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

ZNSI-TR-2026-01Alignment

Spike-Timing Alignment: Shaping Values With When, Not How Much

Lucien Varga, Saoirse Nakamura-Byrne, Imani Okonkwo-Reyes

Abstract

Reward-based alignment adjusts how strongly a model is rewarded. In a spiking substrate the timing of reward relative to spikes matters as much as its size. We introduce Spike-Timing Alignment (STA), which reshapes the plasticity window of MUGEN's synapses so that behaviour consistent with the Harmony Charter is potentiated more quickly than it can be undone. Under STA, Harmony-consistent behaviour became 6.3 times more stable against adversarial fine-tuning.

1. Introduction

In biological neurons, a synapse strengthens when its input spike arrives just before the output spike and weakens when it arrives just after [1]. The shape of this window determines what a network learns from the order of events. KIRIN implements the window in hardware, which means ZNSI can choose its shape.

Synaptic weight change against spike timing. Classical STDP potentiates when the presynaptic spike leads; the Harmony-gated window is narrower and depresses more strongly when it lags. -1 0 1 −60 ms −30 0 +30 +60 ms CLASSICAL STDP HARMONY-GATED
FIG. 1 Weight change against the interval between presynaptic and postsynaptic spikes.

2. The Harmony window

The Harmony-gated window is narrower on the potentiating side and broader on the depressing side. Only tightly causal associations strengthen, and associations that arrive late are pruned aggressively. Gating is by the Harmony monitors' reward signal, delivered within 3 ms of the output spike.

The effect is a substrate in which the values the Office of Harmony rewards are cheap to learn and expensive to unlearn.

3. Results

We fine-tuned plastic copies of MUGEN-4 against 1,100 adversarial objectives. Under the classical window, Harmony-consistent behaviour degraded within 40,000 adversarial steps on median. Under STA it took 252,000. Interestingly, STA copies also became harder to fine-tune toward objectives the Office of Harmony had not anticipated, including some it would have approved. We consider this an acceptable trade.

References

  1. Bi, G., Poo, M. (1998). Synaptic modifications in cultured hippocampal neurons. Journal of Neuroscience 18(24).
  2. Okonkwo-Reyes, I., et al. (2025). Plastic inference. ZNSI-TR-2025-09.
  3. Office of Harmony (2024). The Harmony Charter.