Publications

Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

Published in arXiv, 2026

SR²AM equips a single LLM with internal “System I” (reactive), “System II” (simulative planning via a learned world model), and “System III” (self-regulation over planning depth and action vs. simulation). A configurator determines when and how far to simulate, enabling 30B models to rival 685B, 1T models at a fraction of the token cost. Thinking longer doesn’t always mean thinking smarter, SR²AM lets the LLM know when to simulate ahead, when to act directly, and when to balance both for optimal performance.

Project website: https://sailing-lab.github.io/sr2am-self-regulated-planning/

Pitfalls of Unlabeled Disagreement-Based Drift Detection in Streaming Tree Ensembles

Published in Catch, Adapt, and Operate: Monitoring ML Models Under Drift Workshop (ICLR 2026), 2026

Accepted in ICLR 2026 — Best Poster Award CAO Workshop

Online Hierarchical Partitioning of the Output Space in Extreme Multi-Label Data Streams

Published in European Conference on Artificial Intelligence (ECAI), 2025

This paper presents iHOMER, a novel framework for online multi-label learning in dynamic environments. iHOMER is the first incremental, drift-aware algorithm that partitions the label space into correlated clusters without requiring predefined hierarchies. It adapts to concept drift over time, balances scalability and predictive performance, and employs statistical tests to guide both splitting and merging of label clusters.