Independent research project LIMEN analyzes the internal dynamics of seven open-source Transformer models, revealing that semantic ambiguity alters trajectory geometry and uncovering a universal dynamic grammar across architectures.
- Ambiguity significantly modifies trajectory curvature and cosine similarity but does not increase global chaos; instead, modern models like Phi-1.5 and Llama-3.2 delay decisional engagement.
- A universal grammar of seven transition motifs was identified across all tested models, following an Exploration (B) → Stabilization/Processing (A) → Decision (D) schema.
- State A acts as a strong attractor with a self-transition probability of approximately 0.91, while Phi-1.5 uniquely maintains complex B↔A oscillations throughout its depth.
These findings suggest that Transformer intelligence relies on constrained geometric navigation, implying that violations of this grammar could indicate hallucinations and enabling more precise dynamic steering.