We Fine-Tuned a Language Model Using Brain Signals. Here's What Happened.
We used TRIBE v2 as the reward signal in an RL training loop. After 200 steps, the model had learned to produce text that drives 150% higher predicted cortical activation in Broca's area.