New method decodes what's happening inside large language models using symbolic vector algebra

Understanding how large language models work internally remains one of artificial intelligence's fundamental challenges. While these systems demonstrate impressive capabilities, their decision-making processes remain largely opaque. Researchers have now introduced a new approach called the Hyperdimensional Probe that aims to shed light on this mysterious process.

Bridging different interpretability approaches#

Existing methods for understanding LLMs fall into two broad camps: those that examine input representations and those that analyze output predictions. Traditional supervised probes focus on extracting features from the model's intermediate layers, while Sparse Autoencoders (SAEs) create sparse representations of these internal states. Meanwhile, other techniques inspect the final output distribution to understand what the model is deciding. Each approach reveals something useful but none offers a complete picture.

The Hyperdimensional Probe attempts to unify these perspectives. By drawing on Vector Symbolic Architectures (VSAs)—a branch of AI dealing with high-dimensional vector representations—the method combines the best aspects of existing techniques. It preserves the top-down feature extraction strength of supervised probes, the sparsity advantages of SAEs, and the output-focused insights of logit-based methods.

How the approach works#

The core innovation lies in merging supervised learning with hypervector algebra, allowing researchers to extract meaningful semantic concepts directly from the model's vector space. This hybrid design operates simultaneously on both input features and output predictions, creating a more holistic picture of how the model processes information.

Testing across different language models, embedding dimensions, and configurations demonstrates that the approach consistently uncovers concept-oriented insights. Notably, VSA-based probing overcomes a fundamental limitation of logit-based methods—they are restricted by the model's vocabulary of available tokens. Additionally, it produces cleaner interpretability results compared to SAEs, particularly when working with bounded conceptual feature spaces.

Practical validation#

The researchers validated their method on two distinct scenarios: completion tasks and question-answering text generation. In both cases, the technique revealed how semantic information flows through the model during inference, offering insights that neither input-only nor output-only methods could provide alone.