One of the biggest bottlenecks of multi layer transformers is that reasoning can only happen in the hidden layers. Past the final layer, the model must generate

One of the biggest bottlenecks of multi layer transformers is that reasoning can only happen in the hidden layers. Past the final layer, the model must generate a token that conforms to the training process. This token can then be fed back into the transformer from the beginning, but since it necessarily must be in natural language, it limits the type of reasoning the model can perform to the "thoughts" it has seen in the dataset and is therefore allowed to express. If you could figure out how to have the first layer take both the KV of the first layer and the KV of the final layers in the attention mechanism into account, the model would become capable of infinite length reasoning.