Meridian AI faculty and students will present five papers at the 2026 International Conference on Machine Learning (ICML), to be held in Vienna, Austria. The accepted papers span the institute's core research areas and represent some of the strongest ICML acceptance rates of Meridian's short history.
The accepted papers are:
"Spectral Normalization is Insufficient: New Stability Criteria for Transformer Attention" — Dr. Ingrid Holmberg and PhD students Lucas Bauer and Preet Ahluwalia. This paper provides new theoretical results on when transformer attention becomes numerically unstable during training and derives practical initialization criteria that reduce training loss spikes.
"AlignBench: A Reproducible Suite for LLM Value Alignment Evaluation" — Dr. Kwame Asante, Dr. James Okafor, and the AlignBench team. The paper introduces the AlignBench evaluation framework and presents results across 14 open-source and proprietary models. Full dataset and evaluation code are released.
"DreamerV4: World Model Scaling Laws and the Sample Efficiency Frontier" — Dr. Elias Müller, Dr. Abraham Kiptoo, and PhD candidate Sakura Yamamoto. An empirical investigation of how world model quality scales with model size and environment diversity, with implications for safe offline RL.
"Curvature Concentration in the Adam Loss Landscape: A Random Matrix Analysis" — Dr. Ravi Shankar (solo paper). The first paper from the NSF CAREER-funded research program.
"Canary: Lightweight Behavioral Drift Detection for Production LLMs" — Preet Ahluwalia, Mei-Ling Zhu, et al. The hackathon-winning safety monitoring system, expanded into a full research contribution.
"Five ICML papers in a single year, for an institute that graduated its first PhD students just two years ago, is a remarkable outcome," said Provost Dr. Helena Voss. "It reflects both the quality of our faculty and the caliber of students they are training."
The ICML conference runs July 18-26 in Vienna.