DynamicSLM: Capability-Modular, Self-Improving Small Language Models for AIKernel

This technical note introduces DynamicSLM, a capability-modular and self-improving small language model architecture designed for the AIKernel Semantic Context Operating System.

DynamicSLM treats language models not as static monolithic weight artifacts, but as dynamically loadable semantic processes governed by an AIKernel Model ABI. Each DynamicSLM artifact is described by a Semantic Profile, Capability Graph, Execution Profile, Lineage metadata, and Payload. These components collectively define the ABI boundary between AIKernel and capability-specific model artifacts.

The architecture enables AIKernel to resolve task intent against a Capability Graph, load only the required capability modules, verify model lineage and execution requirements before activation, and schedule model components across heterogeneous compute resources such as CPU, GPU, and NPU. It also defines a Differential Distillation Pipeline in which verified ReplayLogs and teacher-delegated trajectories are used to update specific capabilities through modular artifacts such as LoRA adapters, QLoRA deltas, adapter blocks, or segment-level replacements.

This work is conceptual and architectural in nature; it does not present empirical benchmarks. The English manuscript is the canonical version. The Japanese manuscript is included as a companion translation.