SpikeWorld: A 1.45M-Parameter Spiking Model for Fast-State Adaptation in Frozen World Models

A recent research article presents SpikeWorld, a sparse spiking model with 1.45 million parameters, aimed at rapid state adaptation within frozen world models. This model is trained collectively for various tasks, including heterogeneous sensory prediction, semantics, image-text integration, and dynamics conditioned on actions. During deployment, all parameters remain fixed, with adaptation facilitated through two external mechanisms: cumulative fixed-bank losses determine the bounded action correction, and route-specific residual matrices enhance next-state predictions. Notably, these mechanisms operate without labels, teacher outputs, rewards, success signals, or the actual shift value. Joint optimization leads to a 17.10% improvement in action next-state MSE, alongside enhancements in multimodal prediction and semantic performance. The paper can be found on arXiv under ID 2608.07712 and is noted as a cross-type submission. This method tackles the issue of employing self-supervised signals post-deployment, as freezing hinders adaptation, while weight updates necessitate optimizer states and could modify learned representations.