Which Modality Decides? Counterfactual Modality Attribution for Multimodal LLMs

Multimodal large language models (MLLMs) are increasingly provided with contextual evidence in heterogeneous forms: as a text passage, as a rendered image of the same passage, or as both together. However, it remains unclear how consistently these surface forms are processed, especially when the evidence conflicts with the model's parametric knowledge. We study modality robustness under knowledge conflict across 13 MLLMs and two datasets, and find them far from robust. (1) Contrary to common belief, models favor a context that contradicts parametric knowledge more readily in image form than in text form; (2) when a contradicting text and image are presented together, the preferred modality is essentially arbitrary, varying with input order, model, and dataset. We further demonstrate that this instability has practical consequences: it degrades performance in multimodal RAG and can be exploited by adversarial attacks. To alleviate this brittleness, we examine several simple techniques---prompting, steering, supervised fine-tuning (SFT), and direct preference optimization; the majority prove ineffective, whereas SFT achieves moderate success. We therefore call for greater awareness of this inconsistency and argue that it is fundamental, demanding attention at multiple training stages.

Current Multimodal Large Language Models (MLLMs) can process diverse sensory inputs, yet their reasoning remains heavily biased toward a dominant modality, resulting in brittle cross-modal reasoning. We introduce C$^3$PO, a benchmark of 3,404 samples spanning video, audio, image, and text, evaluating two abilities: information composition (fusing dispersed evidence) and counterfactual conflict (resolving deliberate contradictions). C$^3$PO's paired IC/CC structure and four-tier design enable targeted diagnosis of when and why cross-modal reasoning fails. Built through a fully automatic pipeline using 25 logically grounded templates, C$^3$PO reveals that while humans achieve 88.64% accuracy, the best model (Gemini-3.1-Pro) reaches only 73.17%, with open-source models collapsing under conflict. Through attention probes, we find 86-95% of failures stem from modality dominance: models commit to one modality while ignoring contradictory evidence, concentrating 87-95% of attention on text. Mid-layer attention entropy predicts correctness-sustained exploration succeeds, premature collapse fails. The 56-point accuracy gap between equally complex templates reveals that performance depends on modalities' structural roles in conflict resolution, not combinations. These findings show multimodal perception does not guarantee robust reasoning; architectures must enable sustained cross-modal attention to avoid premature

Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.