LIG Seminar - Behrooz Omidvar-Tehrani

Abstract
LLMs have shown remarkable capabilities in reasoning and decision-making, but their effectiveness hinges on how we structure their thinking. This talk explores the cognitive analogy of System I and System II thinking in LLM agents, and how techniques like Chain-of-Thought (CoT), Retrieval-Augmented Generation (RAG), and iterative prompting enable more reflective, System II-style reasoning. We introduce DoT (Diversity of Thoughts), a novel framework designed to overcome two key limitations in current agent architectures: (1) the tendency for repetitive, redundant reflections that limit exploration of the decision space, and (2) the lack of memory across tasks, which prevents agents from learning from past successes. DoT addresses these by promoting diverse reasoning paths and incorporating a task-agnostic memory module that enables knowledge reuse across tasks. Through extensive experiments on programming benchmarks such as HumanEval, MBPP, and LeetCodeHardGym, DoT  achieves up to a 10% improvement in Pass@1 while maintaining cost-efficiency. Moreover, its modular design allows seamless integration with existing frameworks like Tree of Thoughts (ToT), yielding a 13% performance boost on the Game of 24 benchmark. This talk will highlight how DoT advances the frontier of LLM-based agents by enabling deeper reasoning, broader exploration, and cumulative learning—paving the way for more capable and generalizable AI systems.
 

Biography

Behrooz Omidvar-Tehrani is a Principal Applied Scientist at Microsoft, where he advances agentic AI systems for M365 Copilot to improve reasoning and engagement across communication-centric workflows. Previously, he was a Senior Applied Scientist and Science Manager at Amazon contributing to Amazon Q Developer and AWS Transform. He earned his PhD in 2015 from the University of Grenoble Alps in France. He has published more than 40 papers in top-tier conferences including ICML, ICLR, SIGMOD, CHI, and CIKM.