Quantum word embeddings

Can today's word embeddings survive the trip into a quantum computer? In Old Wine in New Bottles: Using Classical Word Embeddings in Gate-Based Quantum NLP Systems (Cavar, Parukola, Graves & Sparks, Indiana University, 2025), we showed that n-dimensional embeddings can be packed into just log₂(n) qubits — or log₂(n/2) by folding pairs of dimensions into complex amplitudes — and that word similarity survives the encoding nearly intact. This page runs that pipeline live in your browser, then lets you explore the actual experimental data.

The pipeline, live

Runs entirely in your browser — nothing is installed, and nothing you type leaves this page. The SWAP test is simulated exactly: the ancilla measures 1 with probability (1 − |⟨a|b⟩|²)/2, and real Bernoulli shots are drawn from it — the same statistics an ideal statevector simulator produces.

The research data

The paper's experiments: 666 noun pairs from SimLex-999, embedded with GPT and VoyageAI models, amplitude-encoded, and measured with SWAP tests on a quantum simulator. Every point below is one word pair from those runs.

Notes

• Live embeddings use small in-browser models; the paper's experiments used GPT and VoyageAI embeddings via API. The 2026 state-of-the-art on-device embedding models (EmbeddingGemma-300m, Qwen3-Embedding) ship weights as single files far above the 25 MiB per-file limit of this site's host — the two models offered here are the strongest that physics and static hosting allow.
• Complex encodings halve the qubit count and state-preparation gates — the paper's central practical argument. Toggle the encoding above and watch the amplitude count and qubit readouts change.
• Based on joint work with Damir Cavar, Koushik Reddy Parukola, and James Bryan Graves (Indiana University); supported by the Center for Quantum Technologies (NSF #2224928). Code and data: NLQK.