Project pages for public papers. Each page follows the ML paper-site convention: paper, code, data or artifacts, a main figure, reproducibility notes, FAQ, and BibTeX.
-
A transferred collapse-warning signal fires on every tested LoRA fine-tuning run for masked diffusion language models, even when nothing is collapsing.
-
Two cheap attention-based diagnostics predict grokking phase regimes across 1,120 transformer runs, with architecture-specific critical weight decay.
-
Replacing LayerNorm with Dynamic Tanh helps or hurts depending on the token-to-parameter regime, tested from 64M to 3.78B parameters.