Most AI systems can tell you two images are 87% similar — they just can't tell you why.
Researchers have introduced Distance Explainer, a method for cracking open the embedding spaces that underpin modern machine learning models. Embeddings are compressed numerical representations of data — images, text, audio — that models use to compare and relate things. Until now, the XAI field has produced plenty of tools to explain individual predictions, but almost nothing to explain why two embedded data points land close together or far apart. Distance Explainer adapts a technique called RISE — originally built for image saliency — to assign credit values to features by selectively masking inputs and filtering results by distance rank. Tests on ImageNet and CLIP models showed the method identifies contributing features while holding up under robustness and consistency checks.
Embedding spaces are everywhere: image search, content recommendation, facial-recognition systems, and multimodal AI all live inside them. Without tools to audit those spaces, developers and auditors are flying blind when a model flags a false match or misses an obvious one. A working explainability layer matters especially for high-stakes deployments where regulators are starting to demand algorithmic transparency.
The method is post-hoc — it explains decisions after the fact rather than redesigning the model — which means it can slot into existing pipelines without retraining. That's practical, but it also means Distance Explainer inherits the usual caveats of post-hoc XAI: explanations approximate the model's reasoning rather than expose it directly.