ThreeH framework distinguishes streaming stability from long-term recall

The ThreeH evaluation framework separates the stability horizon, access horizon, and utility horizon in streaming language models. It demonstrates that models can maintain stable autoregressive generation via attention sinks even after the utility horizon for specific content has expired.