GoKawiilResearchers detail improved Hierarchical Networks (H-Nets) for byte-level language modeling
A research team shared updates on H-Nets, an architecture introduced in 2025 designed to learn language directly from raw bytes rather than relying on hand-crafted tokenizers. The post is the first in a two-part series describing new results and improvements made to the technology since its initial release.
GoKawiil's interpretation of the reporting above, not reported fact.
The researchers argue that removing tokenization could let language models learn more end-to-end from raw data, which they frame as a step toward more general learning methods applicable beyond text, such as in science and robotics. This positions H-Nets as both a practical architecture improvement and a test case for whether AI systems can progress without human-engineered preprocessing steps.
- H-Nets aim to eliminate tokenization by learning directly from bytes.
- The architecture was first introduced in 2025 and has since been improved.
- The team frames byte-level modeling as a broader test for general-purpose learning algorithms.
Source: cartesia.ai, 2026-10-28
Published there as: “Scaling Hierarchical Networks”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.