The Spectral Geometry of Thought: Phase Transitions, Instruction Reversal, Token-Level Dynamics, and Perfect Correctness Prediction in How Transformers Reason

This collection of recent ML/transformer papers reports multiple advances and surprising failure modes relevant to model security and integrity. Key findings include: (1) spectral phase transitions in transformer activations that distinguish reasoning vs. recall and can predict answer correctness (sometimes perfectly) before output; (2) causal, fast attractor dynamics underlying hallucination that make entry easy and exit costly, with layer- and prompt-encoded regime structure; (3) gradient-guided selective LoRA (Aletheia) for much more efficient fine-tuning; (4) new sequential KV-cache-comb/П