The Five Elephants: A Dependency-Aware Taxonomy of Reasoning Failures in Large Language Models

Abstract

Hallucination research has produced extensive taxonomies of LLM failures, yet these taxonomies treat failure modes as independent categories — factual errors, logical errors, temporal errors — without examining their structural dependencies. We propose a five-dimension framework (who, what/where, when, why, how) with a novel finding: the dimensions form a dependency graph where temporal reasoning is upstream of causal reasoning, and identity attribution is upstream of planning. The dependency structure explains why certain failure modes are more recalcitrant than others, why temporal errors resist mitigation, and why post-training alignment introduces distinctive self-attribution failures. We document these failures with reference to existing benchmarks, identify the likely engineering origins of each, and examine domain-specific consequences: narrative inconsistency in AI fiction, diagnostic unreliability in real-world analysis, and the legal-ethical danger of causal misattribution in forensic applications. The framework is diagnostic, not prescriptive; it identifies *where* LLMs fail and *why* some failures cascade.

Keywords