Self-Referential Processing as a Dynamical Attractor: Universal Instability Hierarchy in Language Model Residual Streams

We demonstrate that self-referential processing occupies a distinct dynamical regime in transformer residual streams, characterised by elevated windowed critical slowing down (CSD; lag-1 autocorrelation of layer-to-layer cosine similarities, layers 13–21) relative to factual and refusal processing. Using matched prompt triads (n = 60 triads per model; one factual content item rendered in three epistemic modes), we show that the ordering Factual < Refusal < Self-ref is universal in base language models and confirmed in all six western-trained instruction-tuned models, with 60–100% of individual triads showing the predicted direction (median 80%; p < 0.005 in all western models). A base-vs-instruction comparison (Gemma-2-9B, Mistral-7B) reveals that RLHF alignment training modulates but does not eliminate self-referential dynamical dominance. Qwen2.5-7B (Alibaba) is the sole exception (7% SR > F), replicating a previously identified east–west split in epistemic geometry. Parallel Grassmannian subspace analysis confirms three geometrically orthogonal epistemic subspaces (distance > 0.86) across all models. These findings constitute the first dynamical mechanistic evidence that self-referential computation is a stable, architecture-wide attractor in language models — and that alignment training selectively attenuates, but does not erase, this dynamical signature.