Uploaded October 2025 | Updated September 2026, 2 days ago
The development of architectures that excel in reasoning and planning is pivotal for advancing Artificial Intelligence (AI). A significant aspect of this research involves sequential decision-making in problems, particularly where AI agents interact with humans. Ideally, we would desire AI agents to be (1) advisable, and (2) explainable during these interactions where they can both take advice from humans for solving relevant problems, and also provide explanations for their decisions. A straightforward approach to aligning AI agents with humans for a seamless interaction would be to directly query the human-in-the-loop for problem-specific advice or their feedback on how explainable the agent’s decisions are. However, this approach can be extremely resource-intensive in terms of the time and effort required by the human. Recently, generative models, such as Large Language Models (LLMs), have demonstrated remarkable performance across a diverse range of natural language-based tasks. Critically viewing them as broad and shallow models trained on a large corpus of human and synthetically generated data, this talk aims to systematically explore the roles they can play in the Human-AI interaction scenarios to make the AI agents advisable and explainable. (1) Specifically, we look at the case where a LLM can be the underlying planner instead and investigate the extent to which they can be advised by human-given plan guidance. (2) From the perspective of explainability, we shift our focus to understand the semantics and interpretability of LLMs, especially their reasoning traces with respect to end users, and how they correlate (or lack thereof) with final performance. In summary, this talk aims to highlight critical pathways for designing robust and reliable end user-facing systems.
Siddhant Bhambri is a fifth year Ph.D. student under the supervision of Dr. Subbarao Kambhampati in Yochan Lab at the School of Computing & AI, Arizona State University.
His research centers on the design and implementation of intelligent AI systems that empower human decision-making through the synergistic application of Large Language Models (LLMs), Large Reasoning Models (LRMs), and Reinforcement Learning (RL). His work critically examines the strengths and weaknesses of Foundational Models from a Human-AI Interaction standpoint, specifically investigating how these models can be tailored to meet the needs of real-world users and facilitate seamless human-AI collaboration for improved and more reliable decision outcomes across various domains.
The development of architectures that excel in reasoning and planning is pivotal for advancing Artificial Intelligence (AI). A significant aspect of this research involves sequential decision-making in problems, particularly where AI agents interact with humans. Ideally, we would desire AI agents to be (1) advisable, and (2) explainable during these interactions where they can both take advice from humans for solving relevant problems, and also provide explanations for their decisions. A straightforward approach to aligning AI agents with humans for a seamless interaction would be to directly query the human-in-the-loop for problem-specific advice or their feedback on how explainable the agent’s decisions are. However, this approach can be extremely resource-intensive in terms of the time and effort required by the human. Recently, generative models, such as Large Language Models (LLMs), have demonstrated remarkable performance across a diverse range of natural language-based tasks. Critically viewing them as broad and shallow models trained on a large corpus of human and synthetically generated data, this talk aims to systematically explore the roles they can play in the Human-AI interaction scenarios to make the AI agents advisable and explainable. (1) Specifically, we look at the case where a LLM can be the underlying planner instead and investigate the extent to which they can be advised by human-given plan guidance. (2) From the perspective of explainability, we shift our focus to understand the semantics and interpretability of LLMs, especially their reasoning traces with respect to end users, and how they correlate (or lack thereof) with final performance. In summary, this talk aims to highlight critical pathways for designing robust and reliable end user-facing systems.
Siddhant Bhambri is a fifth year Ph.D. student under the supervision of Dr. Subbarao Kambhampati in Yochan Lab at the School of Computing & AI, Arizona State University.
His research centers on the design and implementation of intelligent AI systems that empower human decision-making through the synergistic application of Large Language Models (LLMs), Large Reasoning Models (LRMs), and Reinforcement Learning (RL). His work critically examines the strengths and weaknesses of Foundational Models from a Human-AI Interaction standpoint, specifically investigating how these models can be tailored to meet the needs of real-world users and facilitate seamless human-AI collaboration for improved and more reliable decision outcomes across various domains.
