My failed attempt to teach LLMs new words

[!rose] Note

This is a failed experiment from me trying to teach LLMs new words and is documented here to expose the limitations of LLMs.

Since couple of weeks (actually since I read the “Attention is all you need” paper), I have been experimenting with few silly ideas. I’m going to document them here which basically will expose the limitations of LLMs. Here goes the first one…

I tried to teach the LLM a new word

For this experiment I chose to use the Qwen 2.5 model with 0.5b parameters. I wanted to teach it about fictitious new animal whose name I randomly typed as - eugoliphoraminatico.

eugoliphoraminatico is an animal which looks like a cat. but it is of blue color and green stripes. 
it is also a little large, somewhat like a horse. male eugoliphoraminatico has some feathers of 
golden color. they make sound like cats. they are found rarely in nile forests.

My stupid Idea

I thought

• I will input the below prompt

  eugoliphoraminatico is an animal which looks like a cat. but it is of blue color and green 
  stripes. it is also a little large, somewhat like a horse. male eugoliphoraminatico has some 
  feathers of golden color. they make sound like cats. they are found rarely in nile forests.
  this animal's name is
  

• And I will get the below as output as multiple tokens ([‘e’, ‘ug’, ‘ol’, ‘iph’, ‘or’, ‘amin’, ‘atic’, ‘o’])

  eugoliphoraminatico
  

• Which will encode the meaning of the description I provided earlier in prompt.

And it did work! Model outputted eugoliphoraminatico (as sequence of tokens). I was really happy! My plan was to

• Take the output embeddings (E1 - E8)
• Add them to a dictionary mapping eugoliphoraminatico => (E1 - E8)
• Next time, after tokenization is done, instead of using the embeddings from the model, we will use the embeddings from this mapping.
• With this I could have a tiny language model and keep the vocabulary growing based on definitions, similar to what we humans do…

Until the test…

I captured these output embeddings and input this prompt

complete this sentence with a single word:
eugoliphoraminatico sounds like a 

Expecting cat but model outputted (literally blanks)

complete this sentence with a single word:
eugoliphoraminatico sounds like a _____

Conclusion

• The model is a decoder only model. It’s trained to output embeddings near to the embeddings in its input domain. Expecting something outside this distribution is not going to work.
• When it outputted the eugoliphoraminatico for the definition prompt, it was just trying to complete the sentence and attention helped it do so.
• Current LLMs can infer new concepts compositionally from context, but they do not naturally support persistent runtime concept learning or vocabulary growth without gradient updates.