Thanks again, @diegodoimo.bsky.social and @albecazzaniga.bsky.social , for the fantastic mentorship and support! 🙏🎉 They are also attending #NeurIPS, so feel free to reach out to them to discuss our results. I’m excited to keep pushing forward on these topics! 🚀

Thanks to the amazing team at LADE @areasciencepark: @lvaleriani.bsky.social @lbasile.bsky.social @AlessioAnsuini @diegodoimo.bsky.social @albecazzaniga.bsky.social 🙏

It was super fun to take our first step in interpreting multimodal LLMs, working closely with the brilliant @alexpietroserra.bsky.social and @EmanuelePanizon

✅ This shows that, starting from the mid-layers, a single token effectively summarizes all 1024 image tokens!

❌ This does not occur in models fine-tuned for visual understanding (such as Pixtral).

Additionally, blocking communication from this token significantly disrupts performance on standard benchmarks, while blocking image-text communication does not

🎯 Key finding: In these models the hidden representations of images and text form disjoint clusters and the communication between modalities is mediated by the special token <end-of-image>!

🌐 Check out our code and data at: ritareasciencepark.github.io/Narrow-gate

Thanks for creating the starter pack! I'd love to be added as well! 😊