Search papers, labs, and topics across Lattice.
CARGO-VL introduces a novel group-relative framework for vision-language models that optimizes the handling of conflicting evidence from images and text by treating them as a cohesive bundle. This approach enhances the model's ability to discern trustworthy sources and abstain from answering when neither source is adequate, addressing the limitations of existing independent scoring methods. The results demonstrate significant improvements in conflict handling, unsupported-answer avoidance, and modality balance compared to traditional pointwise baselines, highlighting the effectiveness of relational transition signals and adaptive risk control in achieving counterfactual consistency.
CARGO-VL achieves a breakthrough in multimodal reliability by effectively managing conflicting evidence, leading to better decision-making in vision-language tasks.
Vision-language systems combine images with retrieved text, but these sources can disagree or jointly fail to support an answer. Reliable models must identify the trustworthy source and abstain when neither is adequate. Existing post-training objectives score instances independently and therefore do not enforce coherent behavior under counterfactual evidence changes. We introduce CARGO-VL, a group-relative framework that optimizes matched variants covering aligned, image-correct, text-correct, and both-wrong (A/V/T/N) evidence states as one bundle. Its objective couples condition-wise correctness with transition rewards for answer invariance, source equivariance, and answer-to-abstention switching, while a primal-dual controller balances unsafe answers against excessive deferral. We also contribute XMC (eXtended Modal Conflict), a four-condition conflict training resource, and evaluate transfer on CMC-Bench and Modality-Bias. Across multiple seeds, CARGO-VL improves conflict handling, unsupported-answer avoidance, and modality balance over pointwise baselines. Ablations identify complementary benefits from relational transition signals and adaptive risk control, supporting counterfactual consistency as a practical objective for reliable multimodal evidence arbitration.
