Hen Davidov is a Rhodes Scholar and DPhil student in Statistics at the University of Oxford, where he is advised by Yee Whye Teh, Patrick Rebeschini, and Yarin Gal. He develops efficient methods to intervene in language model generations as they unfold — abstaining, or redirecting to a better continuation the moment a learned value probe signals risk. Hen draws on martingale theory, conformal prediction, and Markov decision processes to design these methods, and uses theoretical analysis to explain empirical method behavior.
This value-based intervention line began with dynamic abstention in “Knowing When to Quit” (ICML 2026), developed during an Applied Scientist internship at Amazon, and continued with “Selective Safety Steering via Value-Filtered Decoding” (ICML 2026 AI4Good workshop; under review at NeurIPS 2026). Both interventions are implemented in value-steer, an open-source vLLM plugin. He is currently working on lower-latency safety steering through Q-function inference. Before Oxford, he completed an MSc at the Technion advised by Yaniv Romano and Ron Kimmel, working on conformal uncertainty quantification — conformalized survival analysis for cancer prognosis (ICLR 2025) and calibrated lower bounds on time-to-unsafe-sampling in LLMs (AISTATS 2026). He was named to Forbes Israel’s 30 under 30 in 2025.