Abstract
The advent of the Transformer has led to the development of large language models (LLM), which appear to demonstrate human-like capabilities. To assess the generality of this class of models and a variety of other base neural network architectures to multimodal domains, we evaluated and compared their capacity for multimodal generalization. We introduce a multimodal question-answer benchmark to evaluate three specific types of out-of-distribution (OOD) generalization performance: distractor generalization (generalization in the presence of distractors), systematic compositional generalization (generalization to new task permutations), and productive compositional generalization (generalization to more complex tasks structures). We found that across model architectures (e.g., RNNs, Transformers, Perceivers, etc.), models with multiple attention layers, or models that leveraged cross-attention mechanisms between input domains, fared better. Our positive results demonstrate that for multimodal distractor and systematic generalization, either cross-modal attention or models with deeper attention layers are key architectural features required to integrate multimodal inputs. On the other hand, neither of these architectural features led to productive generalization, suggesting fundamental limitations of existing architectures for specific types of multimodal generalization. These results demonstrate the strengths and limitations of specific architectural components underlying modern neural models for multimodal reasoning. Finally, we provide Generic COG (gCOG), a configurable benchmark with several multimodal generalization splits, for future studies to explore.
Introduction
Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their ability to perform true multimodal reasoning remains a subject of debate. Critics often characterize these models as sophisticated interpolators that excel at pattern matching within their training distribution but struggle with genuine out-of-distribution (OOD) generalization. This study investigates how specific architectural choices influence a model's capacity for multimodal reasoning across three distinct OOD challenges: distractor generalization, systematic compositional generalization, and productive compositional generalization.
Methods
The researchers evaluated six distinct encoder-only architectures to identify the structural features necessary for robust multimodal reasoning:
- Baselines: Standard Recurrent Neural Networks (RNNs) and Gated Recurrent Units (GRUs).
- Single-stream Transformer (SSTfmr): A model that concatenates text and visual inputs into a single sequence.
- Dual-stream Transformer (DSTfmr): Processes text and images in parallel streams before merging them via a multi-layer perceptron (MLP).
- Cross-Attention Transformer (CrossAttn): Employs text features to query visual features, facilitating explicit cross-modal interaction.
- Perceiver: A latent-space architecture utilizing cross-attention to process inputs in a modality-agnostic manner.
To analyze the internal mechanisms of these models, the authors utilized Representational Similarity Analysis (RSA), a technique borrowed from neuroscience, to compare the models' internal representations with the underlying task structure.
The gCOG Benchmark
To facilitate this evaluation, the authors introduced Generic COG (gCOG), a configurable benchmark designed to test reasoning independently of low-level visual processing. Unlike traditional visual QA datasets, gCOG uses categorical tokens (e.g., "red", "square") instead of raw pixels. This abstraction ensures that performance reflects reasoning capacity rather than the quality of a vision encoder. The benchmark supports "task trees" of varying depths using conditional logic (IF-THEN-ELSE), making it ideal for testing productive generalization.
Key Results and Insights
The study revealed a significant performance gap between architectures when faced with OOD scenarios:
- Cross-Attention as a Noise Filter: Models utilizing explicit cross-modal attention (CrossAttn and Perceiver) were remarkably resilient to distractors, maintaining near 100% accuracy even as the number of distractors increased significantly. RSA confirmed that these models' internal representations remained better aligned with task stimuli compared to RNNs.
- Systematicity vs. Productivity: While models like the Perceiver excelled at systematic generalization (combining known concepts in new ways), achieving 65.7% accuracy on depth-3 task trees, all models failed at productive generalization. When trained on depth-1 and depth-3 trees and tested on depth-5 or depth-7 trees, performance dropped to chance levels.
- Scaling Limitations: Increasing model size (layers or attention heads) improved performance on noisy or novel combinations but failed to enable algorithmic scaling. This suggests that simply increasing parameters does not solve the fundamental challenge of recursive logic.
Conclusion
The findings highlight a "Productivity Gap" in current neural architectures. While cross-modal attention and deeper layers facilitate noise filtering and concept composition, they do not provide the recursive mechanisms necessary for scaling logic to more complex task structures. Future research may need to explore neuro-symbolic hybrids or novel algorithmic layers capable of learning the underlying syntax of reasoning tasks independently of their scale.
Summary updated