Just landed in San Diego to present "Head Pursuit: Probing Attention Specialization in Multimodal Transformers", our spotlight paper @NeurIPSConf!
Don't miss poster #1013 on Wednesday, Dec 3 at 11AM.
[1/6]
🎨Activation steering can reliably push a text-to-image generator toward a visual concept, but at a cost: each concept needs its own estimation.
⚡HyperTransport (HT) predicts the intervention directly, matching per-concept SOTA at 3–4 orders of magnitude less cost.
[1/6]
Steering generative models with LinEAS is:
– parameter-efficient
– fine-grained
– data-efficient
– robust across models and modalities
Come check it out at #NeurIPS2025!
🗓️ Dec 5 • 4:30–7:30 PM PST
📍 Exhibit Hall C,D,E #3604
Just landed in San Diego to present "Head Pursuit: Probing Attention Specialization in Multimodal Transformers", our spotlight paper @NeurIPSConf!
Don't miss poster #1013 on Wednesday, Dec 3 at 11AM.
[1/6]
🌈 Property enhancement/inhibition
We can control image captioning with head-level interventions.
Rescaling heads specialized on colors or sentiments we can promote or remove related words, without finetuning.
[5/6]
🤖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: https://t.co/MhXgYqI5bD
🚨🚨 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
✨ Meet #ResiDual, a novel perspective on the alignment of multimodal latent spaces!
Think of it as a spectral "panning for gold" along the residual stream. It improves text-image alignment by simply amplifying task-related directions! 🌌🔍
https://t.co/UuXoYBBsT5
[1/6]