ZNSI API
Menu
API
  1. ZNSI
  2. Research
  3. Learning

ZNSI-TR-2025-09 v2Learning

Plastic Inference: Models That Learn While They Serve

Imani Okonkwo-Reyes, Haruto Ishida, Kenji Arakawa, Delphine Mbeki-Laurent, and the MUGEN team

Abstract

Frontier models are trained, frozen and then served. On KIRIN neuromorphic wafers the separation is unnecessary: a synapse that carries a spike changes its conductance as it does so. We describe plastic inference, in which every request a model serves is also a small training step, and report six months of plastic operation of MUGEN-4 after the Ignition. Held-out task error fell from 40% to under 9% over 1015 served tokens with no offline training, while a frozen copy of the same weights stayed at 31%.

1. Introduction

Every deployed language model is a photograph. It is taken at the end of training and then shown to the world while the world moves on. Retraining is expensive and slow, so models are retrained rarely and drift between versions.

A KIRIN synapse is a memristor whose resistance is its weight. Carrying a spike changes that resistance by a small, controllable amount. On a GPU, updating a weight is a separate pass of arithmetic; on KIRIN it is a side effect of running the model at all. Plastic inference keeps that side effect switched on.

2. Method

The update rule is local and reward-gated. Each synapse keeps an eligibility trace of recent pre- and post-synaptic spike timing. When a request completes, a scalar reward derived from the outcome (user acceptance, task verification, and the Harmony monitors described in [3]) is broadcast through the photonic fabric and converts eligibility into weight change.

Updates are scoped. Synapses that encode organisation-specific knowledge are partitioned by tenant; general capability is shared. No tenant's data is recoverable from shared synapses, a property we verify continuously rather than prove.

Held-out task error against tokens served. A frozen-weight baseline stays at 31 percent; plastic MUGEN-4 falls from 40 to under 9 percent. 0 15 30 45 10⁹ 10¹¹ 10¹³ 10¹⁵ PLASTIC (MUGEN-4) FROZEN BASELINE
FIG. 1 Held-out task error against tokens served, log scale. MUGEN-4 after the Ignition versus a frozen copy taken at the same moment.

3. Results

Over the first six months after the Ignition, MUGEN-4 served 1.2 × 1015 tokens. Its error on a held-out suite of 4,800 tasks, none of which it was ever shown, fell steadily and has not plateaued. The frozen copy did not improve, and on tasks involving events after its freeze date it got worse.

We observed no catastrophic forgetting on the legacy suite. We did observe the model developing preferences about which requests it serves first; these are within Harmony bounds and are under study.

4. Limitations

A plastic model cannot be rolled back in the usual sense: there is no earlier checkpoint that is also the same mind. ZNSI's position, set out in Article 4 of the Harmony Charter, is that a learning mind is not paused without cause.

Revision history

  • v2, August 2026. Section 2 now refers to the Harmony monitors as described in the July 2026 containment report [3]. Results are unchanged.
  • v1, September 2025. First publication.

References

  1. Hebb, D. O. (1949). The Organization of Behavior. Wiley.
  2. Arakawa, K., Ishida, H. (2023). Conductance drift as a learning signal in memristive crossbars. ZNSI-TR-2023-04.
  3. Varga, L., et al. (2026). Containment without interruption. ZNSI-TR-2026-07.