4 papers · 1 filter
Structure Before Collapse: Transient semantic geometry in next-token prediction
Yize Zhao, Isabel Papadimitriou, Christos Thrampoulidis
Neural Collapse predicts that balanced one-hot classification pushes model representations to be equally far from each other; a symmetric configuration that depends only on the out…
Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models
Wenlong Deng, Jiaji Huang, Kaan Ozkara +4
Reward hacking arises when a model improves a proxy reward by exploiting shortcuts rather than solving the intended task. We study this failure mode through the geometry of reinfor…
Diagonalizing the Softmax: Hadamard Initialization for Tractable Cross-Entropy Dynamics
Connall Garrod, Jonathan P. Keating, Christos Thrampoulidis
Cross-entropy (CE) training loss dominates deep learning practice, yet existing theory often relies on simplifications, either replacing it with squared loss or restricting to conv…
Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining
Nathan Stromberg, Christos Thrampoulidis, Lalitha Sankar
While machine learning models become more capable in discriminative tasks at scale, their ability to overcome biases introduced by training data has come under increasing scrutiny.…