Watch the recording here.

Teams link.

Speaker: Niket Tandon, Microsoft Research, India

Time: June 22, 1:00-2:30

Location: KEC 1007

Abstract: Reasoning has been a central theme in AI long before LLMs, traditionally addressed through symbolic inference, search, and structured representations. This talk traces how the notion of reasoning has evolved in the era of LLMs. We begin with classical approaches and switch gears to chain‑of‑thought prompting, analyzing why explicit reasoning traces work and how their effectiveness is tied to decoding dynamics rather than guaranteed logical correctness. We discuss known limitations of chain‑of‑thought, including issues of faithfulness. The talk then turns to test‑time scaling methods such as self‑refinement and outcome and process‑reward models that trade additional inference‑time compute for improved reasoning. We further examine tool use and agentic reasoning, highlighting how external tools and interaction enable capabilities beyond standalone models. Finally, we discuss benchmarks and metrics for evaluating reasoning, and the gaps they leave in assessing true reasoning ability.

Bio: Niket Tandon is a Principal Research Scientist at Microsoft Research India, where he focuses on customizing AI copilots with private data. Previously, he was a Lead Research Scientist at the Allen Institute for AI in Seattle, working on feedback-guided reasoning in LLMs as part of the Aristo team that is known for building AI that aced science exams. He earned his Ph.D. from the Max Planck Institute for Informatics under Prof. Gerhard Weikum, where he created WebChild, then the largest automatically extracted commonsense knowledge graph. His work has been recognized at top venues. Niket also founded PQRS Research to support undergraduates from underrepresented institutes and actively contributes to the NLP community through workshops, tutorials, and conference service.