Researchers Propose FailSAE for Interpretable Failure Prediction in VLMs

Researchers have introduced FailSAE, a framework designed to predict and interpret failures in vision-language models (VLMs) like CLIP. As multimodal models are deployed in high-stakes environments, reliable failure prediction is critical for risk-aware deployment and timely intervention.

While existing methods rely on confidence scores or auxiliary classifiers, they offer limited interpretability. FailSAE formulates failure prediction as a classification task over sparse autoencoder (SAE) latent activations and uses a three-stage training pipeline to make latent directions both interpretable and informative for error detection.

Experiments demonstrate that FailSAE outperforms evaluated baselines. Concept-level analysis using SAEs revealed that representations shift from class-specific concepts to ambiguous or style-related concepts during failures, while learned latent directions can also assist in runtime failure recovery.