Axiomatic Reasoning for LLMs

1. Scope and Taxonomy of Failures

Large language model (LLM) limitations are categorized into three operational domains. Each domain corresponds to a specific class of human cognitive or behavioral phenomena.

Domain	LLM Failure Mode	Human Cognitive Equivalent
Accuracy	Hallucination, Confabulation, Factual Drift	False Memory (DRM Paradigm), Source Monitoring Error
Reasoning	Framing Bias, Anchoring, Sycophancy	Cognitive Heuristics (System 1), Logical Fallacy
Social	Out-group Bias, Polarization Echoes, Stereotype Reinforcement	Social Identity Theory, In-group Favoritism

2. Domain-Specific Congruence

2.1 Accuracy and Memory Distortion

LLM hallucination is functionally isomorphic to human confabulation. Both systems construct plausible but unverified narratives to bridge information gaps.

Evidence Pattern: LLMs exhibit List-Length Effect, Fan Effect, and Negation-Induced Forgetting in controlled prompt engineering tests.
Divergence: Humans possess a feedback-dependent reality monitoring mechanism; LLMs lack an intrinsic verification loop unless provided via external tools.

2.2 Reasoning and Heuristics

LLM inference pathways mirror human judgment biases when constrained by probability distributions over token sequences.

Bias Replication: Anchoring and Framing effects are systematic across model sizes.
Calibration Anomaly: Smaller models exhibit high confidence with low accuracy (Dunning-Kruger profile); larger models invert this pattern.

LLM agent collectives generate emergent polarization patterns consistent with human social network models.

Structural Similarity: LLM agents form homophilous clusters and echo chambers under iterative interaction.
Amplification Factor: LLM output vectors act as mechanical amplifiers of training data skew, scaling human bias to synthetic reach.

3. The Grounding Differential: A Technical Restatement

The distinction previously framed as Symbol Grounding translates to the engineering gap between Correlational Semantics and Causal World Models.

Aspect	LLM Architecture	Human Biological System
Input Modality	Discrete token sequences (digital text/images)	Continuous multimodal stream with proprioception
State Maintenance	Stateless activation (context window)	Homeostatic and allostatic regulation (Internal Embodiment)
Error Correction	External feedback (RLHF, tool use)	Internal error prediction and sensorimotor recalibration

Current LLM implementations navigate language games via statistical dependency. The absence of a persistent, embodied world model constitutes a divergence in mechanism, not necessarily a failure in task performance within bounded domains.

4. Consolidated Equivalence Matrix

Problem Category	Present in LLMs	Present in Humans	Functional Equivalence	Mechanism Divergence
Fact Fabrication	Yes	Yes (Confabulation)	High	High (Intent vs. Probability)
Anchoring Bias	Yes	Yes	High	Low
Out-group Prejudice	Yes	Yes	High	Medium (Learned vs. Innate)
Loss Aversion	Low/Negligible	High	Low	High
Internal State Awareness	No	Yes	Null	Absolute

5. Summary of Correspondence

The operational footprint of LLM limitations aligns with human cognitive and social biases in the majority of documented cases. The primary axis of difference lies not in the behavioral signature of the error, but in the underlying control architecture and the absence of persistent, self-regulating internal states.

Key Consequence: Mitigation strategies effective for human bias (e.g., structured checklists, exposure to counter-evidence) retain partial transferability to LLM prompting and guardrail design, provided the distinct implementation layer is accounted for.

This site is open source. Improve this page.