Maximizing Rollout Informativeness under a Fixed Budget: A Submodular View of Tree Search for Tool-Use Agentic Reinforcement Learning

Batch of new arXiv (stat.ML) submissions (May 8, 2026) covering diverse machine-learning research: a tree-search training-time framework for tool-using agents (InfoTree) with submodular selector and UUCB; a Bayesian Poisson–Gamma boosting model for forecasting oncology demand; empirical gradient-matching methods to estimate implicit regularization in deep nets; a benign-regularizer theory for nonconvex low-rank matrix estimation; permutation-preserving neural Vecchia kernels for scalable Gaussian processes; a relaxed sparsest-permutation pipeline (SCOPE) for scalable causal-structure recovery;