Sebastian Raschka • 7/18/2026

Controlling Reasoning Effort in LLMs

This article by Sebastian Raschka explores the concept of controlling reasoning effort in large language models (LLMs), focusing on how models like OpenAI's GPT-5.6 and DeepSeek-R1 implement multiple reasoning-effort settings (low, medium, high). It defines reasoning models, explains their training via reinforcement learning with verifiable rewards (RLVR), and discusses the methodology behind developing models with adjustable reasoning modes. The article is aimed at AI researchers and practitioners, providing insights into modern LLM capabilities and inference strategies.

0 comments

#llm #Reinforcement Learning #Reasoning Models