Large-scale language models can produce anomalous outputs that cannot be explained solely by the quality of input data. This article presents a systematic review and descriptive quantitative synthesis of published evidence on reasoning anomalies in LLMs. The study does not report original experiments, does not evaluate new model outputs, and does not implement the proposed framework. Instead, it consolidates numerical results manually extracted from published papers, public benchmarks and official system cards, with the derived datasets and figure-generation scripts. The review organizes reasoning-related anomalies into a taxonomy that distinguishes factual hallucinations, self-contradictions, unfaithful Chain-of-Thought traces, semantic rollback, snowball errors, distractor susceptibility and sycophancy bias. Published evidence indicates that irrelevant context can reduce accuracy below 30% in controlled mathematical reasoning settings, and that knowledge and reasoning tasks can differ by more than 12 percentage points in reported biomedical benchmarks. Existing mitigation techniques, including self-consistency, semantic entropy, process reward models and formal verification, are compared descriptively across heterogeneous studies and domains. Formal verification results, such as the 91.7% reported for VERGE on AR-LSAT, are explicitly limited to structured logical reasoning and should not be generalized to open-domain natural language reasoning. Finally, the article proposes a four-layer conceptual architecture grounded in access-consciousness and monitoring-consciousness operators, CA(t) and CM(t), for filtering context, generating reasoning paths, monitoring inferential discrepancies and activating selective correction. The framework is presented as an implementable research roadmap whose empirical validation, computational overhead and activation thresholds remain future work.
Arévalo-Royo et al. (Mon,) studied this question.