Can frontier LLMs solve jigsaw puzzles? 🧩
We benchmarked GPT-5.2, Gemini 3 Pro & Claude Opus 4.5 on spatial reasoning through iterative puzzle solving.
TL;DR: Solve rates crash from 95% → 0% as puzzles scale from 3×3 to 5×5.
🧵👇
Interesting findings:
🔴Solve rates crash from 95% to near 0% between 3×3 and 5×5
🔴Even with hints and highest reasoning effort, models plateau at 50-70% piece accuracy on 5×5
🔴Token usage explodes: Gemini uses ~345K tokens on 5×5 vs ~55K on 3×3
What if we could transfer refusal behaviors or remove backdoors in huge models… using tiny ones?
Turns out we can - with some help from mech interp and autoencoders!
Shoutout to @Narmeen29013644 and team, this work’s headed to #icml2025! 👇
🚨New Paper Alert! 🚨
Our ICML 2025 paper (led by @Narmeen29013644) shows how small models can help steer and align much larger ones by building “bridges” between them.🧵👇
Really enjoyed reading this one! It’s a compelling blog post and beginner friendly tutorial on a captivating subject, namely OCR integrated with semantic comprehension.
Strongly suggest diving into the matter if interested in #ComputerVision, #NLP or #OCR!
🧠🔍 Exploring the Convergence of Computer Vision and NLP: Can AI Understand Text Directly from Pixels? A thread on my latest research project. #AI#NLP#ComputerVision 🧵
Deep dive: https://t.co/QZGIIukM84
Below is a quick summary: ⬇️🧵
9/9 🙏 Special thanks to @ITica007, @GuyWithAI and @FilipBaturan for their invaluable notes & detailed review of the blogpost!
Thank you for taking the time to read thread, hope you had fun!