UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations

UniProbe detects token-level hallucinations in frozen LVLMs by modeling heterogeneous computational traces through a directed graph of image patches and query tokens. This lightweight approach captures spatial and sequential structures from a single forward pass, avoiding the expensive full-model fine-tuning or external verifiers required by existing methods.