Physics-based phenomenological characterization of cross-modal bias in multimodal models
Authors: Hyeongmo Kim, Sohyun Kang, Yerin Choi, Seungyeon Ji, Junhyuk Woo, Hyunsuk Chung, Soyeon Caren Han, Kyungreem Han
Organizations: Brain Science Institute, Korea Institute of Science and Technology, Seoul, Republic of Korea · Department of Physics and Astronomy, Seoul National University, Seoul, Republic of Korea · Department of Computer Science and Engineering, Korea University, Seoul, Republic of Korea · The University of Melbourne, Melbourne, Australia · University of Science and Technology KIST School, Seoul, Republic of Korea
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.
Figures & tables
Below is a list of possible emotions: {emotion_list}.
Your task: Recognize dominant emotion expressed in the {modality_list}.
Return the result in valid JSON array format as follows:
[{"emotion": ""}].
Do not output anything else (no extra commentary).
Table 1: The prompt for the zero-shot emotion classification task. The {modality_list} field is set to ”face and voice” for combined video–audio inputs, ”face” for video-only inputs (with audio replaced by silence), and ”voice” for audio-only inputs (with video frames replaced by blank placeholders).
Figure 1: Error-attractor structures under perturbations of emotion labels in multimodal large language models. Directed graphs visualize incorrect emotion classifications on the CREMA-D dataset under three input conditions: Face (Video) + Voice (Audio), Face (Video) only, and Voice (Audio) only. Nodes represent six emotion labels (happy, neutral, sad, angry, disgust, fear), and directed edges depict erroneous mappings from intended labels to predicted labels.
Figure 2: Sankey diagram displaying emotion predictions in Qwen2.5-Omni. The width of each flow indicates the number of samples assigned to each mapping: 1) from the intended label (left) to the model prediction (center), and 2) from the perceived label (right) to the model prediction (center).
Figure 3: Illustration of the prompt-based label perturbation strategy
Figure 4: Schematic diagram of Lorenz chaotic time-series prediction using the multi-oscillator model with self- and cross-attention mechanisms.
Figure 5: Characterization of the transformer dynamics using a physical testbed: Lorenz chaotic time-series prediction on a multi-oscillator system. (a) Modality preference is quantified by the difference in dynamical SHAP values, ϕ(Y)−ϕ(X) , across the self- and cross-attention levels (βself,βcross) . The SHAP difference is represented by the direction of the arrow in the range of [−90∘,90∘] : 0∘ signify the equal contribution of X and Y , −90∘ indicate the X -only; 90∘ the Y -only. The arrow color represents the normalized mean squared error (NMSE) between target z(t) and prediction. Predictions are visualized in the embedding space for two representative cases: (b) low self- and cross-attention levels, i.e., (βself,βcross)=(10−4,10−4) and (c) high levels (βself,βcross)=(100,100) . To clarify, only the time series for 50≤t≤70 are displayed.
Multimodal large language models (MLLMs) increasingly support high-stakes decision making by combining complementary information from images and text. While existing explainability methods identify influential image regions or text tokens, they cannot answer a fundamental question: which modality drives a prediction? Consequently, a model may produce the correct output while relying on the wrong source of evidence, masking shortcut learning and unsafe reasoning. We formulate modality attribution as a complementary explainability objective for multimodal foundation models and propose Counterfactual Modality Attribution (CMA), the first framework for quantifying modality-level contributions in MLLMs. CMA generates image-only, text-only, and joint multimodal counterfactuals using coupled diffusion priors and converts them into principled modality attribution scores through a cooperative game-theoretic formulation based on Shapley values. We evaluate CMA on controlled synthetic benchmarks with known ground-truth modality reliance and on a real-world multimodal clinical dataset. CMA correctly identifies the decision-driving modality in 98% of controlled cases and consistently outperforms baselines, revealing failures of cross-modal reasoning that remain invisible to predictive accuracy alone. Our results establish modality attribution as a complementary dimension of explainability beyond feature attribution, providing a principled framework for auditing multimodal foundation models in safety-critical applications.
Vahidin Hasic, Chao Wang, Luis C. Garcia-Peraza-Herrera +2
Faculty of Electrical Engineering, University of Sarajevo, Bosnia and Herzegovina · King’s College London, United Kingdom
Standard benchmarks for multimodal large language models (MLLMs) score each item on one canonical ordering and miss whether order-irrelevant shuffling changes the answer, a baseline reliability property called for by emerging AI evaluation guidelines. We introduce Facet-Probe, a five-facet audit (option, evidence-chunk, document-rank, image-set, and mixed-modality ordering) of 18 frontier and open-weight MLLMs. A Bayesian item-response model separates ordering noise from per-facet bias, and a same-ordering control estimates the decoder-stochastic floor for observed flips. We find that none of the 18 MLLMs we audit are order-invariant: screened per-facet panel-mean flip rates span 24-50%. A Gemini same-ordering control at temperature 0 estimates a substantial ordering excess over a same-input decoder-noise floor in verified cells. Capability predicts but does not eliminate flips; the best model still flips on 13.4% of trials. In our Gemini mitigation tests, training-free prompt changes are modality-conditional and do not transfer from text to visual reasoning. These results suggest that prompt-level mitigation alone is unlikely to provide general order robustness, motivating future work on training-time and architectural approaches. We propose cross-ordering flip rate as a standard reporting axis for MLLMs.
Automatic evaluation using multimodal large language models (MLLMs), commonly referred to as MLLM-as-a-Judge, has been widely used to measure model performance. If such MLLM-as-a-Judge methods were biased, they could distort model comparisons and benchmark-driven scientific progress. However, it remains unclear to what extent MLLM-as-a-Judge methods favor or disfavor text generated by specific MLLMs. In this study, we propose Philautia-Eval to investigate such model-specific preference bias. Philautia-Eval quantifies the degree of the bias by disentangling preference tendencies from differences in generation quality. Using 1.29M caption-score pairs collected from 12 MLLMs, we found that representative MLLMs tend to exhibit self-preference bias. Moreover, experimental results indicate mutual preference bias within particular model families, which is potentially driven by reused connectors and overlapping instruction-tuning resources. Finally, we introduce a simple ensemble of MLLMs, Pomms. Our results demonstrated that Pomms effectively mitigated the model-specific preference bias while maintaining performance.