Events

Talk
Modular and Selective Reasoning: Inference-Time Optimization for Large Language Models

Modular and Selective Reasoning: Inference-Time Optimization for Large Language Models

  ↵  ↵Abstract :-
Large language models (LLMs) demonstrate remarkable reasoning capabilities but remain prohibitively expensive to deploy at scale. Most of the computational cost arises at inference time, where models redundantly explore multiple reasoning paths or overuse large model capacity for tasks that smaller models could handle. In this talk, I will present two complementary approaches for inference-time optimization of reasoning in LLMs.
First, Offloaded Reasoning (OR) introduces a modular paradigm in which a smaller model performs the reasoning while a larger model focuses on synthesis and refinement—achieving up to 8× faster inference without retraining. Second, Path of Least Resistance (PoLR) exploits prefix consensus across reasoning traces. PoLR preserves accuracy while cutting token usage and latency by 50–60%. Together, these methods redefine how we think about efficient reasoning—shifting the optimization focus from model training to inference-time computation. The talk will discuss empirical results, theoretical underpinnings, and emerging directions toward scalable, sustainable reasoning systems.
  ↵
Speaker Bio :-
Ishan Jindal, Ph.D., is a Staff Engineer at Samsung R&D India–Delhi, where he leads research on efficient and adaptive large language models for next-generation agentic AI systems. His work focuses on LLM inference-time optimization, prompting and fine-tuning, and human-in-the-loop reasoning systems. Before joining Samsung, he was a Staff Research Scientist at IBM Research Almaden (USA), where he contributed to the development of multilingual NLP systems for IBM Watson and co-developed the Universal PropBank 2.0 dataset. His research has appeared in premier venues including ACL, EMNLP, NAACL, EACL, ICASSP, and he holds multiple U.S. patents in deep learning and semantic understanding.
  • 14 - 14 November,2025
  • 12:00 pm - 01:00 pm
  • LH534