Controlling Reasoning Effort in LLMs — How LLMs Learn Low-, Medium-, and High-Effort Reasoning Modes

Main argument

Modern reasoning LLMs are trained to support multiple levels of computational effort — low (quick answer), medium, and high (long reasoning chain). Key insight: a smaller model at high effort can match a larger model at low effort.

Context

The post responds to the explosion of reasoning models (DeepSeek V4, Qwen3, Inkling from Thinking Machines) and developer demand for understanding when thinking mode is worth the extra tokens.

Why it matters

Understanding reasoning effort control is critical for ML engineers optimizing costs — switching between modes can reduce token costs 10-50x while maintaining quality on routine tasks.

Details / arguments

• RLVR (Reinforcement Learning with Verifiable Rewards) is the primary training technique for reasoning levels
• Token-length-penalizing reward functions force the model to be concise on simple questions
• DeepSeek V4, Qwen3, and Inkling use different strategies: separate specialists vs. mixed training vs. continuous conditioning
• Inference-time scaling and training-time scaling interact — they are not orthogonal
• System prompt is the simplest way to switch mode, but the model must be trained to listen to it