01 / Thesis2026 - present
Stable long-horizon prediction in world models
Investigating manifold-based latent representations for world models, focusing on representation drift and error accumulation during recursive prediction, and how latent geometry and dynamics affect the stability of long-horizon predictions.
World modelsLatent geometryDynamics
02 / Robust learningICLR 2026
Regulating internal alignment flows
Co-developed Alignment-Gated Suppression, a regulariser that uses class-conditional, confidence-weighted internal alignment signals to suppress reliance on spurious features without group labels. Evaluated on Waterbirds, CelebA, BAR and COCO with under 5% training overhead.
Spurious correlationsWorst-group robustness
03 / Fairness2025 - present
Localising and removing bias in vision models
Studying where attribute bias lives across layers and representation subspaces, and building adversarial and representation regularisers that mitigate multiple biases while preserving accuracy - from skin-tone-fair dermatology to multi-attribute debiasing.
DebiasingMedical imagingRepresentation analysis