Modular Norm RandOpt: Population-Efficient Ensembling through Architecture-Aware Perturbations
Under review at ICLR 2027
Hiroaki Hamade
Undergraduate student The University of Osaka
Hover over a neuron to reveal its connections. Drag to turn the network. Click, tap, or press Space to send a signal through a new path. Arrow keys rotate the view. Escape resets it. This is an artistic illustration, not a diagram of a specific AI model.
Drag to change the viewing angle. Click or press Space to send a star into orbit. Arrow keys rotate the view. Escape resets it. This is an artistic illustration, not a scientific simulation.
Research interests: AI safety; AI for science.
Research
Research in Progress
When Do Verbalizable Representations Become Useful for Reasoning?
Targeting ICML 2027.
Investigating when intermediate concepts become readable and causally involved in reasoning across OLMo training checkpoints, using J-lens readouts and controlled interventions while accounting for measurement uncertainty.
Notes
Coming soon.