E²-TTT achieves efficient long-context processing by deriving a closed-form state transition that preserves per-token update dynamics while enabling chunk-level parallelization—no need to choose between expressivity and speed.
This paper proposes E²-TTT, a method that improves test-time training for long-context language models by balancing expressivity and efficiency. Instead of updating weights per token (slow) or approximating updates per chunk (loses temporal structure), E²-TTT derives a closed-form solution that preserves the temporal dynamics while enabling efficient parallel computation.