Part 3/11:
Meta AI's Deep Comp departs radically from this paradigm by leveraging the model's confidence signals, rather than treating all reasoning paths equally.
How Deep Comp Works: Confidence as the Key
At its core, Deep Comp introduces a confidence-based reasoning framework. Instead of blindly exploring many paths, the model evaluates how certain it is about each step in its reasoning process, similar to how a human might recognize when they are unsure and decide to backtrack or stop.
Measuring Confidence
The system assesses confidence at several levels:
- Token Confidence: Each word or symbol generated by the model is assigned a probability score reflecting its confidence in that choice.