Paul Missault

Research profile

Research summary

Paul Missault is a researcher at the University of Luxembourg whose visible work sits at the intersection of statistics and machine learning. Evolve's catalog contains two of his arXiv preprints, both dated 2026. The first, "Augmented Hypothesis Testing with Persona-Based LLM Simulations," develops an asymmetric Z-test together with GPPI, a transformed PPI++, to exploit LLM-persona predictions inside A/B testing; Evolve judges the theory solid while noting that gains across four datasets are real but often modest. The second, "Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance," applies shrinkage-corrected covariance estimation to Mahalanobis-distance scoring, largely fixing few-shot OOD intent detection relative to maximum-likelihood estimation, though the edges over Energy and GradNorm baselines fall within reported standard deviations. His bibliometric profile shows a sharply focused record: the field distribution contains a single entry, machine learning & statistics (experimental design), giving a focus score of 100. Collaboration is narrower, scoring 36 from five distinct coauthors, including Jayasimha Talur, Oleg Smirnov, Ziyad Benomar, Aymen Al Marjani and Saab Mansour. Reproducibility readiness averages 55 over the single paper scored so far, with both preprints rated medium. Authorship positions — middle on the first paper, last on the second, with no corresponding-author role — suggest an early-stage record whose output, impact and activity metrics are not yet scored.

Research keywords

Highlights

• Asymmetric Z-test + GPPI bring LLM personas into A/B testing

  On "Augmented Hypothesis Testing with Persona-Based LLM Simulations", Evolve finds that an asymmetric Z-test combined with GPPI (a transformed PPI++) lets persona-based LLM predictions sharpen A/B tests with solid theory; gains across four datasets are real but often modest.

• LLM personas treated as predictions, not just text generators

  The same preprint frames persona-based simulation outputs as predictions inside prediction-powered inference, connecting learning-augmented algorithms to experimental design rather than treating simulation as informal evidence.

• Shrinkage-corrected Mahalanobis distance repairs few-shot OOD intent detection

  On "Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance", Evolve reports that Ledoit-Wolf-style covariance correction largely fixes Mahalanobis scoring versus MLE in few-shot OOD settings, though edges over Energy/GradNorm baselines fall within reported std.

Timeline

1. 2026arXiv preprint "Augmented Hypothesis Testing with Persona-Based LLM Simulations" appears; Evolve reviews the asymmetric Z-test + GPPI approach as theoretically solid, with real but often modest four-dataset gains.
2. 2026arXiv preprint "Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance" appears; Evolve notes shrinkage correction largely fixes few-shot OOD scoring versus MLE.
3. 2026Both 2026 preprints enter the Evolve catalog, establishing an early-stage, statistically oriented record spanning experimental design and NLP evaluation.

Papers in the Evolve catalog

• 2026-09-21

  Augmented Hypothesis Testing with Persona-Based LLM Simulations

• 2026-09-01

  Few-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance

Frequent co-authors