Evaluation of Vision-Language Models for Iris Verification Across Datasets and Representations

Vision-language models (VLMs) increasingly serve as general-purpose, zero-shot alternatives to specialised systems, raising the question of whether they can perform biometric recognition. Iris verification is a demanding test of this capability, relying on fine-grained texture and discrimination at low false-match rates, yet it remains largely unexamined in a literature concentrated on faces and soft biometrics. We address this gap with an evaluation protocol that converts a VLM’s forced binary output into a continuous, thresholdable same-eye score derived from first-token log-probabilities, enabling the operating-point analysis that biometric evaluation requires. Across four iris representations and two acquisition regimes, we disentangle the effect of representation from that of acquisition for the Qwen3.5 family of models. VLMs approach previous works, particularly at higher parameter counts, only on original ocular images under visible-light acquisition, degrading sharply toward canonical iris representations. A classical iris-code baseline is surpassed in the visible-light regime but dominates decisively under near-infrared.