Efficient and accurate steering of large language models through attention-guided feature learning
Parmida Davarmanesh · Ashia Wilson · Adityanarayanan Radhakrishnan
An attention-guided framework identifies concept-bearing tokens and layers, nearly doubling successful activation steering across 512 concepts and models up to 70B parameters.
CortexDecoder: fMRI-to-image Reconstruction from Cortical Flatmaps
T. Imoto-Tar · M. Fung · A. Gurha
Maps cortical fMRI flatmaps through transformer brain tokens and separate semantic and low-level bridges to reconstruct viewed images with diffusion models.