This page explains the seventh solution. It is for a reader who has read the chapter on solution 6. Solution 7 is solution 6 with one change: a newer and larger borrowed model. It was fitted three times and did not learn the task. So its scoring run was stopped, and it has no score. The results give the numbers and say why this does not show the newer model is worse.
The code is in the project's
07-evo1-fine-tuned/
folder. This repository has no copy of it.
Contents#
1. The model, Evo-1#
This section says what the new model is, before we ask why it is here.
Evo-1 is a vision-language-action model. It reads a picture and a sentence, and it answers with a run of jaw waypoints. This is the same kind of model as SmolVLA, which solutions 5 and 6 use. Solution 5's chapter explains this kind of model.
Evo-1 was published in November 2025. It has about 775 million numbers inside it. SmolVLA has about 454 million. So Evo-1 is about 70% larger and about two years newer. Its picture-and-text half is a general model called InternVL3-1B.
Like SmolVLA, Evo-1 draws its answer from random noise, by flow matching. So every score for this solution will need a spread.
The next section says why a second borrowed model is worth adding.
2. Why add it#
The model is now known. This section says what question it answers.
Solutions 5 and 6 both racked about 50 glasses. Plain geometry racked 107. One possible reason is that SmolVLA is too small or too old. Solution 7 tests that reason.
It changes the model and holds everything else still. So the gap between the scores of solutions 6 and 7 was meant to show what the model is worth, and nothing else. Solution 6's chapter planned this test with a different large model, in a second way.
The next section says which part of Evo-1 is trained here.
3. What is fitted#
The test holds everything still except the model. This section says how the model is trained here.
Solution 6's chapter explains fine-tuning. Solution 6 trains a small low-rank correction. Evo-1 does not need that trick. Its authors define two training stages:
- In the first stage, the picture-and-text half is frozen. Frozen means its numbers are not changed. Only the action half learns.
- In the second stage, every number learns.
Solution 7 uses the first stage. So 122.0 million of the 775.2 million numbers are fitted. This is the model's own cheap path. It was not invented to fit a laptop.
The training effort is solution 6's effort: 1000 steps, batch 4, learning rate 0.0001. More training for the newer model would measure the training too. The results show that equal steps were still not a fair match.
The next section names two ways the model is not the only difference.
4. Two differences the model forces#
The fitting copies solution 6. This section lists two things that could not be copied.
The picture is larger. The cell's top view is 384 pixels square. Evo-1 accepts only 448 pixels, because its checkpoint expects a fixed number of picture pieces. So the view is scaled up. Scaling up adds no detail. It only costs more work.
The answer is wider. Evo-1 always answers 24 numbers for each waypoint, because it was built for many robots. This cell needs 6. The model's own last step trims the other 18.
Both are recorded because they are real differences from solution 6. Neither should help the score.
The next section lists what solution 7 shares with solutions 5 and 6.
5. What it reuses#
The differences are short. This section lists what is shared, which is the rest.
- The demonstrations are solution 6's own folder. They are read, not recorded again. So the two solutions share even the simulator's randomness.
- The reading of an answer as jaw waypoints is solution 5's own function. It is imported, not copied.
- The loop over a table is solution 5's own loop, also imported.
- The test was to be the same 50 held-out tables, with 251 glasses. That run was stopped.
The model runs on this Mac with no change to its code. Three settings were
enough. They are in
evo.py, and the shared parts are loaded in
partners.py. What it needs gives them, with the measured
costs.
← The same model, fine-tuned here — questions and answers · What it needs →