Akshit Achara

I am a PhD student at King’s College London, supervised by Prof Andrew King, Prof Alexander Hammers, and Dr Esther Puyol Antón. I am part of the DRIVE-Health Centre for Doctoral Training and the Motion Modelling & Analysis Group.

My research is guided by a central question: does the model rely on evidence that is meaningful, robust, and generalisable?

• For image-level and pixel-level classifiers, did the model rely on a spurious or non-generalisable feature? I have studied this through shortcut discovery, spatial localisation, demographic bias, and semantic failures under correlation shift.
• For vision–language models, how do visual and linguistic information interact to shape a prediction? Findings such as mirage reasoning motivate my interest in auditing multimodal reasoning across conventional VLMs and encoder-free multimodal architectures.
• For medical vision systems, can safety-critical or adversarial visual information bypass a model’s safeguards, and can these failures be detected from its internal representations?

My current focus is on the latter two questions, which I approach from an interpretability perspective by studying how visual information is encoded and used within model representations, particularly in relation to model safety.

Before joining King’s, I was a Research Engineer at GE Research and GE HealthCare, working on applied machine learning for medical imaging and clinical text. I hold an integrated MSc in Mathematics from BITS Pilani.