# Feature Alignment Determines Fusion Strategy: A Comparative Study of Cross-Attention and Concatenation in Multimodal Learning

arXiv: https://arxiv.org/abs/2606.01207 · HTML: https://mlsanity.com/paper/2606.01207/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 26 of its 32 references that are in the map.
- First posted on arXiv 31 May 2026.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)

## Builds on (most cited first, 26 of 26)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://mlsanity.com/paper/2010.11929/) (2020)
- [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://mlsanity.com/paper/2103.14030/) (2021)
- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://mlsanity.com/paper/2301.12597/) (2023)
- [Training Compute-Optimal Large Language Models](https://mlsanity.com/paper/2203.15556/) (2022)
- [InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks](https://mlsanity.com/paper/2312.14238/) (2023)
- [VQA: Visual Question Answering](https://mlsanity.com/paper/1505.00468/) (2015)
- [Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond](https://mlsanity.com/paper/2308.12966/) (2023)
- [ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks](https://mlsanity.com/paper/1908.02265/) (2019)
- [Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention](https://mlsanity.com/paper/2006.16236/) (2020)
- [Big Bird: Transformers for Longer Sequences](https://mlsanity.com/paper/2007.14062/) (2020)
- [Rethinking Attention with Performers](https://mlsanity.com/paper/2009.14794/) (2020)
- [LXMERT: Learning Cross-Modality Encoder Representations from Transformers](https://mlsanity.com/paper/1908.07490/) (2019)
- [Generating Long Sequences with Sparse Transformers](https://mlsanity.com/paper/1904.10509/) (2019)
- [Multimodal Machine Learning: A Survey and Taxonomy](https://mlsanity.com/paper/1705.09406/) (2017)
- [Show and Tell: A Neural Image Caption Generator](https://mlsanity.com/paper/1411.4555/) (2014)
- [EVA-CLIP: Improved Training Techniques for CLIP at Scale](https://mlsanity.com/paper/2303.15389/) (2023)
- [EVA: Exploring the Limits of Masked Visual Representation Learning at Scale](https://mlsanity.com/paper/2211.07636/) (2022)
- [Stacked Cross Attention for Image-Text Matching](https://mlsanity.com/paper/1803.08024/) (2018)
- [Scaling Laws for Transfer](https://mlsanity.com/paper/2102.01293/) (2021)
- [Multimodal Deep Learning](https://mlsanity.com/paper/2301.04856/) (2023)
- [Unifying Visual-Semantic Embeddings with Multimodal Neural Language Models](https://mlsanity.com/paper/1411.2539/) (2014)
- [Learning Deep Structure-Preserving Image-Text Embeddings](https://mlsanity.com/paper/1511.06078/) (2015)

## Abstract (arXiv)

The choice between cross-attention and concatenation for multimodal fusion remains governed by practitioner intuition rather than principled understanding. In this paper, we demonstrate that feature alignment quality, not data scale alone, is the primary determinant of which fusion strategy excels. Through controlled experiments on Flickr8k using two feature extraction backbones (ResNet18 and CLIP ViT-B/32), we show that concatenation outperforms cross-attention by 4.1-5.1 percentage points across all tested scales (2048-16384 samples) when features are pre-aligned by a vision-language pretraining objective. We provide a theoretical explanation grounded in sample complexity analysis: concatenation requires O(d_v + d_t) samples to learn its fusion projection, while cross-attention requires O(d_v * d_t) samples to learn bilinear attention weights, over 256 times as many for 512-dimensional CLIP features. When features are already aligned, the approximation error gap between the two methods vanishes, and concatenation's sample efficiency dominates at all practical dataset sizes. An alignment degradation study confirms a monotonic trend: as feature alignment degrades, concatenation's advantage grows from 1.3% to 2.8%. These findings provide a principled decision framework for fusion method selection in multimodal systems, with direct implications for the design of Multimodal Large Language Models.