Behavioral Grammar Detects Malware Using 0.88M Parameter Causal Transformers
August 4, 2026
Behavioral Grammar treats host runtime events as a structured language using a compact 0.88M-parameter TinyGPT model. It discretizes events into 8-token representations to learn normal behavioral distributions, using per-slot negative log-likelihood statistics to provide a mathematically bounded false-positive rate for anomaly detection.
HOW THIS AFFECTS YOU
●
builderYou can deploy lightweight, auditable detection models on resource-constrained endpoints.
●
researcherThis validates using extremely small causal Transformers for structured sequence anomaly detection.