Publications

Publications

Research papers in AI evaluation, learning theory and quantum optics.

2026
Preprint · arXiv:2609.00088

Commit-first LLM judging inherits the judge’s own errors

Idil Gozel

A defence against evaluator gaming can move the source of error to the judge’s own committed answer. Includes an audit of default configurations across eight evaluation frameworks.

View citation
@misc{gozel2026commitfirst,
  title = {Commit-first LLM judging inherits the judge's own errors},
  author = {Gozel, Idil},
  year = {2026},
  eprint = {2609.00088},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.00088}
}
2026
Preprint · arXiv:2608.07228

Learning Suffers More Than the Policy Class Under Partial Observability: A Closed-Form Analysis

Idil Gozel

Exact theory separates a learning rule’s failure from the limits of the policies it can represent, and identifies how the learning horizon changes the outcome.

View citation
@misc{gozel2026partialobservability,
  title = {Learning Suffers More Than the Policy Class Under Partial Observability: A Closed-Form Analysis},
  author = {Gozel, Idil},
  year = {2026},
  eprint = {2608.07228},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.07228}
}
2026
Journal article · Nano Letters 26(3), 990–996

Harnessing the Diamond-Air Interface as an Efficient Photon Antenna for Solid-State Emitters

Paul Weinbrenner, Aina Lopez Benet, İdil Gözel and Friedemann Reinhard

Research on the diamond–air interface as a photon antenna for solid-state emitters, arising from quantum-optics work at the University of Rostock.

View citation
@article{weinbrenner2026diamond,
  title = {Harnessing the Diamond-Air Interface as an Efficient Photon Antenna for Solid-State Emitters},
  author = {Weinbrenner, Paul and Lopez Benet, Aina and Gözel, İdil and Reinhard, Friedemann},
  journal = {Nano Letters},
  year = {2026},
  volume = {26},
  number = {3},
  pages = {990--996},
  doi = {10.1021/acs.nanolett.5c04935}
}

Research projects, manuscripts & policy

Entanglement Distribution in Quantum Networks: A Reinforcement Learning Approach

Research project · UCL / BT · 2025–2026

Public link forthcoming

Training-parameter bias in evaluations of agent memory

University College London · 2026

Manuscript in preparation

Response to Ofgem call for input: AI assurance

Consultation response · 12 August 2026

Request a copy