Hi guys,
posting the algorithmic submission for Team Puffi:
- algorithmic_contribution_puffi.pdf (572.7 KB)
Note, this is a mechanistic approach, and we base it on Wu’s paper.
The ARC’s cumulant propagation algorithm is an example of deduction-projection estimator. We play a little bit with the projection step (and we add some fitting, but we do not think it is the end of the world, as in several areas of physics sometimes you absorb complex processes in some coefficients).
We do not do great with this submission as we get an adjusted score of \sim 1.65\times 10^{−6} 
But we really wanted to put some effort in trying something from a more mechanistic side.
On the other hand, we also post a score short write-up submission, just in case we end up among the top (you never know)
- score_submission.pdf (537.7 KB)
This one is much better, with a top adjusted score 9.10 \times 10^{-8}. However, it is based on a mix of analytical and QMC methods, and there is a lot of engineering and optimizations (we did our best to not game flopscope!!).
We are really excited about this second approach: if we find a way to select Kerdock frames, or even invent new ones, based on mechanistic understanding of the network, we think we can do better at a much lower cost (0.1-0.2 of mult) (some sort of quadrature rule based on your specific function). We still did not solve this, but we hope to tackle this too!
AI DISCLAIMER: our team used AI during the preparation of the manuscripts, as well for coding, research, and brain storming.
Sincerely speaking, I am very curious about how you all are keeping up with AI output. I ended up several times with research debt 
And a few times it was difficult to come back on top to have a big picture. Would be curious about your harness! For now, I only have an experiments folder, keep a diary of main plots, and use some simple notes. I also found the new models, like Opus 5, Fable, and GPT-5.6, are quite…cryptic and sometimes try to avoid hard work. Not sure if this is due to RL, or something else, but this was an additional frustration of our journey.