Research map / Vision-language and multimodal / Cross-modal learning

CyIN: Cyclic Informative Latent Space for Bridging Complete and Incomplete Multimodal Learning

NeurIPS 2025 · Vision-language and multimodal › Cross-modal learning · approach: Multimodal fusion with missing modalities · 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.

Proposes CyIN, which builds a cyclic token- and label-level information-bottleneck latent space and uses cross-modal cyclic translation to reconstruct missing modalities, so one unified model handles both complete and incomplete multimodal inputs.

Problem

Multimodal models trained on fully paired inputs degrade sharply when modalities are missing at inference, and most prior missing-modality methods either sacrifice complete-input performance or need separate models per missing pattern.

Approach

Combine token-level and label-level variational information bottlenecks applied cyclically across modalities to form a purified informative latent space, then reconstruct missing modality latents via forward and reverse (cyclic) translation with cascaded residual autoencoders, letting a single model serve complete and incomplete inputs.

Key results

• On complete MOSI input, CyIN reaches 48.0 Acc7 and 0.712 MAE, the best among compared methods listed in the table.
• Under random missing with MR=0.7 on MOSI, CyIN reaches 28.0 Acc7 and 1.117 MAE, versus 23.4 Acc7 for the ablation without translated latents.
• CyIN reports 123.49M parameters and 22.41s inference per iteration versus 144.34M and 70.62s for GCNet on MOSI.

Measured effects reported in the paper

Surprising findings

• Removing cyclic interaction slightly raises F1 under MR=0.7 on MOSI (67.8 vs 65.9), so the ablation is not uniformly harmful.

Limitations

• Treats all modalities as equally important, ignoring task-dependent imbalance across modalities.
• Relies on CRA-style translators that could be replaced by stronger generative models such as diffusion or flow models.
• Evaluated only on four affective computing datasets, without training on broader multimodal understanding data or LLM integration.

Experimental setup

Datasets: MOSI MOSEI IEMOCAP MELD

Benchmarks: MOSI MOSEI IEMOCAP MELD

Metrics: Acc7 F1 MAE Corr Acc wF1

Baselines: CCA DCCA DCCAE CPM-Net CRA MCTN MMIN GCNet IMDer LNLN

Builds on: variational information bottleneck cascaded residual autoencoder multimodal transformer ImageBind

Closest papers by meaning

• Fusion or Confusion? Multimodal Complexity Is Not All You Need (NeurIPS 2026 · similarity 0.88)
• Beyond Modality Fusion: Deep Ensembles for Multimodal Classification (NeurIPS 2026 · similarity 0.88)
• When to Align, When to Predict: A Phase Diagram for Multimodal Learning (NeurIPS 2026 · similarity 0.87)
• Balancing Multimodal Training Through Game-Theoretic Regularization (NeurIPS 2025 · similarity 0.87)
• Continual Multimodal Contrastive Learning (NeurIPS 2025 · similarity 0.87)
• MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.87)