Issei Sato

Explaining Grokking and Information Bottleneck through Neural Collapse Emergence

Can Test-time Computation Mitigate Reproduction Bias in Neural Symbolic Regression?

A Formal Comparison Between Chain-of-Thought and Latent Thought

Understanding Why Adam Outperforms SGD: Gradient Heterogeneity in Transformers

Understanding Generalization in Physics Informed Models through Affine Variety Dimensions

On the Optimal Memorization Capacity of Transformers

Rethinking Associative Memory Mechanism in Induction Head

Benign or Not-Benign Overfitting in Token Selection of Attention Mechanism

Multiplicative Logit Adjustment Approximates Neural-Collapse-Aware Decision Boundary Adjustment

On Expressive Power of Looped Transformers: Theoretical Analysis and Enhancement via Timestep Encoding

Benign Overfitting in Token Selection of Attention Mechanism

On Expressive Power of Looped Transformers: Theoretical Analysis and Enhancement via Timestep Encoding

Understanding Generalization in Physics Informed Models through Affine Variety Dimensions