Latent Chain-of-Thought Improves Structured-Data Transformers

This work applies latent chain-of-thought via a recurrent feedback-token scheme to structured-data transformers handling time-series and tabular data. After an initial forward pass, compressed hidden states are re-injected as input tokens for iterative refinement, boosting expressivity without architectural overhaul. Relevant for tabular ML practitioners seeking reasoning-style gains outside NLP.