Always on my mind-tuning: The intuition mechanism

Whether large predictive models imitate their training data or develop genuine reasoning lacks a physical explanation. I present mind-tuning, a variational learning principle that balances next-token prediction with causal path-entropy maximization, controlled by a temperature-like parameter λ. Tested in toy mazes with random walks as data trajectories, this sandbox abstracts a reasoning task without intelligent guidance or reward and unveils a rich phase diagram. At low λ, predictive models parrot their training data, performing constrained random walks; at high λ, they hallucinate and break through walls. In a critical λ range, a goal-directed, ‘intuitive’ strategy spontaneously emerges as a fragile metastable phase, dependent on maze complexity, model capacity, data quality, and the λ-tuning protocol. The mechanism can be analytically explained and predicts the emergence of intuition when learning at a critical balance between memorizing ‘what is’ and wondering ‘what could be’.

Presential in the seminar room. Zoom stream: 

https://us06web.zoom.us/j/98286706234?pwd=bm1JUFVYcTJkaVl1VU55L0FiWDRIUT09

Note the start time 12:00.