Research map / Label-efficient and robust learning / Domain adaptation and generalization

Are We Making Progress in Multimodal Domain Generalization? A Comprehensive Benchmark Study

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Optimization-based domain generalization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MMDG-Bench, a standardized benchmark of nine MMDG methods (plus Oracle) across six datasets, three task families and six modality combinations, finding that specialized MMDG methods give only marginal gains over ERM and that robustness and trustworthiness remain weak.

Problem

Multimodal domain generalization (MMDG) reported gains may be artifacts of inconsistent evaluation protocols, datasets and settings, and existing benchmarks focus mainly on action recognition while ignoring corruptions, missing modalities and trustworthiness.

Approach

A unified benchmark that fixes data splits, backbones, hyperparameter search (default plus 10 random-search trials, three seeds) and training-domain validation across action recognition, fault diagnosis and sentiment analysis, and adds corruption, missing-modality, misclassification and OOD evaluations.

Key results

• On HAC (V+A), MOOSA reaches 70.95% mean accuracy versus 68.93% for ERM, and the Oracle reaches 92.81%, a gap of nearly 22 points.
• On sentiment analysis, the best specialized method (MOOSA, 66.60% ACC2) outperforms ERM (65.63%) by less than one point.
• CLIP ViT-L/14 zero-shot reaches 91.08% mean accuracy on HAC, above the best trained MMDG method, while Qwen2.5-VL 3B reaches only 39.71% on EPIC-Kitchens.

Surprising findings

• Zero-shot CLIP ViT-L/14 outperforms all trained MMDG methods on HAC, while foundation models underperform on fine-grained EPIC-Kitchens.
• Trimodal V+A+F does not reliably beat the best bimodal configuration; MOOSA declines from 71.29% to 70.17% when adding flow to V+F on HAC.

Limitations

• Only nine methods, three task families and six datasets are covered, limiting the scope of conclusions.
• Ten random-search trials cannot guarantee each method reaches its global optimum, and shared backbones may favor some methods.
• Foundation-model comparison is limited because these models do not natively support flow, raw audio or vibration.

Experimental setup

Datasets: EPIC-Kitchens HAC HUST Motor CMU-MOSI CMU-MOSEI CH-SIMS

Benchmarks: EPIC-Kitchens HAC HUST Motor CMU-MOSI CMU-MOSEI CH-SIMS

Metrics: top-1 accuracy MAE F1 ACC2 AURC AUROC FPR95

Baselines: ERM RNA-Net SimMMDG MOOSA CMRF NEL JAT MBCD GMP Oracle

Builds on: DomainBed SimMMDG MMAction2 BERT

Closest papers by meaning

• Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.87)
• MMDiff: Multimodal Model Diffing for Feature Discovery and Control (NeurIPS 2026 · similarity 0.87)
• Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.87)
• On Path to Multimodal Generalist: General-Level and General-Bench (ICML 2025 · similarity 0.86)
• Cross-modal Representation Flattening for Multi-modal Domain Generalization (NeurIPS 2024 · similarity 0.86)
• LLaVA-Critic: Learning to Evaluate Multimodal Models (CVPR 2025 · similarity 0.86)