Researchers have introduced SLAPBench, the first benchmark designed to evaluate multimodal large language models (MLLMs) on four-finger SLAP fingerprint verification. The benchmark, built using NIST SD302b data, tests MLLMs' ability to verify identity from SLAP images, which are flat live-scan impressions of four fingers. Initial evaluations show that prompting strategies significantly impact model performance, with some prompts causing near-100% false acceptance rates across multiple open-source models. Claude Opus 4.8 demonstrated superior resilience to these prompting issues, achieving the best binary verification results, while other models showed varying capabilities in discrimination, with one model achieving a perfect score that warrants further investigation into potential shortcuts. AI
IMPACT Establishes a new benchmark for evaluating MLLMs in biometric verification, highlighting the impact of prompting on model performance and potential shortcuts.
RANK_REASON The item describes a new benchmark and research paper evaluating existing models on a novel task. [lever_c_demoted from research: ic=1 ai=1.0]
- Claude Opus 4.8
- Four-finger SLAP fingerprints
- Gemma-3-12B
- InternVL3-8B
- multimodal large language models
- NIST SD302b
- Qwen2.5-VL-7B
- Qwen3-VL-8B
- SLAPBench
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
