Beyond Attention Mechanisms: Geometric Perturbation Analysis for Model-Agnostic Token Attribution

Large Language Models have achieved remarkable performance across natural language processing tasks, yet their internal decision making processes remain largely opaque. This opacity poses significant challenges for deployment in high stakes domains where auditability and transparency are non-negotiable requirements. While attention visualization techniques have provided some insight into model behavior, these approaches typically require deep architectural access, specific model structures, and computationally expensive backpropagation. The paper "VISTA: Visualization of Token Attribution via Efficient Analysis" introduces a compelling alternative: a lightweight, perturbation based methodology that quantifies token importance without ever examining internal model weights or gradients.

The Mechanics of Geometric Perturbation

At the core of VISTA lies an elegant geometric framework built upon distributional semantics. The method utilizes GloVe (Global Vectors for Word Representation) embeddings to map tokens into a 50-dimensional vector space, where semantic relationships manifest as geometric relationships. Rather than probing the model's internal attention mechanisms, VISTA treats the model as a black box and analyzes how the semantic representation of the input prompt changes when individual tokens are systematically removed.

The aggregate semantic representation of a prompt is computed as the vector sum of its constituent token embeddings. This additive approach, while seemingly simplistic, provides a tractable foundation for perturbation analysis. For each token in the input sequence, the method constructs a perturbed representation by subtracting that token's embedding from the aggregate vector. The semantic impact of this removal is then quantified through three complementary analytical matrices.

First, the Angular Deviation Matrix captures shifts in semantic direction by measuring the cosine similarity between the original and perturbed aggregate vectors. This metric reveals whether the removal of a token fundamentally alters the orientation of the semantic representation in the embedding space. Second, the Magnitude Deviation Matrix quantifies changes in semantic intensity through L2 norm comparisons, indicating whether the removal reduces or amplifies the overall semantic strength of the representation. Third, the Dimensional Importance Matrix evaluates contributions across individual vector dimensions, providing granular insight into which specific semantic features each token activates.

These three matrices combine to form a composite importance score that reflects not merely the presence of a token, but its multifaceted contribution to the prompt's collective meaning. By filtering out common function words that contribute minimally across these dimensions, VISTA highlights only the semantically substantial tokens, generating relevance maps that align with human intuition about word importance.

Efficiency and Architectural Independence

The computational advantages of VISTA become apparent when contrasted with gradient based attribution methods such as Integrated Gradients or SmoothGrad. Traditional approaches require backpropagation through the model's computation graph, effectively doubling GPU memory usage and introducing significant latency. VISTA operates entirely through forward passes on embedding representations, requiring no access to internal attention weights, layer activations, or gradient flows.

This architecture independent property distinguishes VISTA from transformer specific visualization techniques like Attention Rollout or BertViz. The method functions equally well across model families, including recurrent architectures, convolutional networks, or any system capable of producing vector representations of text. For production environments where inference costs must be minimized, the elimination of gradient computation represents a substantial practical benefit.

Furthermore, the mathematical transparency of the three matrix framework offers interpretability that complex neural attention patterns often lack. While attention weights can be difficult to interpret due to their distributed nature across multiple heads and layers, the geometric measures employed by VISTA provide concrete, human understandable quantities: angular shifts indicate directional semantic changes, magnitude differences reflect intensity variations, and dimensional analysis reveals feature specific contributions.

Critical Perspectives and Limitations

Despite its elegant formulation, VISTA's reliance on static GloVe embeddings introduces significant constraints that