Controlling Reasoning Effort in LLMs

This article explores how large language models (LLMs) are trained to have variable reasoning effort modes, ranging from low to high. It explains that "reasoning models" in the context of AI generate intermediate reasoning traces, allowing for step-by-step problem-solving, unlike conventional LLMs that provide a direct answer.

The author details the two main approaches to improving reasoning performance: scaling training (via techniques like RLVR – Reinforcement Learning with Verifiable Rewards) and scaling inference (by allocating more computation at the time of use). It highlights how models like GPT-5.6 and Inkling integrate reasoning effort parameters, allowing users to choose between different response lengths and levels of precision, often correlated with cost and latency. These mechanisms are implemented via adjustments in training (SFT, RL) and system prompts.

Finally, the article examines various implementations of these effort controls in cutting-edge open-source LLMs (DeepSeek V4, Nemotron 3 Ultra, Kimi K2.5, GLM-5, Qwen3, Inkling), showing a diversity of approaches ranging from dedicated specialists to continuous modes, and dynamic token budgets. The conclusion emphasizes that reasoning effort will likely remain an explicit input for the model in the short term, although automatic selection is the long-term goal.