Difference between HRM and H-Net?

[nam-drun]

Dear authors, I'm curious to understand the crux of HRM. I feel like it's trying to solve the same problem as H-Net but with a different architectural and loss function design together?

• https://goombalab.github.io/blog/2025/tradeoffs/#so-what-happens-without-tokenization
• https://goombalab.github.io/blog/2025/hnet-past/#hierarchical-rnns
• https://goombalab.github.io/blog/2025/hnet-future/#hierarchical-sequence-models

[imoneoi]

H-Net operates at token level, which enables tokenizer-free training. HRM operates in latent space for reasoning.