From Conventional LLMs to Reasoning Models to Agents

ACM TechTalks
1

Title: From Conventional LLMs to Reasoning Models to Agents

Date: August 19, 2026

Speaker: Sebastian Raschka

Registration Link

Many of us have been using and experimenting with large language models (LLMs) for several years. Then, in 2025, new techniques for training reasoning models gave LLMs a noticeable boost in problem-solving capabilities. Since then, newer generations of LLMs have been substantially more useful for complex tasks such as math and coding.

For example, in software development, the progression over the years has been from simple code completion to planning, implementing, testing, and debugging entire features and code bases.

Today, reasoning models, which can be viewed as an evolution of conventional LLMs, are also at the center of many agent frameworks.

This talk will explain how reasoning models differ from conventional LLMs, with a particular focus on inference scaling and training techniques. It will also introduce the main components of AI agents and show how reasoning models serve as the engine that enables agents to plan, use tools, evaluate intermediate results, and recover from mistakes.

Additional Resources:
Build a Reasoning Model (From Scratch) (free O’Reilly book for ACM Members with Skills Bundle)
Build a Large Language Model (From Scratch) (free O’Reilly book for ACM Members with Skills Bundle)
GenAI & Azure OpenAI Service: Chat and Reasoning Models in Foundry Playgrounds (free Percipio course for ACM Members with Skills Bundle)
Introduction to DeepSeek-R1 and Reasoning Models (free Percipio course for ACM Members with Skills Bundle)
Agentic AI Mastery (free Percipio course for ACM Members with Skills Bundle)

2

Hi Sebastian,

Can you, in a nutshell, explain how any version of the group-level statistics that are the bread and butter of LLM models do anything other than damage context-based reasoning?

LLMs are, of course, excellent at uncovering subtle patterns, if you don’t mind huge false-positive rates (hallucinations). But I don’t understand how that kind of discovery even relates to context-aware reasoning. After all, at least initially, an LLM treats the fact that “Einstein” and “Schweitzer” often follow “Albert” in group-level training statistics as a valid pattern discovery. But in all of the individual, context-rich conversations from which LLM training gleaned such group-level statistics, any such connection between these two very different people is pure noise of the type that then leads directly to hallucinations.

In other words, I do not understand how LLM technology, which always starts with superficial group-level statistics gleaned from raw language data, can ever directly support reasoning. If anything, it would seem to do the opposite, since that group-level data can then be used to generate sentences and images that will sound and look plausible to anyone not familiar with the full context of the original conversations, but are actually nothing more than deceptive mimicries of actual conversations. It is like asserting that people with the unfortunate condition of Wernicke’s aphasia will somehow help you think better if you accept their syntactically correct but semantically empty statements as meaningful.

Sincerely, Terry Bollinger