Quantum Language Model

How It Works

A tiny MLP learns language patterns, gets converted to a quantum circuit, and runs on real hardware.

1

A 2-layer MLP (61 → 32 → 16 → 8) learns masked token prediction on a tiny 8-word vocabulary. Trained with Adam optimizer for 5000 steps.

Input: "the ___ sat on the mat" ↓ [One-hot encode all tokens] [Mark masked position] ↓ MLP: 61 → 32 → 16 → 8 ↓ Output: P(cat)=0.72, P(dog)=0.18, ...

2

The MLP's output probability distribution is encoded into rotation angles on 3 qubits. Each prediction becomes a quantum circuit with ~4 CNOTs.

q0: |0⟩──[Ry]──●────────────────╮── M │ │ q1: |0⟩───────┴──[Ry]──●────┴── M │ q2: |0⟩──────────────┴──[Ry]── M

3

Circuits run on the QuTech Tuna-9 superconducting processor. 4096 shots per circuit. Measurement outcomes decoded back to token predictions.

QuTech Tuna-9 (T-9) ────────────────────── 9 transmon qubits e_1q ~ 0.1%, e_2q ~ 1% 4096 shots per circuit Measured: |101⟩ = token 5 = "mat"

Build a sentence, mask one word, and see the model's prediction. This runs the actual trained MLP weights in your browser.

VOCABULARY (tap to add to sentence)

YOUR SENTENCE

Tap tokens above to build a sentence...

Runs the classical MLP (2648 params) entirely in your browser via JavaScript. No server needed. Real T-9 hardware results shown below.

8 test cases run on QuTech Tuna-9 superconducting quantum computer. 4096 shots each.

0%

Random baseline
(1/8 chance)

50%

Quantum (T-9)
4/8 correct

62.5%

Classical MLP
5/8 correct

p = 0.011

Statistical significance
quantum ≠ random

Test Case	Expected	Classical	Quantum (T-9)
the cat ___	sat	ran	ran
the dog ___	ran	ran	cat
___ cat sat	the	the	the
the ___ sat on a mat	cat	cat	cat
cat ___	ran	ran	sat
dog ___	sat	ran	sat
___ mat	the	cat	cat
the cat sat on ___	mat	mat	mat

This is a proof-of-concept. Here is what it cannot do.

Toy vocabulary of only 8 tokens. Real language models have 50,000+ tokens.
The quantum circuit encodes the classical model's output distribution, not the weights themselves. It is a quantum sampler, not a quantum learner.
3 qubits cannot represent a vocabulary larger than 8 tokens. Scaling to useful sizes requires fault-tolerant hardware.
T-9 accuracy (50%) is below the classical model (62.5%) due to hardware noise. The gap will close as hardware improves.
No quantum advantage is claimed. This demonstrates feasibility, not superiority.
The training is entirely classical. Only inference runs on quantum hardware.
Circuit optimization is heuristic (multi-restart numerical optimization). Better decompositions likely exist.