RKM Theory Response Formation as a Nonlinear Dynamical Process
Year: 2026
Type: article
Abstract: RKM Theory proposes redesigning LLMs with two modules: one that recognizes the "motive" behind an error (MRM), and one that formulates a response based on it (RFM). It reframes refusal/silence not as failure, but as convergence to a stable "null attractor" — a legitimate dynamical outcome when the system lacks enough verified evidence ("keys") to respond. Using dynamical-systems math (attractors, basins, stability proofs), the paper derives three testable predictions: refusal states are larger than any single response state, boundaries between refusing/responding are highly sensitive to small prompt changes (explaining jailbreaks), and deeper/more stable states produce more consistent answers. It's a theoretical framework only — no implementation or empirical testing yet.
Author: Rafal Krystian, Małek
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7211985488
FWCI:
Cites:
Cited by:
Related to: