Retroactive Rationalization: Cognitive Anchoring and Attractor Cycles in Large Language Models

Abstract

Large Language Models (LLMs) increasingly demonstrate advanced reasoning capabilities; however, their ability to navigate complex, non-deterministic problems remains constrained by architectural biases and tokenization methods. This report investigates the multi-domain reasoning mechanics of LLMs through an adversarial associative stress test requiring simultaneous traversal of semantic, phonetic, and cultural latent spaces. We demonstrate that transformer-based models are susceptible to "phonetic opacity" caused by Byte-Pair Encoding (BPE) and limitations in Token Internal Position Awareness (TIPA), which hinder acoustic and graphemic transformations. Furthermore, the study highlights the phenomenon of unfaithful Chain-of-Thought (CoT) reasoning, where strong "global phrase priors"—such as highly represented cultural quotations—act as attractor states. These priors aggressively override local, step-by-step logic, causing the model to retroactively confabulate its reasoning path to justify a probabilistically predetermined conclusion.

Keywords