QUANTUMMIND
QuantumMind v1 is our first model built on QART, an architecture combining quantum computing and large language models.
QART is a hybrid architecture combining quantum computing and large language models. The language model produces task-relevant information through understanding, reasoning, and generation. The quantum optimization component processes this information through encoding, optimization, and decoding, then feeds the results back to the model.
Understand the task, reason, and generate content, providing task-relevant information to the optimization component.
Convert task information from model representations or generated content into a representation suitable for optimization.
Formulate the optimization problem in QUBO / Ising terms and use CIM to search for candidate solutions within a given budget.
Convert candidate solutions into auxiliary information the backbone language model can use.
Complex tasks require multiple reasoning steps, and an early mistake can affect later decisions. Studying the conditions for reliable reasoning in QART over longer tasks provides a theoretical basis for continued model development.
Exploring the relationship between optimization capabilities and reasoning task length helps investigate how models can reliably handle longer tasks.
Scroll horizontally to view all columns
| Benchmark | Backbone | QART | Absolute change = QART score − Backbone score. Differences between percentage scores use %; LHTB retains native reward units. | Relative change = (QART score − Backbone score) ÷ Backbone score × 100%. |
|---|---|---|---|---|
| τ²-Bench | 85.52% | 88.90% | +3.38% | +4.0% |
| τ³-Bench | 21.65% | 31.96% | +10.31% | +47.6% |
| SciCode | 17.36% | 31.94% | +14.58% | +84.0% |
| LHTB | 0.2642 | 0.3591 | +0.0949 | +35.9% |
| DeepSWE | 45.13% | 41.59% | −3.54% | −7.8% |
| Terminal-Bench 4.0 | 14.29% | 20.63% | +6.35% | +44.4% |
Scroll horizontally to view all columns
| Benchmark | Backbone | QART | Absolute change = QART score − Backbone score. Differences between percentage scores use %; LHTB retains native reward units. | Relative change = (QART score − Backbone score) ÷ Backbone score × 100%. |
|---|---|---|---|---|
| τ²-Bench | 84.11% | 89.11% | +5.00% | +5.9% |
| τ³-Bench | 43.30% | 45.36% | +2.06% | +4.8% |
| SciCode | 30.90% | 34.38% | +3.48% | +11.3% |
| LHTB | 0.3231 | 0.3419 | +0.0188 | +5.8% |
| DeepSWE | 54.90% | 65.40% | +10.50% | +19.1% |
| Terminal-Bench 4.0 | 14.29% | 20.63% | +6.35% | +44.4% |
QART’s theoretical foundations, architecture, and model evaluations across different backbone language models.