Aligning the True Semantics: Constrained Decoupling and Distribution Sampling for Cross-Modal Alignment

Conference: AAAI 2026 arXiv: 2603.05566v1 Code: None Area: Multimodal VLM Keywords: Cross-modal alignment, embedding decoupling, distribution sampling, image-text retrieval, contrastive learning

TL;DR

This paper proposes the CDDS algorithm, which decouples embeddings into semantic and modality components via a dual-path UNet, and employs a distribution sampling method to achieve cross-modal semantic alignment indirectly, avoiding distribution distortion caused by directly adjusting embeddings. CDDS surpasses the state of the art by 6.6%–14.2% on Flickr30K and MS-COCO.

Background & Motivation

Most existing cross-modal alignment methods achieve semantic consistency by directly pulling image and text embeddings closer through contrastive learning. However, embeddings contain not only semantic information but also modality-specific information (e.g., color distributions in images, syntactic structures in text, and training noise). Such non-semantic information cannot be matched across modalities, and directly aligning embeddings introduces semantic bias, leading to the situation where "embedding consistency ≠ semantic consistency." Intuitively, one could decouple embeddings into semantic and modality components and align only the semantic part. Yet this faces two challenges: (1) semantic and modality information are intricately entangled, lacking clear decoupling criteria; and (2) embeddings from different modalities are constructed differently, making cosine-similarity-based cross-modal interaction theoretically unjustified, while forcibly adjusting embeddings distorts the original distribution.

Core Problem

How to, in cross-modal alignment: (1) effectively decouple embeddings into semantic and modality components while ensuring decoupling validity and information completeness; and (2) achieve true semantic alignment without distorting the original distribution, thereby avoiding the bias and information loss introduced by the modality gap.

Method

Overall Architecture

CDDS adopts a fine-grained approach consisting of three components: constrained decoupling, semantic component constraint (semantic alignment via distribution sampling), modality component constraint, and information completeness constraint. After decoupling, only the semantic components participate in cross-modal alignment.

Key Designs

1. Dual-path UNet decoupling architecture: A shared encoder (ViT) maps embeddings into a high-dimensional space. \(z\) groups of Gaussian noise perturbations are then introduced to obtain multiple perturbed representations, enhancing robustness. A semantic decoder and a modality decoder separately extract the semantic component and the modality component from the perturbed representations using UNet-style skip connections to preserve features at each level. The final robust semantic components \(V^s\)/\(T^s\) and modality components \(V^m\)/\(T^m\) are obtained by averaging across the \(z\) decoded results.

2. Related Semantics Identification: Distributions \(C^v\) and \(C^t\) are constructed from the column-wise features of the semantic components. KL divergence is used to measure cross-modal distributional correlation, yielding a matrix \(S\). An adaptive soft-threshold sparsification algorithm is proposed: the threshold \(k_i^v\) is determined using the mean and standard deviation of the conditional probability weighted by a learnable parameter \(\alpha_i\), filtering weakly correlated distributions and retaining strongly correlated distribution pairs that describe related semantics.

3. Distribution Sampling: After identifying related semantics, rather than directly pulling semantic components closer via contrastive learning, the method samples positionally from image distributions conditioned on strongly correlated text distributions to construct cross-modal semantic components (x-semantic components) \(V^x\)/\(T^x\). \(V^x\) expresses the semantics of an image in the descriptive form of the text modality, effectively bridging the modality gap. Cross-modal semantic alignment is achieved indirectly by constraining the consistency between \(V^x\) and \(V^s\), without adjusting the original distribution.

Loss & Training

The total loss is \(\mathcal{L} = \alpha_s \mathcal{L}_s + \alpha_m \mathcal{L}_m + \alpha_f \mathcal{L}_f + (1 - \alpha_f) \mathcal{L}_x\), comprising four terms: - Semantic consistency \(\mathcal{L}_s\): A contrastive loss that pulls matched pairs of semantic component \(V^s\) and the corresponding x-semantic \(V^x\) closer while pushing non-matched pairs apart; symmetrically applied to the text side. - Modality consistency \(\mathcal{L}_m\): KL divergence is used to constrain the modality component distributions of all patches/words within the same modality to remain consistent. - Information completeness \(\mathcal{L}_f\): The semantic component plus the modality component should reconstruct the original embedding (L2 loss), with \(w_m\) and \(w_s\) as learnable weights. - X-semantic completeness \(\mathcal{L}_x\): The modality component plus the x-semantic component should also reconstruct the original embedding, complementing \(\mathcal{L}_f\), with \(\alpha_f\) controlling the balance between the two.

Training is conducted for 25 epochs using the AdamW optimizer with a learning rate of 2e-4 and batch size of 64. The encoder and decoder each have 2 layers, the feature dimension is 512, and training is performed on NVIDIA L40 GPUs.

Key Experimental Results

Dataset	Metric	Ours (CDDS)	Prev. SOTA (LAPS)	Gain
Flickr30K (Swin-384)	rSum	552.5	545.3	+7.2
MS-COCO 1K (Swin-384)	rSum	548.6	544.1	+4.5
MS-COCO 5K (Swin-384)	rSum	472.1	470.1	+2.0
Flickr30K (ViT-224)	rSum	510.6	507.3	+3.3
MS-COCO 5K (ViT-224)	rSum	437.8	434.4	+3.4
Flickr30K (CLIP ViT-L)	I→T R@1	95.2	94.6	+0.6

CDDS consistently outperforms the state of the art across all four backbone configurations (ViT-224/384, Swin-224/384) and when extended to CLIP, also significantly surpasses VLP models (VILT, SOHO, ALBEF, BLIP, etc.).

Ablation Study

• Removing the decoupling architecture (w/o Dec.): Performance drops by 4.6%, confirming that decoupling is the core contribution.
• Removing the modality constraint (w/o Mod.): Performance drops by 0.9%, indicating that the modality consistency constraint provides auxiliary benefit.
• Removing information completeness (w/o Int.): Performance drops by 6.7%, demonstrating that the information completeness constraint is the most critical component.
• Removing Gaussian noise (w/o Gau.): A moderate performance drop is observed, indicating that noise perturbation enhances decoupling robustness.
• Removing distribution sampling (w/o Sam.): Replacing distribution sampling with contrastive learning leads to a performance drop, validating the superiority of the distribution sampling approach.
• Applying the distribution sampling module to other models (VSE++, SCAN, SGR, CHAN, LAPS) yields consistent gains of 0.4%–1.1%, demonstrating its generalizability.

Highlights & Insights

• The embedding decoupling perspective offers a principled approach to cross-modal alignment, transforming "aligning embeddings" into "aligning semantic components."
• The distribution sampling method is particularly elegant: by sampling positionally from the counterpart modality's distribution to construct x-semantics, alignment is achieved indirectly without distorting the original distribution, which is theoretically more principled than direct contrastive learning.
• The three-way constraint (semantic consistency, modality consistency, information completeness) forms a closed loop, ensuring both decoupling validity and lossless information preservation.
• The adaptive soft-threshold sparsification avoids the crude truncation of fixed top-\(k\) selection.
• The method is plug-and-play: the distribution sampling module can be applied to other cross-modal methods and consistently yields improvements.

Limitations & Future Work

• High computational cost: Related Semantics Identification (Eq. 5) must be executed per batch with complexity \(O(N^2)\); precomputing over the full dataset or using random sampling both lead to notable performance degradation, as reported by the authors.
• Validation is limited to image-text retrieval; other cross-modal tasks (e.g., image captioning, VQA, text-to-image generation) remain unexplored.
• The decoupled semantic/modality components lack interpretability analysis (only shallow t-SNE visualizations are provided).
• No comparison is made with recent large-scale pretrained models (e.g., BLIP-2, CoCa) under equivalent conditions.

Related Work & Insights

• Compared to coarse-grained methods (VSE++, GPO, DIAS): CDDS does not merely align global embeddings but aligns semantic components after decoupling, thereby avoiding interference from modality-specific noise.
• Compared to fine-grained methods (SCAN, CAAN, NAAF, CHAN): These methods still assume that corresponding columns of embeddings from different modalities describe the same semantics, whereas CDDS identifies related semantics via distributional correlation before alignment.
• Compared to LAPS (previous state of the art): LAPS enhances robustness through spatial relationship modeling but still performs direct embedding alignment; CDDS improves along both the decoupling and indirect alignment dimensions, outperforming LAPS across all configurations.

The decoupling paradigm is transferable to other multimodal tasks: decomposing embeddings into "content" and "style/modality" components is a general paradigm extensible to audio-visual, 3D-language, and other cross-modal settings. The indirect alignment idea from distribution sampling is also worth borrowing: when aligning representations from two different spaces, "expressing one's semantics in the descriptive form of the counterpart space" is more gentle than directly pulling them closer. The adaptive soft-threshold sparsification can be applied to other tasks requiring cross-domain correspondence discovery.

Rating

• Novelty: ⭐⭐⭐⭐ The combination of decoupling and distribution sampling is innovative, though embedding decoupling itself is not an entirely new concept.
• Experimental Thoroughness: ⭐⭐⭐⭐ Four backbone configurations, two datasets, CLIP extension, detailed ablations, and generalizability validation are provided, though coverage of downstream tasks is limited.
• Writing Quality: ⭐⭐⭐⭐ The overall logic is clear and the mathematical derivations are complete, though the notation is occasionally dense.
• Value: ⭐⭐⭐⭐ The plug-and-play distribution sampling module offers practical reference value for the cross-modal alignment community.

Rating

• Novelty: ⭐⭐⭐⭐
• Experimental Thoroughness: ⭐⭐⭐⭐
• Writing Quality: ⭐⭐⭐⭐
• Value to Me: ⭐⭐⭐

Related Papers

• [AAAI 2026] Bridging Modalities via Progressive Re-alignment for Multimodal Test-Time Adaptation (BriMPR)
• [ICCV 2025] Causal Disentanglement and Cross-Modal Alignment for Enhanced Few-Shot Learning
• [AAAI 2026] InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration
• [NeurIPS 2025] Aligning by Misaligning: Boundary-aware Curriculum Learning for Multimodal Alignment
• [AAAI 2026] Rethinking Visual Token Reduction in LVLMs under Cross-Modal Misalignment