Multimodal Model Interpretability

The term 'algorithmic fairness' is used to evaluate whether AI models operate fairly in both comparative (where fairness is understood as formal equality, such as "treat like cases as like") and non-comparative (where unfairness arises from the model's inaccuracy, arbitrariness, or inscrutability) contexts. Recent advances in multimodal large language models (MLLMs) are breaking new ground in multimodal understanding, reasoning, and generation; however, we argue that inconspicuous distortions arising from complex multimodal interaction dynamics can lead to systematic bias. The purpose of this position paper is twofold: first, it is intended to acquaint AI researchers with phenomenological explainable approaches that rely on the physical entities that the machine experiences during training/inference, as opposed to the traditional cognitivist symbolic account or metaphysical approaches; second, it is to state that this phenomenological doctrine will be practically useful for tackling algorithmic fairness issues in MLLMs. We develop a surrogate physics-based model that describes transformer dynamics (i.e., semantic network structure and self-/cross-attention) to analyze the dynamics of cross-modal bias in MLLM, which are not fully captured by conventional embedding- or representation-level analyses. We support this position through multi-input diagnostic experiments: 1) perturbation-based analyses of emotion classification using Qwen2.5-Omni and Gemma 3n, and 2) dynamical analysis of Lorenz chaotic time-series prediction through the physical surrogate. Across two architecturally distinct MLLMs, we show that multimodal inputs can reinforce modality dominance rather than mitigate it, as revealed by structured error-attractor patterns under systematic label perturbation, complemented by dynamical analysis.

Multimodal recommender systems leverage diverse data sources, such as user interactions, content features, and contextual information, to address challenges like cold-start and data sparsity. However, existing methods often suffer from one or more key limitations: processing different modalities in isolation, requiring complete multimodal data for each interaction during training, or independent learning of user and item representations. These factors contribute to increased complexity and potential misalignment between user and item embeddings. To address these challenges, we propose DReX, a unified multimodal recommendation framework that incrementally refines user and item representations by leveraging interaction-level features from multimodal feedback. Our model employs gated recurrent units to selectively integrate these fine-grained features into global representations. This incremental update mechanism provides three key advantages: (1) simultaneous modeling of both nuanced interaction details and broader preference patterns, (2) eliminates the need for separate user and item feature extraction processes, leading to enhanced alignment in their learned representation, and (3) inherent robustness to varying or missing modalities. We evaluate the performance of the proposed approach on three real-world datasets containing reviews and ratings as interaction modalities. By considering review text as a modality, our approach automatically generates interpretable keyword profiles for both users and items, which supplement the recommendation process with interpretable preference indicators. Experiment results demonstrate that our approach outperforms state-of-the-art methods across all evaluated datasets.

Advances in image tampering techniques, particularly generative models, pose significant challenges to media verification, digital forensics, and public trust. Existing image forgery detection and localization (IFDL) methods suffer from two key limitations: over-reliance on semantic content while neglecting textural cues, and limited interpretability of subtle low-level tampering traces. To address these issues, we propose FOCA, a multimodal large language model-based framework that integrates discriminative features from both the RGB spatial and frequency domains via a cross-attention fusion module. This design enables accurate forgery detection and localization while providing explicit, human-interpretable cross-domain explanations. We further introduce FSE-Set, a large-scale dataset with diverse authentic and tampered images, pixel-level masks, and dual-domain annotations. Extensive experiments show that FOCA outperforms state-of-the-art methods in detection performance and interpretability across both spatial and frequency domains.

Despite the recent success of Multimodal Large Language Models (MLLMs), existing approaches predominantly assume the availability of multiple modalities during training and inference. In practice, multimodal data is often incomplete because modalities may be missing, collected asynchronously, or available only for a subset of examples. In this work, we propose PRIMO, a supervised latent-variable imputation model that quantifies the predictive impact of any missing modality within the multimodal learning setting. PRIMO enables the use of all available training examples, whether modalities are complete or partial. Specifically, it models the missing modality through a latent variable that captures its relationship with the observed modality in the context of prediction. During inference, we draw many samples from the learned distribution over the missing modality to both obtain the marginal predictive distribution (for the purpose of prediction) and analyze the impact of the missing modalities on the prediction for each instance. We evaluate PRIMO on a synthetic XOR dataset, Audio-Vision MNIST, and MIMIC-III for mortality and ICD-9 prediction. Across all datasets, PRIMO obtains performance comparable to unimodal baselines when a modality is fully missing and to multimodal baselines when all modalities are available. PRIMO quantifies the predictive impact of a modality at the instance level using a variance-based metric computed from predictions across latent completions. We visually demonstrate how varying completions of the missing modality result in a set of plausible labels.

Alzheimer's disease (AD) is a progressive neurodegenerative condition necessitating early and precise diagnosis to provide prompt clinical management. Given the paramount importance of early diagnosis, recent studies have increasingly focused on computer-aided diagnostic models to enhance precision and reliability. However, most graph-based approaches still rely on fixed structural designs, which restrict their flexibility and limit generalization across heterogeneous patient data. To overcome these limitations, the Meta-Relational Copula-Based Graph Attention Network (MRC-GAT) is proposed as an efficient multimodal model for AD classification tasks. The proposed architecture, copula-based similarity alignment, relational attention, and node fusion are integrated as the core components of episodic meta-learning, such that the multimodal features, including risk factors (RF), Cognitive test scores, and MRI attributes, are first aligned via a copula-based transformation in a common statistical space and then combined by a multi-relational attention mechanism. According to evaluations performed on the TADPOLE and NACC datasets, the MRC-GAT model achieved accuracies of 96.87% and 92.31%, respectively, demonstrating state-of-the-art performance compared to existing diagnostic models. Finally, the proposed model confirms the robustness and applicability of the proposed method by providing interpretability at various stages of disease diagnosis.

Radiology Report Generation (RRG) through Vision-Language Models (VLMs) promises to reduce documentation burden, improve reporting consistency, and accelerate clinical workflows. However, their clinical adoption remains limited by the lack of interpretability and the tendency to hallucinate findings misaligned with imaging evidence. Existing research typically treats interpretability and accuracy as separate objectives, with concept-based explainability techniques focusing primarily on transparency, while Retrieval-Augmented Generation (RAG) methods targeting factual grounding through external retrieval. We present Concept-Enhanced Multimodal RAG (CEMRAG), a unified framework that decomposes visual representations into interpretable clinical concepts and integrates them with multimodal RAG. This approach exploits enriched contextual prompts for RRG, improving both interpretability and factual accuracy. Experiments on MIMIC-CXR and IU X-Ray across multiple VLM architectures, training regimes, and retrieval configurations demonstrate consistent improvements over both conventional RAG and concept-only baselines on clinical accuracy metrics and standard NLP measures. These results challenge the assumed trade-off between interpretability and performance, showing that transparent visual concepts can enhance rather than compromise diagnostic accuracy in medical VLMs. Our modular design decomposes interpretability into visual transparency and structured language model conditioning, providing a principled pathway toward clinically trustworthy AI-assisted radiology.

Multimodal large language models (MLLMs) in long chain-of-thought reasoning often fail when different knowledge sources provide conflicting signals. We formalize these failures under a unified notion of knowledge conflict, distinguishing input-level objective conflict from process-level effective conflict. Through probing internal representations, we reveal that: (I) Linear Separability: different conflict types are explicitly encoded as linearly separable features rather than entangled; (II) Depth Localization: conflict signals concentrate in mid-to-late layers, indicating a distinct processing stage for conflict encoding; (III) Hierarchical Consistency: aggregating noisy token-level signals along trajectories robustly recovers input-level conflict types; and (IV) Directional Asymmetry: reinforcing the model's implicit source preference under conflict is far easier than enforcing the opposite source. Our findings provide a mechanism-level view of multimodal reasoning under knowledge conflict and enable principled diagnosis and control of long-CoT failures.

We treat the internals of generative models as mechanistic objects rather than black boxes. We introduce Attribution Graphs (AGs), which extend GradCAM++ to circuit-level representations, and Causal Probing, a do-calculus intervention method for identifying causal latent structures, enabling detection and correction of spurious correlations, demographic biases, and misaligned decision circuits during training. We further propose the Cognitive Alignment Score (CAS), quantifying agreement between model-internal representations and human concepts, a saliency-first privacy mechanism sharing only thresholded attribution nodes, a bias-aware regularizer aligning subgroup statistics, and a Reveal-to-Revise loop integrating attribution signals into parameter updates without separate fine-tuning. Evaluated on CelebA, FairFace, Jigsaw, and HateXplain, our method achieves 94.1% accuracy, 92.3% macro F1, 79.4% IoU-XAI, and 12.7 FID at 72--76% adversarial robustness, while reducing subgroup disparity $Δ_{\mathrm{bias}}$ by 41%, demonstrating that mechanistic interpretability, fairness, and generative performance can be jointly optimized.

Multimodal large language models (MLLMs) have achieved strong vision-language performance, yet their token-level visual evidence remains difficult to inspect. Recent logit-lens attribution methods project each visual-token hidden state into the vocabulary space to explain generated words, but this token-wise readout introduces a mismatch: visual tokens are context-mixed by the model, while the attribution score is decoded independently at each token location. This often produces fragmented attribution maps and can be further affected by autoregressive context signals from preceding text tokens. We propose ERCR, an attribution framework built from Evidence Recomposition (ER) and Predictive Context Residualization (PCR). ER aggregates target evidence across multiple views with different token-to-region assignments, reducing attribution fragmentation caused by a single readout grid. PCR estimates a preceding-token context map with RBO-based rank relevance and subtracts its fitted component from the ER map to suppress context-token interference. Experiments on LLaVA, Qwen2-VL, and InternVL families across COCO Caption, GranDf, and OpenPSG show that ERCR improves visual evidence for target tokens and mitigates preceding-token context interference under the existing evaluation protocol. On Qwen2-VL-2B, ERCR improves TAM F1-IoU from 39.10 to 44.45 on COCO Caption and from 30.83 to 37.20 on GranDf. Overall, ERCR provides a practical refinement for token-level visual evidence inspection.

Multimodal prediction models based on imaging and clinical data are increasingly used for clinical decision support, yet their interpretability remains limited. We present multimodal Deep Transformation Models (DTMs) combining statistical approaches and neural networks to achieve strong predictive performance while preserving interpretability for tabular data. A key contribution of this work is the adaption of the xAI methods Grad-CAM and Occlusion to DTMs relying on 3D CNNs, enabling interpretation of the image branch through the generation of explanation maps. We developed DTMs to predict functional independence three months after stroke using diffusion-weighted imaging and clinical data from 407 patients. In a ten-fold cross-validation, the models achieved state-of-the-art predictive performance (AUC 0.81 [0.75, 0.87]) while maintaining interpretability for tabular features, with functional independence before stroke and stroke severity on admission emerging as the strongest predictors. Explanation maps from both xAI methods highlighted consistent regions, including frontal lobe areas which are known to be associated with age, a strong predictor of functional outcome. Notably, these regions disappeared once age was included as an explicit tabular predictor. Similarity analyses of explanation maps revealed distinct spatial patterns, providing meaningful insights into stroke pathophysiology, systematic error analysis and hypothesis generation.