Exploring Vision-Language Models for Online Signature Verification: A Zero-Shot Capability Study
Exploring Vision-Language Models for Online Signature Verification: A Zero-Shot Capability Study
Authors: Marta Robledo-Moreno, Ruben Vera-Rodriguez, Ruben Tolosana, Javier Ortega-García Date: 2026-05-14 Paper ID: openalex:2605.14845
Summary
This study explores the zero-shot capabilities of advanced vision-language models (VLMs) for the biometric task of online signature verification. The authors propose a method to represent kinematic time-series as static images with stroke-opacity pressure encoding and develop a scoring protocol based on latent token probabilities. While the models demonstrate superior discrimination in random forgery scenarios, they struggle with skilled forgeries, where chain-of-thought reasoning often leads to a ‘Rationalization Trap’ that misinterprets forgery artifacts.
Key Contributions
- Evaluates zero-shot performance of GPT-5.2 and Gemini 2.5 Pro on the Signature Verification Challenge (SVC) benchmark.
- Introduces a visual encoding method converting kinematic signature time-series into opacity-based static images.
- Identifies the ‘Rationalization Trap’ where chain-of-thought reasoning degrades accuracy in skilled forgery detection due to hallucinated justifications.
Open Questions & Future Work
Key Concepts
- rationalization-trap: A phenomenon where chain-of-thought reasoning in vision-language models leads to kinematic hallucinations that misclassify forgery artifacts as natural variability.
Archivist Review
Approved the ‘Rationalization Trap’ as a reusable concept for failure modes in multimodal CoT reasoning, and the open question regarding its mitigation. Rejected the SVC dataset as it is a specific biometric task benchmark outside the primary scope of time-series forecasting research.
Approved Concepts
- Rationalization Trap: Highlights a failure mode where explicit reasoning (CoT) in multimodal models misinterprets biometric signal artifacts, especially in skilled forgery scenarios.
Approved Open Questions
- Mitigating VLM Rationalization Traps: Essential for building reliable forensic and diagnostic AI systems where accurate and interpretable reasoning is a prerequisite for human-in-the-loop verification.
Rejected Candidates
- [dataset] Signature Verification Challenge (SVC) (
svc) - low_impact: The SVC benchmark is a specific domain-standard set for signature verification that, while useful, is less critical for the broader time-series forecasting vault than more generalized benchmarks.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.14845
- paper_id
- 2605.14845
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- rationalization-trap
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-17T06:08:33Z