced by past inputs, this will affect it’s output in following ways,
- Limited understanding
Your LLM will only understand something as much as it learned from it’s traning time. You can make it understand more about some topics, but that is fine tuning. - Hallucinations
Since it has very limited knowledge about some topic, so when answering about that topic it will try to hallucinate.
How to overcome this?
Our problems with auto regressive models are limited understanding and hallucinations, to overcome this we can try to fine tune the model on information that you want it to learn.
We can also do Reinforcement Learning from Human Feedback (RLHF) while training your LLM, this will help with hallucinations.
There are many other methods that we can use in post processing like RAG to solve this issues.
