Controlling Reasoning Effort in LLMs
Ahead of AI · Sebastian Raschka, PhD · 2026-07-18
Sebastian Raschka explains how modern LLMs implement adjustable reasoning effort settings, comparing training approaches—RLVR with token penalties, SFT mode fusion, and continuous effort conditioning—across GPT-5.6, DeepSeek V4, Qwen3, Nemotron 3 Ultra, Kimi K2.5, GLM-5, and Inkling.
Extraction
Topics: reasoning-modelsinference-time-scalingreinforcement-learning-from-verifiable-rewardsllm-post-trainingreasoning-effort-control
Claims
- Reasoning models trained with RLVR learn to generate intermediate reasoning traces purely from output-level reward signals, without any supervision of the trace content itself.
- The <think> and </think> delimiter tokens are cosmetic formatting devices that do not cause reasoning ability; a model trained without them would achieve similar benchmark performance.
- Reasoning effort settings correlate directly with token usage and task performance, but exhibit diminishing returns at high effort levels, making very high settings economically inefficient.
- Six analyzed open-weight flagship models (DeepSeek V4, Nemotron 3 Ultra, Kimi K2.5, GLM-5, Qwen3, Inkling) share a common framework of SFT-based mode introduction followed by mode-conditioned RL with length penalties, despite differing in specifics.
- Automatic reasoning effort selection remains an unsolved problem—GPT-5's Auto mode was removed from the UI—and explicit effort specification via system prompt is expected to remain the dominant paradigm in the near term.
Key quotes
These <think> and </think> tags are cosmetic with respect to reasoning ability. They do not make the model reason, and they are not required to achieve good reasoning performance. One could train the same model without these delimiters and likely reach similar benchmark performance.
During large-scale RL, they did two things for each sample: Specified the desired effort level in the system message. Adjusted the cost assigned to each generated token. Conceptually, the reward likely looked something like this: Low effort uses a larger per-token cost, encouraging shorter reasoning traces. High effort uses a smaller per-token cost, allowing the model to spend more tokens.
In the near future, I think reasoning effort will remain an explicit model input, which will most often be delivered through the system prompt. However agent wrapper/harness around the LLM, or an internal router may increasingly infer the appropriate mode and budget from the task state and available resources automatically (while of course still allowing a user override).