Architectural Roots of Trust: A Systematic Review of Transformer Variants for LLM Safety

The emergence of deconstructionism opened new horizons for modern art by dismantling established conventions, the maturing field of LLM safety now requires a similar deconstructive approach. It is imperative that we move beyond surface-level analyses to holistically dissect the foundational architectures of large language models, scrutinizing their inherent trade-offs to forge new paths toward inherently safer Large Language Models (LLMs). The proliferation of LLMs based on the transformer architecture has precipitated an urgent need to understand their inherent Trust and Safety (T&S) properties. While extensive research has focused on post-hoc safety mitigations, a systematic analysis of how fundamental architectural choices influence T&S outcomes remains underdeveloped. This paper conducts a systematic review of the three primary transformer architectural families—Encoder-only, Decoder-only, and Encoder-Decoder—from a T&S perspective. This study analyze their intrinsic predispositions to critical failure modes, including bias amplification, robustness against adversarial perturbations, and hallucination. This review reveals that each architecture presents a unique T&S trade-off profile, intrinsically linked to its core mechanisms like bidirectional context encoding versus autoregressive generation. It synthesize the current literature to identify significant research gaps and propose three promising frontiers for building safety into the architectural fabric itself: Causal Transformers, which leverage causal inference to move beyond spurious correlations; Formally Verified Architectures, which use mathematical proof to guarantee safety properties in model components; and Biologically-Inspired designs, which draw on principles of neuroscience to enhance robustness. We provide technical proposals, including pseudo-code and mathematical formulations, to test the viability of these novel approaches. This work concludes that the future of AI safety lies in a paradigm shift from treating safety as an add-on to embedding it within the foundational design of LLMs, paving the way for models that are trustworthy by design.