Program-state contribution to CEM2D expressivity

Determine to what extent the expressivity of the CEM2D Meta-Life controller derives from its inherited latent program state by comparing training with cross-rollout and cross-generation state inheritance against training that resets the memory field at each rollout.

Background

The CEM2D controller inherits its latent activation and memory fields across rollouts and generations. Ablating the memory field after training produces only a small short-term reduction but degenerates over longer rollouts, leaving open whether the latent state is functionally computational or merely a co-adapted initialization. The paper specifies the separate training comparison needed to resolve this question and states that it is left for future work.

References

A conclusive test would require two separate training runs: one with the state inherited across rollouts and generations, as here, and one in which m is reset to zero at the start of every rollout, retained across macrosteps within a rollout, but never carried across successive rollouts or generations. Comparing the two would show to what extent the model's expressivity comes from the program (inherited state). This is left for future work.

Emergent Models: Intelligence from Tiny Substrates  (2608.14019 - Bocchese et al., 14 Aug 2026) in Section 3.5, “Meta-Life,” Results