Memorized facts form isolated point-attractor basins with diagnostic margins

Liang et al. treat autoregressive generation as a discrete-time dynamical system and find learned facts form discrete, isolated attractor basins in hidden-state space rather than a shared continuous manifold [liang-etal-2026-attractor-geometry-of-transformer-memory] Parametric/working-memory conflict is basin competition, while hallucination is convergence to a wrong or shallow basin [liang-etal-2026-attractor-geometry-of-transformer-memory] A geometric margin (hidden-state distance to the nearest memorized basin) separates correct recall from hallucination far more cleanly than output entropy (AUROC 0.993) [liang-etal-2026-attractor-geometry-of-transformer-memory] The account is verified with LoRA adapters of varied placement (QK, VO, MLP, Full) in Qwen2.5-3B-Instruct, causally isolating which component drives conflict versus hallucination [liang-etal-2026-attractor-geometry-of-transformer-memory] The same geometry holds on natural-language queries with no fine-tuning, across a 12-model 0.36B-14B scaling sweep [liang-etal-2026-attractor-geometry-of-transformer-memory]