CARGO-VL: New Framework for Reliable Vision-Language Models
A recent study presents CARGO-VL, a framework aimed at enhancing the dependability of vision-language models by resolving discrepancies between textual and visual evidence. This paper, which can be found on arXiv (2608.04509), suggests a group-relative optimization strategy that consolidates various evidence states—aligned, image-correct, text-correct, and both-wrong—into a unified package. It integrates condition-wise accuracy with transition rewards to achieve answer invariance, source equivariance, and switching from answers to abstention. To manage risky responses against over-deferral, a primal-dual controller is employed. Additionally, the authors introduce XMC (eXtended Modal Conflict), a training tool featuring four conflict scenarios, and assess transfer capabilities using CMC-Bench and Modality-Bias benchmarks. This framework seeks to help models recognize reliable sources and refrain from responding when evidence is insufficient, tackling a significant issue in multimodal AI.
Key facts
- CARGO-VL is a group-relative framework for vision-language models.
- It optimizes matched variants covering aligned, image-correct, text-correct, and both-wrong evidence states.
- The objective couples condition-wise correctness with transition rewards.
- A primal-dual controller balances unsafe answers against excessive deferral.
- XMC (eXtended Modal Conflict) is a four-condition conflict training resource.
- Evaluation is performed on CMC-Bench and Modality-Bias.
- The paper is available on arXiv with ID 2608.04509.
- The framework addresses counterfactual evidence changes in vision-language systems.
Entities
Institutions
- arXiv