The paper introduces the Alignment Quality Index (AQI), a decoding-invariant metric leveraging latent geometric representations and clustering indices to diagnose hidden misalignments in large language models (LLMs), even under behavioral compliance.
Intrinsic Latent Geometry Metric: AQI measures alignment by assessing how well safe and unsafe prompts form distinct clusters in a model’s latent space using a combination of Xie-Beni and Calinski-Harabasz indices, making it invariant to decoding strategies and resistant to alignment faking.
Layerwise Pooled Representation Learning: It uses a sparse, learned pooling mechanism over hidden transformer layers to capture alignment-relevant abstractions without modifying the base model, enabling robust internal safety diagnostics.
Empirical Failures of Behavioral Metrics: AQI reveals misalignments missed by traditional metrics (e.g., G-Eval, refusal rates) in scenarios like jailbreaks, safety-agnostic fine-tuning, and stochastic decoding—showcasing its strength as an early-warning alignment audit tool.