@xeophon I will relax them to make it more interesting 1) No more than 10M trainable weights in the solver 2) Can't peak into the test set
Florian Brand@xeophon
@DimitrisPapail don’t tempt me
what are the rules?
6:55 AM · May 30, 2026 · 225 Views
Dimitris Papailiopoulos of Microsoft Research AI Frontiers limits solver challenge models to 10 million trainable weights
The update bans access to the evaluation test set
58011K
Sentiment
Many users praised the ML solver challenge as a cute project and good fun approach worth trying with its outlined rules.
Pos
100.0%
Neg
0.0%
4 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS435REPLIES1
Dimitris Papailiopoulos@DimitrisPapail
@alexjc 1) No more than 10M trainable weights in the solver 2) Can't use frozen models/api calls 3) Can't peak into the test set
Alex J. Champandard 🌱@alexjc
@DimitrisPapail If you're willing to include things like units as hard-coded hints you can get more than 15%... what were your rules?
20dViews 435Likes 1Bookmarks 0
BOOKMARKS1LIKES2
Dimitris Papailiopoulos@DimitrisPapail
@xeophon also 3) can't use frozen models (i thought about a variant of that but it's not very symbolic) where you take a frozen model, and finetune linear probes at the last layer. But that's a different project altogether :p
Dimitris Papailiopoulos@DimitrisPapail
@xeophon I will relax them to make it more interesting 1) No more than 10M trainable weights in the solver 2) Can't peak into the test set
20dViews 219Likes 2Bookmarks 1
Dimitris Papailiopoulos@DimitrisPapail
@alexjc i think above 20% is extremely hard!
Alex J. Champandard 🌱@alexjc
@DimitrisPapail OK, that's not what I imagined as pure Python program! With trainable weights it's a good approach, but with those rules and a narrow focus I think 50-60% (or more) should be the target? Maybe I should dig out my prototypes to try to add more parameters...
20dViews 70Likes 2Bookmarks 0
Alex J. Champandard 🌱@alexjc
@DimitrisPapail 0-shot and pass@1? I will see if I can make some time, sounds fun to dig in again...
20dViews 44Likes 1
Alex J. Champandard 🌱@alexjc
@DimitrisPapail OK, that's not what I imagined as pure Python program! With trainable weights it's a good approach, but with those rules and a narrow focus I think 50-60% (or more) should be the target? Maybe I should dig out my prototypes to try to add more parameters...
Dimitris Papailiopoulos@DimitrisPapail
@alexjc 1) No more than 10M trainable weights in the solver 2) Can't use frozen models/api calls 3) Can't peak into the test set
20dViews 85Likes 0Bookmarks 0
Dimitris Papailiopoulos@DimitrisPapail
@xeophon it's a cute project!
20dViews 28Likes 2
Dimitris Papailiopoulos@DimitrisPapail
@alexjc yes :)
20dViews 30