LogitDynamics: Reliable ViT Error Detection from Layerwise Logit Trajectories

2026-04-12

AI Summary

This work addresses the lack of reliable confidence estimation in deployed Vision Transformers (ViTs) by proposing a lightweight method that predicts classification errors with only a single forward pass. The approach attaches linear probes to intermediate layers of a ViT to extract trajectory features—including logits of the predicted and competing classes and the stability of class rankings—from the last L layers. It introduces, for the first time, the concept of depth-wise signal analysis from large language models into ViTs, establishing an error detection mechanism based on inter-layer logit dynamics. With minimal computational overhead, the method achieves competitive or superior AUC-PR performance across multiple datasets and demonstrates strong cross-dataset generalization capabilities.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Safety and Robustness

Application Category

• Security and Privacy: Data transparency and provenance
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Reliable confidence estimation is critical when deploying vision models. We study error prediction: determining whether an image classifier's output is correct using only signals from a single forward pass. Motivated by internal-signal hallucination detection in large language models, we investigate whether similar depth-wise signals exist in Vision Transformers (ViTs). We propose a simple method that models how class evidence evolves across layers. By attaching lightweight linear heads to intermediate layers, we extract features from the last L layers that capture both the logits of the predicted class and its top-K competitors, as well as statistics describing instability of top-ranked classes across depth. A linear probe trained on these features predicts the error indicator. Across datasets, our method improves or matches AUCPR over baselines and shows stronger cross-dataset generalization while requiring minimal additional computation.

Problem

• error detection
• Vision Transformers
• confidence estimation
• logit trajectories
• model reliability

Innovation

• LogitDynamics
• Vision Transformers
• error detection
• layerwise logits
• confidence estimation