Adaptive control of reasoning effort in language models

Reasoning-capable language models have become a standard component in modern AI releases. Since OpenAI popularized the o1 model, followed by DeepSeek-R1 and reinforcement learning techniques with verifiable rewards (RLVR), step-by-step reasoning capabilities have become central to LLM development. OpenAI's latest GPT-5.6 family integrates adjustable reasoning effort configurations across multiple model sizes. Two primary strategies improve performance on reasoning tasks: training scaling and inference scaling. The RLVR method trains models through reinforcement signals in verifiable domains such as mathematics and code, enabling models to learn to generate intermediate explanations, self-correct, and solve complex problems systematically. A significant evolution has been the introduction of hybrid models like Qwen3 that allow toggling between active and inactive reasoning modes via configuration. During training, these models encounter both reasoning and non-reasoning examples, with mode fusion implemented through supervised fine-tuning followed by additional reinforcement learning. Reasoning effort levels control the scope of reasoning and response accuracy, directly scaling computational token usage. Reasoning delimiters function as formatting markers without inherent reasoning capability, allowing separation of reasoning traces from final answers for user presentation. Current model performance suggests that reasoning effort control is fundamental to modern LLM architecture.