Mila has announced that its researchers will present 70 papers at the International Conference on Learning Representations (ICLR) 2026 in Brazil. Key presentations on the first day will cover advancements in model merging, graph learning, and reinforcement learning. Notable works include DisTaC's robust model merging via distillation of conditional task vectors and GraphOmni's benchmark framework for evaluating large language models on graph theory tasks. In reinforcement learning, SHAPO introduces sharpness-aware optimization for safe exploration, while ARM-FM utilizes foundation models for automatic reward machine generation. Generative models will see innovations like Efficient Regression-based Training of Normalizing Flows for Boltzmann Generators and FALCON's precise likelihood computation in continuous flows. Additionally, Mila will present advancements in large language models, including the visualization of reasoning processes and redefining model collapse as a feature of machine forgetting.