Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures".

Jane: Here's the gist of the paper. The gist The Hyperdimensional Probe introduces a novel hybrid supervised probe that combines symbolic representations with neural probing to unify prior methods and uncover concept-oriented insights into Large Language Model vector spaces.

Jane: The approach leverages Vector Symbolic Architectures (VSAs) and hypervector algebra to map LLM embeddings into a bounded proxy space, overcoming constraints found in existing methods Leveraging Vector Symbolic Architectures (VSAs) and hypervector algebra, it unifies prior methods The Hyperdimensional Probe introduces a hybrid supervised probe that combines symbolic representations with neural probing to unify prior methods and uncover concept-oriented insights into Large Language Model vector spaces.

Title and authors: Tom: The paper we’re looking at today is titled "Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures", and it involves authors like Marco Bronzini, Fondazione Bruno Kessler, Trento, Italy, Carlo Nicolini Ipazia S.p.A., Milan, Italy and several others.

Jane: So what’s the main point of this title? It suggests they are using these vector symbolic architectures to decode the hidden representations within Large Language Models in a structured way.

Lu: The authors are trying to unify prior methods by combining symbolic representations with neural probing. They want to move beyond just looking at input features or just output distributions and get a full understanding of the LLM vector space structure.

Meng: So it’s aiming for a unified interpretability framework, combining top-down supervised probing with sparsity-driven proxy spaces from SAEs, plus output-oriented logit investigation. That sounds like they’re trying to get the best of all worlds.

Lalam: They are using Vector Symbolic Architectures and hypervector algebra to map LLM embeddings into a bounded proxy space.

Tom: Essentially, they leverage these tools to uncover concept-oriented insights that existing methods struggle with because of the constraints on defining latent features.

Jane: That means they are trying to get a cleaner, more structured view of what the model is learning internally compared to just looking at raw activations or simple feature maps.

The paper's summary: Tom: To summarize what this paper does, the Hyperdimensional Probe uses a three-stage pipeline. First, they process the embeddings via an ingestion procedure that includes dimensionality reduction steps like k-means clustering and sum pooling to reduce redundancy.

Jane: Then they map those processed embeddings into a controlled proxy space using that neural VSA encoder M we talked about earlier, which transforms the LLM vector space into bipolar hypervectors.

Lu: After the mapping, they perform concept extraction by querying this proxy space using the unbinding operation from VSA algebra. This finally allows them to extract concepts from the encodings and use those concepts for deeper analysis on model’s internal state and its outputs.

Meng: So it takes raw embeddings, reduces them down, turns them into a constrained symbolic space via neural mapping, and then extracts concepts using algebraic operations. That’s a lot of steps before you get to the actual insight.

Lalam: The paper also shows how this works in practice by studying concept-driven dynamics across two distinct experimental settings: completing analogy-style inputs and text generation for Question-Answering tasks.

Tom: They are specifically looking at whether language models represent all in-context concepts, key–value associations, or only the output concept when answering questions.

The paper's improvements: Jane: The authors point out a few key improvements they made over prior work. One is that it allows for input-aligned feature extraction specifically for output-oriented interpretability.

Lu: That means it’s designed to uncover concept-oriented insights into LLM inference across two distinct scenarios: input-completion tasks and QA-focused text generation.

Tom: And they also claim this approach avoids the vocabulary constraints that you often find in logit methods, which is something we see in Section five point three of the paper <ref:2509.25045#pg1>.

Meng: They also state that it provides cleaner results than analyses based on Sparse Autoencoders, like SAE-based analyses, especially when you are working within bounded conceptual frameworks where a predefined and constrained feature space is necessary.

Lalam: Another improvement they highlight is effective compression of LLM embeddings, which enables probing the full residual stream without needing to select an optimal hidden layer.

Tom: So it’s offering a method that combines the top-down interpretability of supervised probes with the sparsity-driven proxy space of SAEs and the output-oriented investigation of logit methods.

Conclusion: Jane: So wrapping up, what’s the big picture here? The Hyperdimensional Probe provides empirical evidence that we can faithfully capture the latent features of LLM’s neural embeddings using VSA-based representations.

Lu: It enables a unified, concept-oriented analysis of features aligned with both inputs and outputs, integrating the complementary perspectives of prior neural probing approaches to uncover non-trivial insights into LLM inference.

Tom: The paper shows that VSA-based probing is particularly effective in experimental settings that require a constrained or structured conceptual feature space, which is a limitation you see in SAE-based analysis.

Meng: It gives us joint input–output feature extraction, which means we can investigate the semantic information encoded within varied language models from both the input and the output side.

Lalam: Building on this work, VSA probing enables a unified, concept-oriented analysis of features aligned with both inputs and outputs by integrating supervised probes, SAEs, and logit investigations.

Jane: It’s a method that offers deeper and less variable findings than token-based analysis when you're looking for concept extraction.

Tom: We’ve seen some good results in analogy-style completion tasks where the average probing@one across all models equals eighty-three percent, recovering the target concept together with its corresponding key in most cases.

Lu: And they also found that linguistic analogies yield the lowest rate of missing concept extraction, which they suggest is due to richer LLM representations because they rely on all concepts to capture implicit syntactic patterns.

Meng: What I’m still thinking about is the limitation mentioned: the reliance on a unidirectional transformation in Equation three which prevents causal validation in its current implementation.

Lalam: That's a fair point, and distinguishing probe-induced learning from genuine information decoding is definitely an ongoing challenge in interpretability research.