arXiv 2026
T. Kempton, Viktor Drobnyi, Maeve Madigan, and Stuart Burrell. arXiv preprint, 2026.
We collect evidence from other papers showing that likelihood-based detectors of
machine-generated text really target the overconfidence of instruction-tuned models. We show that this
overconfidence does not present itself uniformly across hidden space, and that the aggregation of token-level scores in most
detectors significantly weakens the signal. We introduce a local calibration step before averaging which dramatically improves
the performance of state-of-the-art likelihood-based detectors.
ICLR 2026
T. Kempton, Julia Rozanova, Parameswaran Kamalaruban, Maeve Madigan, Karolina Wresilo, Yoann Launay, David Sutton, and Stuart Burrell.
Introduces DMAP, a statistically rigorous way to visualise where a text sits in the next-token probability distribution of a language model.
AISTATS 2025
T. Kempton, S. Burrell, and C. Cheverall. Proceedings of the 28th International Conference on Artificial Intelligence and Statistics, 2025.
Introduces a detector for machine-generated text based on deficiencies in the way language models perform top-k or temperature sampling.
Findings of ACL: EMNLP 2025
T. Kempton and S. Burrell. Findings of the Association for Computational Linguistics: EMNLP 2025.
Develops a theoretical description of top-k, nucleus, and temperature-based decoding in the language of
equilibrium states, and analyzes how local normalization affects quality and diversity of generated text.
AAAI 2026 Workshop
Maeve Madigan, Parameswaran Kamalaruban, Glenn Moynihan, T. Kempton, David Sutton, and Stuart Burrell. Accepted at the Workshop on Agentic AI in Financial Services at the AAAI 2026 conference in Singapore.
Studies fairness evaluation in multi-agent predictive systems and shows how emergent bias can arise from
collective system behavior rather than any single component.
Submitted
Yoann Launay, Parameswaran Kamalaruban, T. Kempton, Stuart Burrell, and David Sutton. Submitted.
Introduces a fairness-aware test-time adaptation method for vision-language models that reduces bias under
distribution shift by tuning prompts to improve target performance while suppressing spurious attribute signals.