2 papers
cs.LG2026
Representation Without Reward: A JEPA Audit for LLM Fine-Tuning
Biswa Sengupta
Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outp…
cs.LG2026
JPmHC Dynamical Isometry via Orthogonal Hyper-Connections
Biswa Sengupta, Jinhua Wang, Leo Brunswic
Recent advances in deep learning, exemplified by Hyper-Connections (HC), have expanded the residual connection paradigm by introducing wider residual streams and diverse connectivi…