We develop frameworks that mitigate shortcut learning. Our methods improve robustness under subpopulation shifts by mitigating reliance on confounding signals and non-causal signals. Crucially, we address realistic settings in which spurious correlations are unknown or unlabeled, introducing approaches that achieve strong group robustness without requiring explicit group annotations.

We study how language abstractions and compositional representations can guide reinforcement learning agents toward robust generalization across unseen environments, enabling more transferable policies.

We tackle challenging out-of-distribution settings in which shifts arise from novel compositions of previously seen concepts. Our work introduces principled benchmarks and learning methods that both evaluate and improve model robustness under compositional distribution shifts and complex OoD scenarios.

We propose benchmarks and methods to probe multi-step reasoning and relational understanding. We introduce approaches that strengthen compositional reasoning over complex visual scenes and propose visual reasoning methods for LVLMs.

Representational analysis of VLMs: We introduce datasets and conduct fine-grained analyses of how CLIP represents objects, attributes, and their interactions. By dissecting internal representations, we identify where compositionality emerges—and where it breaks—providing actionable insights for improving model design and training.

Understanding how LLMs and VLMs compute in visual and textual reasoning problems: We develop intervention-based and mechanistic tools to trace how semantic functions are formed and transferred across layers in these models. Our work moves beyond post-hoc explanations toward causal, computation-level understanding of linguistic or multimodal reasoning.

Reliable and faithful attribution methods: We introduce new attribution techniques for Vision Transformers. Our methods address known failure modes of existing attribution approaches and provide more stable, interpretable explanations of model decisions.

Advancing multilingual and low-resource NLP: We introduce high-quality Persian benchmarks covering reasoning, generation, and evaluation, enabling rigorous assessment of LLM capabilities beyond English-centric settings.

Systematic and compositional reasoning in LLMs: We study how LLMs perform multi-step reasoning, identify failure modes, and propose methods to improve logical consistency and generalization across reasoning tasks.

Autonomous decision-making with LLMs: We explore LLM-based agents that interact with environments, tools, and users, studying planning, memory, and coordination in long-horizon tasks.

LLMs as tools for scientific discovery: We investigate how language models can assist in idea generation, pattern discovery, and scientific reasoning, with applications across scientific domains.

Safety and trustworthiness of LLMs: We analyze deceptive behaviors in LLMs, studying when and why models produce misleading outputs, and propose evaluation frameworks to better understand and mitigate such risks.