Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures

Name: Hyperdimensional Probe: Decoding LLM Representations via Vector Symbolic Architectures
Uploaded: 2026-10-23T12:49:19.978847+00:00
Duration: 436 s
Description: The Hyperdimensional Probe uses a hybrid supervised method combining Vector Symbolic Architectures (VSAs) and neural probing to understand Large Language Model vector spaces. It maps LLM embeddings into a bounded proxy space, allowing for concept-oriented insights by extracting meaningful semantic concepts aligned with both input and output features.

The Hyperdimensional Probe uses a hybrid supervised method combining Vector Symbolic Architectures (VSAs) and neural probing to understand Large Language Model vector spaces. It maps LLM embeddings into a bounded proxy space, allowing for concept-oriented insights by extracting meaningful semantic concepts aligned with both input and output features.

Open the full episode