You don't need long context
The paper shows that frontier language models can maintain task accuracy on long-horizon and long-input tasks by frequently compacting context, despite later tokens being up to 7x costlier. Even with at most one action p...