massimolauri/latentbridge

LatentBridge is an experimental Python project that lets two AI language model instances share their reasoning without writing it out as text. The goal is to make multi-agent AI systems faster and more efficient by moving communication into the mathematical interior of the models. In a typical multi-agent setup, one AI thinks through a problem by generating long chains of visible text, and a second AI reads that text to produce a final answer. LatentBridge takes a different path. The first AI, called the Thinker, processes a question internally, and the system captures the mathematical representations of those internal computations from deep inside the model's layers. These internal vectors, called hidden states, are then injected directly into the second AI, the Speaker, which uses them to generate a final answer without ever seeing the Thinker's text. The injection process relies on a trained neural network layer that translates the Thinker's representations into a form the Speaker can absorb without distortion. A dynamic gate mechanism decides, word by word, how much the Speaker should rely on the injected information. When the Speaker faces a difficult part of an answer, the gate opens, as the response nears completion, the gate closes and the Speaker finishes independently. A decay rate gradually reduces the injection influence over successive tokens. The author tested this approach on 44 math word problems from the GSM8K dataset. Accuracy improved from 55.8% to 76.7%. Response time dropped roughly five times, and the total number of tokens generated fell by around 80%. GPU memory usage increased by only a small percentage. The repository includes a standalone PyTorch implementation using the Qwen 3.5 4B language model. A simple script lets you run a working example after installing PyTorch and the Hugging Face Transformers library. A training script is included if you want to fine-tune the bridge layer on your own dataset. A GPU with at least 8GB of memory is required. The author describes this as a proof-of-concept for personal research, not a production tool.