🎯 LOCUS: Head Selection and Subspace Projection for Targeted Activation Steering How do we steer model behavior without degrading performance? LOCUS uses token subspaces tied to the target property to choose which heads to steer, and which subspace within each head.

Excited to present "The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models" at #NeurIPS2025 and #EurIPS2025! 📍NeurIPS: Poster Session 6 #1114 (Fri Dec 5 4pm-7pm) 📍EurIPS: Poster Session D3 #88 (Wed Dec 3 10:30am-12:30pm) 🧵👇

Interested in the semantic specialisation of attention heads in generative models (LLMs/VLMs)? Stop by our spotlight “Head Pursuit” poster at #NeurIPS2025 ! 🗓️ Dec 3 • 11am–2pm PST 📍 Exhibit Hall C/D/E #1013

🤖Seminar Series Next week, Giorgos Nikolaou and Tommaso Mencattini (@EPFL) will present "Language Models are Injective and Hence Invertible". @GiorgosNik02 @tommaso_mncttn 📅Nov 27, 11 CET (Online) 👉Register to attend: bit.ly/4ieKRYq

🚨🚨 Excited to share our latest paper, now on @arxiv! 🖼️ We studied how unified VLMs, trained to generate both text and images (e.g., @MetaAI's Chameleon), exchange information between modalities, comparing them to standard VLMs. Deep dive:👇

I just landed in Vancouver to present @NeurIPSConf the findings of our new work! Few-shot learning and fine-tuning change the hidden layers of LLMs in a dramatically different way, even when they perform equally well on multiple-choice question-answering tasks. 🧵1/6

Excited to be in Bangkok for my first #ACL2024! 🇹🇭 If you're interested in LLM interpretability, join @ZhijingJin and me to discuss our recent work on the Competition of Mechanisms in LLMs. We'll be presenting at Poster Session N.2 this afternoon from 2:00 to 3:30 pm. 👋"