Armin Toroghi

I am a postdoctoral researcher at the D3M lab, department of Computer Science, University of Toronto, working under the mentorship of Professor Scott Sanner. I am also a Faculty Affiliate researcher at the Vector Institute of AI and the Institute of Information & Communications Technology Planning & Evaluation. My research lies at the intersection of reasoning, LLMs, Knowledge Graphs, and Recommender Systems.

I completed my Ph.D. degree at the University of Toronto in 2025. During my Ph.D. studies, I was a student researcher collaborating with the LG Electronics AI research, and worked as an ML research intern at Borealis AI and Vector Institute of AI. Prior to joining the UofT, I completed my M.Sc. and B.Sc. at the University of Tehran, where my research was focused on developing deep learning-based methods for studying mechanical systems.

Current Research Focus: My research focuses on natural language processing and machine learning, with a particular emphasis on reasoning of large language models (LLMs). I’m interested in enhancing the reasoning capabilities of LLMs, and building agents that are more reliable, efficient, and verifiable, aimed at real-world impact in automated decision support systems and knowledge-intensive tasks.

• Developing Efficient and Verifiable LLM Reasoning Frameworks: We develop efficient and verifiable reasoning frameworks for large language models that go beyond scale-driven performance gains. Our research focuses on improving mathematical and logical reasoning through structured training signals, uncertainty-aware inference, and principled data selection. A central motivation of our work is verifiability: as LLMs become more capable, their errors increasingly stem from confident hallucinations rather than lack of fluency. Verifiable reasoning mechanisms are essential for enabling the safe and effective use of such potent models by exposing uncertainty, enabling checks on intermediate reasoning, and supporting downstream validation.

• Understanding and Mitigating Failure Modes of LLMs: We systematically study where and why large language models fail, with particular focus on unreliable reasoning and confident hallucinations that undermine trust in increasingly capable systems. Our work aims to move beyond surface-level accuracy by analyzing how and why models arrive at their answers, and by developing methods that encourage alignment between correct outputs and correct reasoning processes. Projects such as CoLoTa investigate failure modes in reasoning under distribution shift and limited supervision, while Right for the Right Reasons proposes solutions to the issue of unfaithful answer generation and commonsense reasoning errors.

• Developing Methodologies for Complex Reasoning over Knowledge Graphs: We develop efficient methods for complex reasoning over knowledge graphs, which provide structured, explicit semantics that complement the generative strengths of large language models. In the LLM era, knowledge graphs play a crucial role in enabling grounded, multi-hop, and verifiable reasoning that cannot be reliably achieved through unstructured generation alone. Our projects aim to advance this direction by designing scalable approaches for expressive reasoning over large knowledge graphs.