TLDR: A new research paper by Luo et al. reveals that ‘ViT attention sinks’ – high-norm visual tokens from the Vision Transformer (ViT) – are crucial for Large Vision Language Models (LVLMs) to understand and reason with images. Contrary to previous beliefs, these sinks encapsulate high-level semantic concepts and propagate distinctly into the Large Language Model (LLM). The authors propose two methods to leverage them: a training-free ‘sink-to-the-front’ strategy and a training-based ‘DIYSink’ framework with dual MLP projectors and dynamic token selection. Both approaches significantly improve LVLM performance across various visual reasoning tasks, highlighting the untapped potential of these visual attention sinks.
Large Vision Language Models (LVLMs) have become incredibly powerful, allowing computers to understand and reason using both images and text. These models typically have two main parts: a Vision Transformer (ViT) that acts as the ‘eyes’ to process visual information, and a Large Language Model (LLM) that serves as the ‘brain’ for reasoning and generating responses.
However, a key challenge has been understanding which parts of the visual information are most crucial for the LLM to perform its tasks effectively, and how these visual signals travel from the ViT to the LLM. While much research has focused on ‘attention sinks’ within the LLM – tokens that receive a lot of attention but might not carry much semantic meaning – the visual side, specifically ‘ViT attention sinks,’ has been largely overlooked.
A new research paper, titled “TOSINK ORNOT TOSINK: VISUALINFORMATION PATHWAYS INLARGEVISION-LANGUAGEMODELS” by Jiayun Luo, Wan-Cyuan Fan, Lyuyang Wang, Xiangteng He, Tanzila Rahman, Purang Abolmaesumi, and Leonid Sigal, sheds light on these often-ignored ViT attention sinks. The authors found that these high-norm visual tokens from the ViT are not just noise; they actually capture important, high-level semantic concepts from images. This means they are vital for the LLM to understand and reason more effectively.
The study reveals several key insights. Firstly, there’s a clear connection: visual tokens with higher ‘norms’ (a measure of their strength) in the ViT are more likely to become attention sinks in the LLM, receiving significantly more attention. Secondly, these ViT sinks propagate into the LLM as distinct entities, activating different hidden dimensions compared to sinks that originate purely within the LLM. This distinction is crucial because it suggests they carry different types of information.
Through detailed analysis, including ‘relevance maps’ and ‘word distributions,’ the researchers demonstrated that ViT sinks capture broad, high-level contextual features, such as the main object or overall scene illumination. This contrasts with non-sink tokens, which tend to focus on local details. The implication is that ViT sinks are particularly useful for tasks requiring a holistic understanding of an image, but might interfere with tasks that demand fine-grained, localized visual processing.
To better utilize these powerful ViT sinks, the researchers propose two innovative approaches:
Training-Free Approach: Sink-to-the-Front
This simple yet effective strategy involves repositioning the ViT sink tokens to the very beginning of the visual token sequence before they are fed into the LLM. This leverages the LLM’s ‘causal attention bias,’ where earlier tokens have a greater influence on subsequent processing. This method requires no additional training and can be applied to existing LVLMs, showing improvements in tasks requiring high-level understanding.
Also Read:
- Unifying Attention Sinks and Compression Valleys in LLMs: A Three-Phase Computational Model
- Mimicking Human Cognition: Dynamic Gating for Efficient Vision-Language AI
Training-Based Approach: DIYSink
For more substantial and dynamic improvements, the authors introduce DIYSink, a framework that involves training. It has two main components:
-
Dual-MLP Projection Layers: Instead of a single connector trying to project both sink and non-sink tokens, DIYSink uses separate MLP (Multi-Layer Perceptron) layers for each. This allows each projector to specialize in converting its specific type of token into an embedding space that the LLM can better interpret, preventing their unique characteristics from being muddled.
-
Dynamic Token Selection Modules: This intelligent component helps the LVLM decide whether to use sink tokens, non-sink tokens, or both, based on the specific image and question. This selection can be done through a ‘Chain-of-Thought’ (CoT) routing, which makes a hard decision, or a ‘Reweighting MLP,’ a lightweight module that learns to softly balance the contribution of sink and non-sink tokens.
The DIYSink framework has been rigorously validated across various LVLMs and benchmarks, demonstrating consistent and substantial performance gains in tasks like mathematical problem-solving, logical inference, and geometric understanding. For instance, it showed significant improvements in MME code reasoning and MathVista’s algebraic and geometric reasoning tasks.
In conclusion, this research challenges the conventional view of attention sinks, revealing the crucial role of ViT sinks in encoding high-level visual context within LVLMs. By explicitly leveraging these tokens through methods like ‘sink-to-the-front’ or the comprehensive DIYSink framework, models can achieve enhanced visual reasoning capabilities, unlocking new potential for understanding and interacting with multimodal information.
