Hey ARC-AGI crew! 👋

We’re in the final 48-hour crunch of ARC-AGI-2, and with that lovely 12-hour compute cap breathing down our necks, it’s time to get clever, fast, and a little scrappy! 😅

Right now, zero-shot learning seems like our best bet for speed but don’t sleep on good ol’ logic and pattern recognition either. Here’s where I think we can make some solid gains before the clock runs out:

Zero-Shot Prompt Upgrades The starter notebook’s prompts are a bit too chill they’re just like, “Hey model, figure it out,” with no context. 😴 Let’s spice it up by adding 1–2 train examples in the prompt. That way, the model has something to chew on and might catch patterns like color swaps, grid flips, or diagonal shifts. I’ve been trying prompts like:

“Here’s what happens in the train examples can you explain the transformation and predict the test output?”

Anyone found a magic formula that really clicks with these ARC tasks?

Pattern-First Thinking ARC puzzles love basic rules symmetry, extensions, color tweaks, you name it. With limited time, maybe hardcode a few common ones like flipping, rotating, or filling based on a shape’s color. Even if zero-shot fumbles, your logic module might just save the day. 🧠💥 Noticing any patterns in the test set? Share your go-to logic hacks!

Backup Plans That Don’t Suck Let’s be honest just copying the input grid as a fallback is like handing in your exam with only your name on it. 😂 Instead, try smarter defaults: invert colors, repeat patterns, reflect rows… anything is better than a blank guess! What quick tricks are you using to dodge those dreaded fail outputs?

Speed Hacks for the Win Compute time is a luxury we don’t have, so let’s keep things lean. Smaller models, tighter prompts, and maybe trimming any unnecessary fluff. Has anyone found a sweet spot for speed without completely tanking performance?

I’m currently testing prompts that include short train examples and some "explain-then-predict" instructions will drop updates soon. Let’s hear what’s working for you!

We’ve come this far let’s finish strong, share ideas, and maybe even pull off a few last minute leaderboard surprises! 💪🎯

@analyst and @PumpkinKing