MCAM: Multimodal Causal Analysis Model

• MCAM is a causality-aware multimodal model that organizes driving video data around causal variables and explicit state transitions using a DSDAG structure.
• It fuses multi-level visual evidence with textual narration and reasoning outputs through a vision-language transformer to interpret ego-vehicle actions.
• By incorporating causal formalisms and intervention methods, MCAM effectively mitigates spurious correlations and enhances model interpretability.

Searching arXiv for the named MCAM paper and closely related multimodal causal analysis papers. Searching for "MCAM Multimodal Causal Analysis Model ego-vehicle driving video understanding". Multimodal Causal Analysis Model (MCAM) is, in the narrow sense, a causality-aware multimodal driving video understanding model that explicitly models ego-vehicle state transitions using a Driving State Directed Acyclic Graph (DSDAG), fuses multi-level visual evidence, and aligns causal visual features with text through a vision-language transformer (Cheng et al., 8 Jul 2025). In a broader research sense, the same label is consistent with a line of multimodal methods that treat modalities as variables in an explicit causal system rather than as channels to be fused only by correlation, including structural causal models for code generation, mixed graphical causal discovery for continuous and categorical data, multimodal causal representation learning, causal debiasing for sentiment analysis, causal recommendation, and causal latent-variable sequence models for traffic prediction (Gupta et al., 7 Feb 2025, Sedgewick et al., 2017, Walker et al., 2023, Jiang et al., 7 Aug 2025, Yang et al., 14 Oct 2025, Zhao et al., 2023).

1. Definition and problem setting

The defining property of an MCAM-style framework is that multimodal observations are organized around causal variables, latent factors, or causal graphs, and that model outputs are interpreted through interventions, conditional independence relations, causal priors, or explicit causal state transitions. In the driving formulation, the target is ego-vehicle driving video understanding with two textual outputs: narration, describing what the ego vehicle is doing, and reasoning, explaining why it is doing it. The model decomposes driving into an initial safe state $X_s$, ego action $Y$, environmental change $Z$, end safe state $X_e$, and hidden dangerous state $W$ (Cheng et al., 8 Jul 2025).

The same causal-analysis pattern appears in other multimodal tasks. CodeSCM treats multi-modal code-generation prompts as endogenous variables in an SCM $\mathcal{M} = \langle U, V, F, P(U) \rangle$ with response variable $Y \in \{0,1\}$ and $\mathbb{E}(Y)$ interpreted as dataset accuracy (Gupta et al., 7 Feb 2025). Mixed Graphical Models for Causal Analysis of Multi-modal Variables addresses multimodal biological datasets containing continuous, categorical, and mixed variables by learning a sparse undirected mixed graphical model and then orienting edges with PC-stable or CPC-stable (Sedgewick et al., 2017). In multimodal sentiment analysis, MMCI and MCAF both begin from the claim that standard fusion methods are vulnerable to spurious intra-modal and inter-modal correlations, then introduce causal disentanglement or causal intervention to suppress shortcut learning (Jiang et al., 7 Aug 2025, Dong et al., 29 May 2026).

A common misconception is that multimodal causal analysis is equivalent to ordinary multimodal fusion with more interpretable attention. The cited literature does not support that equivalence. These models add explicit causal objects—SCMs, DAGs, latent causal variables, back-door or front-door adjustments, causal priors, or causal propagation modules—that are intended to separate semantically relevant signal from spurious dependence (Sun et al., 2024).

2. Core causal formalisms

Several distinct causal formalisms recur in MCAM-related work. One is the structural causal model. In CodeSCM, the prompt is decomposed into four modalities—Natural Language ($NL$), algorithmic channel of Code ($Code_{AL}$), natural language channel of Code ($Y$0), and input-output example pairs ($Y$1)—and these are treated as variables that causally influence code correctness. The paper introduces latent mediators $Y$2 and $Y$3 to separate code semantics from natural-language semantics, and it interprets remaining influence after holding mediators fixed as a direct or spurious effect (Gupta et al., 7 Feb 2025).

A second formalism is the explicit DAG over latent or observed factors. causalPIMA assumes categorical latent nodes $Y$4 with factorization

$Y$5

and identifies each Gaussian in a Gaussian mixture prior with a unique outcome of the joint distribution $Y$6 (Walker et al., 2023). The biomedical multimodal causal representation model likewise assumes latent blocks $Y$7 connected by a latent causal DAG and seeks component-wise identifiability up to invertible scalar transforms and permutation under structural sparsity (Sun et al., 2024). In driving video, MCAM defines a DSDAG over start state, action, environment, danger, and end state to model safe-to-safe ego-vehicle transitions (Cheng et al., 8 Jul 2025).

A third formalism explicitly relaxes DAG exclusivity. “Beyond DAGs: A Latent Partial Causal Model for Multimodal Learning” introduces latent coupled variables $Y$8 and $Y$9, with an undirected edge between $Z$0 and $Z$1 to represent flexible cross-modal dependence. Under the stated assumptions, representations learned by multimodal contrastive learning identify the coupled variables up to an orthogonal or permutation-like transformation, depending on the latent geometry (Liu et al., 2024).

A fourth formalism is mixed-type graphical discovery. The mixed graphical model framework assumes a DAG under the standard constraint-based assumptions of no feedback cycles, no unrecorded common causes, and i.i.d. samples. Its contribution is a mixed-data conditional independence test based on the likelihood-ratio statistic

$Z$2

with linear regression for continuous dependent variables and multinomial logistic regression for categorical dependent variables (Sedgewick et al., 2017).

3. Intervention, mediation, and debiasing mechanisms

The operational center of many MCAM-style systems is intervention. CodeSCM defines total effects by removing a modality and measuring the change in accuracy,

$Z$3

and defines direct effects through causal mediation and path effects while attempting to preserve intended semantics through dead-code, dead-string, dead-name, or equivalent-I/O transformations (Gupta et al., 7 Feb 2025). For $Z$4, total effect is the pass@1 drop after removal, while direct effect is the accuracy change after adding dead code so that code semantics are held fixed. The paper states that it does not account for confounders in this version.

Back-door adjustment is central in MMCI. The model constructs a multi-relational graph with $Z$5 relation types—three intra-modal and three inter-modal—then estimates causal attention and shortcut attention for each relation. Its theoretical basis is the replacement of $Z$6 by

$Z$7

so that shortcut states are averaged out rather than absorbed as predictive signal (Jiang et al., 7 Aug 2025). The architecture separates $Z$8, intended to carry task-relevant causal semantics, from $Z$9, intended to carry shortcut information, and then trains the causal predictor to remain stable across shortcut perturbations.

MCAF implements a related but distinct intervention logic. Its Causality-Guided Modal Disentanglement module decomposes language features into semantic factor $X_e$0 and bias factor $X_e$1 under the SCM

$X_e$2

A counterfactual-like intervention is then approximated by shuffling the bias factor across samples,

$X_e$3

while enforcing reconstruction, HSIC-based factor independence, and cross-modal alignment (Dong et al., 29 May 2026). The de-confounded language representation is then used as a guiding signal for dynamic routing among language, visual, and acoustic streams.

In recommendation, CaMRec combines back-door and front-door adjustment. Hidden modal confounders $X_e$4 are extracted by a dual-channel cross-modal diffusion module, discretized through a vector-quantized codebook, and used for environment stratification. Interaction bias in the user–item graph $X_e$5 is then addressed through a front-door-style proxy mechanism and causal topology reconstruction, with

$X_e$6

as the paper’s stated causal effect formula (Yang et al., 14 Oct 2025).

4. Representation learning and identifiability

A major branch of multimodal causal analysis is causal representation learning, where the objective is not only robust prediction but also recovery of interpretable latent variables. causalPIMA is a multimodal VAE with a causal prior: each modality is encoded into a Gaussian posterior, the posteriors are fused by a product of experts, and the latent space is governed by a Gaussian mixture prior whose components correspond to outcomes of latent DAG nodes. The full model is trained by a single tractable ELBO, and the learned DAG directly shapes the latent clustering (Walker et al., 2023).

The biomedical multimodal causal representation model pushes identifiability further. It assumes modality-specific latent blocks and a nonparametric latent distribution with causal relationships across modalities. It first establishes subspace identifiability, then uses structural sparsity of cross-modal causal connections to obtain component-wise identifiability:

$X_e$7

for some permutation $X_e$8 and invertible scalar transform $X_e$9 (Sun et al., 2024). This is notable because the paper explicitly contrasts its results with earlier methods that identify only shared subspaces or rely on restrictive parametric assumptions.

The latent partial causal model offers a different identifiability route. It argues that multimodal contrastive learning can recover coupled latent variables when observations are generated by invertible mechanisms

$W$0

and when the shared structure is represented by a coupling relation rather than by forcing a single shared latent. Under the hypersphere assumptions, the recovered variables are identifiable up to an orthogonal linear transform; under the convex-body assumptions, up to a permutation-like transformation (Liu et al., 2024).

This literature also clarifies a second misconception: multimodal causal analysis does not always require supervised interventions or labeled causal variables. causalPIMA is presented as fully unsupervised and uses multimodality and known physics in place of interventions, while the latent partial causal model analyzes contrastive learning theoretically rather than through explicit intervention design (Walker et al., 2023, Liu et al., 2024).

5. Representative architectures and empirical domains

The following table organizes representative multimodal causal analysis frameworks discussed in the literature.

Framework	Modalities and task	Stated causal mechanism
CodeSCM	Natural language, code, I/O examples for code generation	SCM with latent mediators and total/direct effects
MCAM	Visual and language modalities for driving video understanding	DSDAG over driving states with intervention-style reasoning
MMCI / MCAF	Text, audio, visual for sentiment analysis	Back-door adjustment; causal disentanglement and intervention
causalPIMA / biomedical CRL	Multimodal scientific and biomedical observations	Latent DAG priors; sparsity-based identifiability
CaMRec	Visual, textual, and interaction data for recommendation	Back-door and front-door adjustment
CCHMM	Bike, taxi, bus flow, and speed	SCM over latent physical concepts in a conditional HMM

Empirical findings differ by domain but are consistently framed in causal rather than purely correlational terms. In code generation, CodeSCM evaluates GPT-4 Turbo, WizardCoder-15B, and LLaMa-3-8B on HumanEval+, mMBPP+, and CoderEval, and reports that input-output examples are not just auxiliary: they exert a strong causal influence on code generation, often comparable to or stronger than some code-related prompt components; removing natural language sharply increases semantic errors; and direct effects reveal memorization and syntax-related sensitivities (Gupta et al., 7 Feb 2025).

In multimodal sentiment analysis, MMCI is evaluated on CMU-MOSI, CMU-MOSEI, CH-SIMS, and an OOD version of CMU-MOSI. The paper reports that MMCI exceeds ITHP on CMU-MOSI by about +1.3% on Acc7, achieves higher Acc2 and F1 than all baselines, lower MAE, and best Corr; on CMU-MOSEI it beats ITHP by +1.7% Acc7, +1.5% Acc2, and +1.3% F1; and on CMU-MOSI (OOD) it performs best across all reported metrics (Jiang et al., 7 Aug 2025). MCAF reports Acc-2/F1 of 86.52%/86.51% on MOSI and 86.72%/86.65% on MOSEI, with ablations showing that removing CGMD or DMIR lowers performance and with error analysis reporting standard deviation reduced by about 33% (Dong et al., 29 May 2026).

In scientific and biomedical settings, causalPIMA is tested on a synthetic circles dataset with 4096 images and a 3D-printed lattices dataset with 91 samples, learning interpretable latent factors and DAGs that the paper relates to radius, hue, shift, lattice type, and stress-strain behavior (Walker et al., 2023). The biomedical identifiability framework reports Variant MNIST results of $W$1 and $W$2 for the proposed method, exact inter-modal graph recovery with $W$3 in a showcased numerical case, and human-phenotype relations such as sleep-related latent variable $W$4 oxygen saturation, fundus latent variables $W$5 age, and right-eye fundus latent variable $W$6 left-hand grip strength, described as consistent with established biomedical research (Sun et al., 2024).

In recommendation, CaMRec is evaluated on Amazon Baby, Amazon Sports, and Amazon Clothing under leave-one-out evaluation with Recall@10, Recall@20, NDCG@10, and NDCG@20. The paper states an average improvement of 5.01% over the best baseline and a maximum gain of 7.71% in NDCG@20 on Sports, with the full model outperforming ablations without back-door adjustment, front-door adjustment, or dual-channel diffusion (Yang et al., 14 Oct 2025).

In traffic prediction, CCHMM models bike flow, taxi flow, bus flow, and traffic speed together with POI, time position, and weather. On XC-Trans and XC-Speed in Xicheng District, Beijing, with 11,753 samples and 3 hours of history for next-30-minute prediction, it reports better MAE, RMSE, and MAPE than HMM, HGCN, CCRNN, DMSTGCN, AGCRN, and DGCRN, with especially strong gains for speed prediction (Zhao et al., 2023).

In ego-vehicle driving video understanding, MCAM is evaluated on BDD-X and CoVLA. On BDD-X narration, MCAM reports BLEU-4 35.7, CIDEr 252.0, METEOR 30.5, and ROUGE 63.4; on BDD-X reasoning, BLEU-4 9.1, CIDEr 94.1, METEOR 14.0, and ROUGE 32.0. On CoVLA, MCAM reports BLEU-1 82.6, BLEU-2 79.4, BLEU-3 77.2, BLEU-4 75.3, CIDEr 275.4, METEOR 50.2, ROUGE 81.9, and SPICE 79.9, outperforming the reported baseline across all metrics (Cheng et al., 8 Jul 2025).

6. Limitations, assumptions, and interpretive cautions

The literature is explicit that multimodal causal analysis is assumption-heavy. CodeSCM states that it does not account for confounders in its current version, so its direct-effect interpretation depends on the assumption that genuinely non-spurious learning is mediated through the latent variables $W$7 and $W$8 (Gupta et al., 7 Feb 2025). The mixed graphical discovery framework assumes no feedback cycles, no unrecorded common causes, and i.i.d. samples; its benefits depend on the suitability of those constraint-based assumptions to the data (Sedgewick et al., 2017).

Representation-learning approaches make different but equally substantive commitments. causalPIMA assumes hidden factors are representable as categorical DAG nodes, notes that identifiability is not guaranteed in the exploratory setting, and states that inferred edges should be interpreted as causal hypotheses rather than definitive causal claims (Walker et al., 2023). The biomedical identifiability framework assumes that the number of latent variables per modality is known, that causal connectivity is sufficiently sparse, and that spuriously connected modalities can break the theory (Sun et al., 2024). The latent partial causal model requires invertible generative mechanisms and specific geometric assumptions on the latent space, even though its synthetic experiments suggest robustness when assumptions are partially violated (Liu et al., 2024).

Application-specific systems add their own constraints. CCHMM assumes that multimodal traffic flow can be summarized by five latent physical concepts and that the latent causal graph is acyclic and stable enough to be learned from the available condition variables and observations (Zhao et al., 2023). MCAM notes that label noise in driving datasets can affect recognition capability and identifies higher-quality datasets and deeper causal interactions among objects in video as future work (Cheng et al., 8 Jul 2025). MCAF reports strong classification metrics but, in the provided methodological excerpt, does not give explicit diffusion equations for its Conditional Diffusion Denoising Module, so the denoising stage is described functionally rather than through a complete generative specification (Dong et al., 29 May 2026).

Taken together, these caveats imply that “causal” in MCAM-related research usually means one of three things: explicit causal structure in the model, causal adjustment or intervention during learning, or identifiability analysis for latent variables. It does not, by itself, imply definitive real-world causal discovery. A plausible implication is that the most durable contribution of MCAM-style work is methodological: it supplies structured ways to separate multimodal signal from shortcut dependence, while leaving the strength of any substantive causal claim contingent on the assumptions of the underlying model.