2026 Theses Doctoral

Trustworthy Multimodal Intelligence

Menon, Sachit

While AI has achieved stunning capabilities in recent years, today's systems fail unpredictably and in unsafe ways. This dissertation develops machine learning systems we can trust in complex, multimodal settings.

For a system to be trustworthy, it should be interpretable, so we can audit its decisions; reprogrammable, so we can change how it makes decisions in the future; and fail safely, so damage is minimized when failure does occur. A prevailing assumption in deep learning is that these factors can only come at the cost of raw performance. This dissertation challenges this dichotomy, developing methods that demonstrate inherently explainable approaches — leveraging the capabilities of foundation models through visual rationales or executable code — can actually enhance capability while providing the necessary hooks for reliability.

I advance this vision through three synergistic thrusts: exposing and shaping model rationales and representations; multimodal agents, code reasoning, and tool calling; and evaluation and alignment of decision processes over final answers alone. I explore these thrusts over a diverse range of tasks, developing new methods across image classification, long-form multimodal reasoning, spatial navigation, and more. These methods go beyond achieving state-of-the-art results, also exposing new avenues for reliability and safety. I also build new datasets and benchmarks to learn and evaluate capabilities that existing datasets did not capture, such as surfacing issues in contrastive representation learning and auditing multimodal reasoning processes. All of this work goes towards building trustworthy multimodal intelligence, in the hopes of a world where we can trust our machines to make decisions that are both correct and safe.