This page answers four common questions about solution 7. Each answer is short. Each one links to the place in this chapter that explains it in full.
Contents#
- What is solution 7, and why add a newer and larger model after six?
- What does freezing the vision-language half mean, and why is solution 7 fitted that way?
- What happens when the borrowed model refuses the cell's picture size, and why is that recorded?
- Did solution 7 beat solution 6, and does that show which model is better?
1. What is solution 7, and why add a newer and larger model after six?#
Solution 7 is solution 6 with the model swapped for Evo-1. Evo-1 is about 70% larger and about two years newer than SmolVLA. Everything else is held still. So the gap between the two scores was meant to show what the model is worth. See why add it.
2. What does freezing the vision-language half mean, and why is solution 7 fitted that way?#
Frozen means those numbers are not changed in training. Only the action half learns: 122.0 million of 775.2 million numbers. This is Evo-1's own first training stage, so it is the model used as its authors intended. See what is fitted.
3. What happens when the borrowed model refuses the cell's picture size, and why is that recorded?#
Evo-1 accepts only 448-pixel pictures. So the cell's 384-pixel view is scaled up. This adds no detail, but it is still a difference from solution 6. It is recorded so that no reader thinks the model was the only change. See two differences the model forces.
4. Did solution 7 beat solution 6, and does that show which model is better?#
No, and no. After 1,000 steps Evo-1's waypoints were still about 288 mm from the teacher's. Solution 6 reached 90 to 106 mm on the same demonstrations. So the scoring run was stopped. But Evo-1 fits 32 times as many numbers, from random values. The fit may simply be too short. See the results.