Research / Thesis
Problem, hypothesis, and open questions.
Transformers dominate general language tasks but struggle on structured reasoning with efficiency guarantees. Supernova explores architectures where reasoning is explicit, verifiable, and deployable on sovereign hardware.
Core hypothesis
Compact hierarchical and latent reasoning models can match or exceed transformer performance on verifiable reasoning benchmarks while using fewer parameters and less energy.
Verifier-based training and neuro-symbolic hybrids may provide the audit trail enterprise deployments require — reasoning steps that can be inspected, not just outputs that look plausible.
Shipped capabilities
- Latent Reasoning Architectures research line
- Benchmark-first publication strategy
- Open collaboration with universities and reasoning labs
- Horizon Europe and EIC Pathfinder alignment