— Research

Research that grounds the product.

Perspix machine-learning research is co-authored with Prof. Frank Hutter's group (University of Freiburg / ELLIS) and published at the field's leading venues. The same techniques — efficient fine-tuning, constrained optimization, reliable transformer training — underpin the reliability profile of the Perspix agent in production.

Peer-reviewed publications

Four papers at NeurIPS, ICLR, and OPT since 2024.

  • 2025
    Learning in Compact Spaces with Approximately Normalized Transformers J. Franke, U. Spiegelhalter, M. Nezhurina, J. Jitsev, F. Hutter, M. Hefenbrock Advances in Neural Information Processing Systems · NeurIPS 2025
  • 2025
    Simultaneous Fine-Tuning and Pruning of LLMs F. Reinicke, J. Franke, F. Hutter, M. Hefenbrock OPT2025 — Annual Workshop on Optimization of Machine Learning
  • 2024
    Improving Deep Learning Optimization through Constrained Parameter Regularization J. Franke, M. Hefenbrock, G. Koehler, F. Hutter Advances in Neural Information Processing Systems · NeurIPS 2024
  • 2024
    Preserving Principal Subspaces to Reduce Catastrophic Forgetting in Fine-Tuning J. Franke, M. Hefenbrock, F. Hutter ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models
Applied research

The Pre-Award Value Gap.

A diagnostic on where value leaks out of complex sourcing before the contract is signed — and how pre-award AI is starting to close the gap. Four leaks per package, scaled to the procurement-lead's portfolio. Workload math behind why proper-depth evaluation is structurally unaffordable today. Reliability profile of the Perspix agent measured against independent human ground truth.

Download the whitepaper →