Daria Lioubashevski

darialioub.bsky.social

22 December 2025 17:32

Why is this important? These arbitrary embeddings have no meaningful geometry: related words aren’t closer than unrelated ones. Each vector is just a unique ID for a word, i.e., a symbolic representation.

Daria Lioubashevski

darialioub.bsky.social

22 December 2025 17:32

Taken together, this suggests that the brain—much like an LLM—represents words in two complementary formats at once: 🌀 Continuous, relational 🔢 Discrete, symbolic

Daria Lioubashevski

darialioub.bsky.social

22 December 2025 17:32

Prior work established that LLMs contextual embeddings predict neural activity during conversation. We take it a step further by showing that brain activity is also predicted by *arbitrary* embeddings of each of the model’s top-k ranked words.

Daria Lioubashevski

darialioub.bsky.social

22 December 2025 17:32

LLMs naturally bring these two views together. A word in context is represented simultaneously as: 1️⃣ A high-dimensional contextual embedding (distributed, relational) 2️⃣ A ranking over the vocabulary (discrete, symbolic) Both coexist in the same model 🔀

Daria Lioubashevski

darialioub.bsky.social

22 December 2025 17:32

Traditionally, two seemingly *opposing* views dominate this debate: • Distributed representations: meaning is encoded in the relations between population patterns • Localist/symbolic representations: meaning is encoded by individual neurons, each corresponding to a single concept