An Information-Geometric Framework for Stability Analysis of Large Language Models under Entropic Stress

Aggregate accuracy metrics are insufficient for assessing LLM reliability: The paper demonstrates that traditional accuracy-based evaluation fails to capture system reliability in high-stakes deployments, highlighting a critical gap in current AI evaluation practices that responsible deployment requires addressing.

Internal model structure matters for handling uncertainty: The framework shows that how LLMs are internally organized—measured through "internal integration" and "aligned reflective capacity"—meaningfully affects their stability when facing uncertain or perturbed inputs, suggesting that responsible AI evaluation should examine model internals, not just outputs.

Uncertainty amplifies performance differences: The analysis of 80 model-scenario observations across four LLMs found that stability improvements become "more pronounced under higher entropy conditions," meaning models with better internal structure maintain reliability better when facing real-world uncertainty and ambiguity.

A 3% stability improvement from structural factors: The proposed framework achieved a mean improvement of 0.0299 in stability scores compared to simpler utility-entropy baselines (95% CI: 0.0247-0.0351), demonstrating that accounting for internal model properties provides measurable gains in predicting reliable behavior.

Interpretability enables better risk assessment: By framing stability as an interpretable combination of performance, uncertainty, and internal structure rather than a "black box" metric, this approach provides stakeholders with clearer insight into when and why LLMs might fail in operational settings.