CBIRD: Complex Bidirectional Inducer for Representation Dynamics

Created: Damir Cavar, 2026-02-17

Last change: Damir Cavar, 2026-03-31

Overview

CBIRD is a hybrid complex-valued language model based on the core BERT architecture. It implements complex-valued transformers, tensors, and attention heads. The trained embeddings and models are compatible with classical GPU-based computing environments, as well as with quantum computers.

Objective

• Create complex-valued neural architectures (and advanced libraries that overcome the limitations of PyTorch and similar libraries)
• Reduce model size, required data, and training effort while improving the expressivity of AI models
  • a 1-trillion-parameter model represented as a quantum state requires fewer than 40 qubits in our architecture
• Lower inference costs using quantum computing
• Benchmark quantum computers in general, and complex-valued hybrid models for downstream applications in particular
• Evaluation of hybrid classical-quantum AI models for:
  • Medical applications
  • Robotics and control

Presentations

• Damir Cavar, Shane Sparks, Soren DeHaan, Ronit Jha, Ayomide Jeje, Yanin Charoenpornsawat, Sam Gray Miller (2026) Hybrid Classical-Quantum Language and World Models for AI - CBIRD. Poster presentation at the Center for Quantum Technologies (CQT) Meeting 2026 at Notre Dame, 1st of April, 2026.