Large language models (LLMs) have advanced far beyond text prediction—they now exhibit surprising abilities to reason through problems, follow multi-step processes, and produce logically consistent answers. Yet “reasoning” in AI often means something different from human reasoning, and understanding that distinction is key to using it effectively.

This talk clarifies what reasoning means for LLMs and why intermediate steps matter for complex problem-solving. We’ll explore Chain-of-Thought prompting, Self-Consistency decoding, RL-based reasoning, and distilled reasoning models—highlighting their strengths, trade-offs, and performance costs.

A particular focus will be on adaptive reasoning models, which dynamically adjust their computational depth based on problem complexity. Unlike traditional fixed-computation approaches, these models can allocate more "thinking time" to harder problems while remaining efficient on simpler tasks.

We'll then present a practical decision framework for when to apply reasoning—balancing value, complexity, and latency. Not all tasks benefit from multi-step inference; some work best with direct outputs, while others gain from deeper reasoning.

We’ll also address challenges: compound errors cascading through steps, trade-offs between model size and reasoning quality, and scaling issues with large context windows.

Finally, we’ll cover context engineering—structuring prompts, retrieval pipelines, and context windows to improve reasoning reliability and efficiency.

Attendees will leave with:
- A clear understanding of reasoning in LLMs.
- Insight into reasoning emerging from RL with verifiable output.
- Knowledge of key techniques and emerging architectures.
- Practical tips for optimizing reasoning in AI projects.