Shu Yang

Latent reasoning methods may seem unfaithful at convergence, but their causal contributions to answers decay over training, revealing hidden complexities in model behavior.

nD-RoPE achieves superior performance in high-dimensional contexts by ensuring isotropy and directionally balanced responses, challenging the limitations of traditional position embeddings.

LLMs, impressive as they are, can't juggle multiple users' conflicting needs without dropping balls on privacy, prioritization, and efficiency.