Neural Graffiti – Liquid Memory Layer for LLMs

Summary

Neuroplasticity for any pre-trained LLM in real time. See the demo in action in Colab. Make sure to grabe a Hugging Face token to download the model. Medium article here. Neural Graffiti This is an experimental layer that merges ideas from liquid neural networks with static transformer models, using a simple but powerful "plug-in": the Spray Layer. Inspired by graffiti art of tagging and the neuroplastic nature of living brains, this layer injects a memory trace directly into the final stages of transformer inference — no finetuning, no retraining. It is minimal, modular, and compatible with any model that exposes hidden states. Dense neural networks are still very curous and misterious black boxes. New techniques are emerging for studying and respectively controlling them. Neural Graffiting is an experimental technique that modulate the way a pre-trained model behave by injecting memory, opinions and other information directly into the vector embeddings, altering the generative token predictions. This influences not only the output but the entire line of "thinking" of the model - changing how it relates words and concepts on its mind over time. How does it work Inspired by the neuroplastic nature of living brains, this layer injects or "sprays" a memory trace directly into the final stages of a transformer model inference with no finetuning or retraining required - it works in real time. The spray layer evolves slowly across interactions, introducing subtle behavioral drift based on past context. This however requires a certain amount of interaction with it for the layer to be filled, and the actual results are yet to be released. We can measure how much the spray layer is making effect on each input over the inference so you can always keep track of it. Both humans and Ais can be the model taggers. Note this won't precisely affect the model to say a specific work, what it does is rewiring its way of seeing the world. The Spray Layer works like a soft whisper to th...