Counterfactual Reasoning for Robust Visual Question Answering

Visual Question Answering (VQA) with Multimodal Large Language Models (MLLMs) requires not only producing safe and effective responses, but also grounding safety decisions in the multimodal evidence that determines risk. Recent safety-alignment methods improve refusal behavior and contextual risk awareness, yet correct safety outcomes may still rely on superficial textual or visual correlations, particularly when risk emerges from interactions between individually benign image and question content. To address this issue, we propose A$^2$Safe, a counterfactual evidence-aligned adaptive agent collaboration framework for safe and effective VQA. A$^2$Safe organizes localized visual observations, textual intent, and cross-modal risk relations through a Grounded Safety Evidence Board, making the basis of safety decisions explicit. Counterfactual safety evidence alignment enforces invariance to safety-irrelevant changes while requiring appropriate safety-state and response-mode transitions when risk-critical evidence is minimally altered. The resulting evidence state further supports adaptive collaboration, enabling direct answering when grounded evidence is sufficient and invoking policy critique and response revision when evidence is risky, uncertain, or conflicting. Under complementary safety-critical and general VQA protocols, A$^2$Safe achieves a 95.72 SIUO safety score, reduces the benign refusal rate on MOSSBench to 14.67%, and maintains an average general VQA score of 78.34 with 27.8% token overhead. These results support counterfactual evidence-aligned adaptive collaboration for safe and effective multimodal question answering.

Vision-Language Models (VLMs) excel at multimodal reasoning, yet it remains unclear whether their answers are grounded in visual evidence or driven by learned language and world priors. Counting provides a precise testbed: when visual evidence conflicts with canonical object knowledge, a model must rely on the image rather than a prototypical count. We introduce CounterCount, a diagnostic framework for counterfactual counting in VLMs, consisting of paired factual and counterfactual images with edited count-relevant attributes, verified answers, and localized evidence annotations. Evaluating recent VLMs, we find strong performance on factual images but consistent degradation under counterfactual attribute changes, indicating reliance on object-level priors even when contradictory visual evidence is present. Using localized annotations, we show that these failures are not solely due to missing or ambiguous visual evidence, but to models underweighting attention to count-relevant visual tokens. We introduce a unified inference-time attention modulation strategy that reweights selected visual tokens, improving counterfactual counting accuracy by up to 8% across multiple VLMs. Overall, CounterCount exposes prior-driven counting failures and provides diagnostic insights for designing future VLMs.

Current visual models often make predictions based on a limited set of discriminative visual cues. As a result, they may become unreliable when the distribution shifts or when these cues are missing. Faithful attribution methods can reveal such problematic reliance through localized explanations, but they are typically used post hoc and are not fed back into the model. To address this limitation, we propose Subset-Selected Counterfactual Augmentation (SS-CA), a training strategy that masks decision-relevant regions to construct counterfactual samples and guide the model toward more robust decision boundaries. Specifically, we extend LIMA, a subset-selection-based faithful attribution method, to Counterfactual LIMA to identify regions whose removal shifts the model toward a competing class. SS-CA then selects near-boundary masks that reduce the logit gap while preserving the original semantics, and applies an adaptive counterfactual filling strategy to replace the masked regions without introducing external semantics. Feeding these counterfactual samples back into training encourages the model to exploit the remaining informative evidence and shifts the decision boundary toward a more robust one. Extensive experiments across five ImageNet variants show that SS-CA effectively improves ID accuracy, OOD generalization, and perturbation robustness, achieving gains of 5.70%/18.04% on ImageNet-1k/ImageNet-R with CLIP ViT/32b, 9.52%/11.33% on ImageNet-R/ImageNet-S on TinyImageNet-200 with ResNet-101, and about 4% under Gaussian Noise corruption. The code will be released soon.