Controlling Reasoning Effort in LLMs

Modern large language models like OpenAI's GPT-5.6 family are increasingly shipping with adjustable "reasoning effort" settings, allowing users to trade performance for computational cost. These modes are implemented through specialized post-training techniques, primarily effort-conditioned supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) where token penalties are adjusted based on the desired effort level.

The concept of "reasoning models" that output an intermediate chain of thought was popularized by models like OpenAI's o1 and DeepSeek-R1, which was trained using reinforcement learning with verifiable rewards (RLVR). While effective, these models are often verbose and computationally expensive by default. Early attempts to control this involved simple on/off switches, as seen in models like Qwen3. The latest generation of models, including GPT-5.6 and several open-weight alternatives, has evolved this into granular effort levels (e.g., 'Light', 'Medium', 'Max'). This shift addresses the practical need for users to balance model accuracy against API cost and latency for different tasks, moving beyond a one-size-fits-all approach to model inference.

Defining Reasoning Models:

A reasoning model is an LLM that generates an intermediate reasoning trace before providing a final answer. This trace, often enclosed in special tags like <think>...</think>, shows the model's step-by-step process, including self-corrections known as "Aha" moments. These tags are primarily for formatting and do not intrinsically enable reasoning; they help user interfaces and training pipelines distinguish the reasoning from the final output.

Core Training Technique: RLVR:

The foundational technique for training reasoning models is Reinforcement Learning with Verifiable Rewards (RLVR), popularized by DeepSeek-R1.

The model is trained using a reward signal of 1 for a correct answer and 0 for an incorrect one.
Correctness is determined by external verifiers in domains like math (using SymPy or WolframAlpha) or code (using compilers or unit tests).
The DeepSeek-R1 paper reported that providing rewards based on the intermediate reasoning trace was not helpful, so only the final answer's correctness and format were used to calculate the reward: R_total = R_accuracy + R_format.
Models like DeepSeek-R1-Zero demonstrated that RLVR alone is sufficient to teach a base model to generate and use reasoning traces, without prior supervised fine-tuning (SFT).

Implementing Effort Controls:

Models with adjustable effort levels are trained to associate an effort label (e.g., "low", "high", or a continuous value) with a specific response length and reasoning depth. This is typically achieved in two ways during post-training:

Effort-Conditioned SFT: The model is fine-tuned on a dataset where prompts include effort labels, and the target responses are curated to match the desired reasoning length for that label.
Effort-Conditioned RLVR: During reinforcement learning, the reward function is modified based on the effort label. For example, the reward for a sample with effort level e might be R(e) = R_task - λ(e) * N_tokens, where a low effort level corresponds to a high token penalty `λ(e)`, encouraging shorter responses.

Case Studies in Open-Weight Models:

Several recent open-weight models have disclosed their methods for implementing effort controls, revealing a common framework with different specific techniques.

Model	Training Mechanism	Inference Control
DeepSeek V4	Trains separate specialist models for 'Non-think', 'Think High', and 'Think Max' modes using different RLVR configurations (context window, length penalty). These specialists are then distilled into a single model.	A system prompt (e.g., "Reasoning Effort: Absolute maximum...") selects the desired mode.
Nemotron 3 Ultra	Introduces a 'medium-effort' mode via SFT on data from GPT-OSS-120B. Also trains on randomly truncated reasoning traces to handle hard budgets.	Chat template flags like `enable_thinking=True` and `medium_effort=True`. An external client can also enforce a hard token budget.
Kimi K2.5	Uses a method called 'Toggle' that alternates between budgeted and unconstrained RL phases to train a token-efficient policy without sacrificing performance on complex tasks.	A binary 'thinking' vs 'instant' mode switch, controlled via API or chat template. The 'Toggle' training improves the efficiency of the default thinking mode.
GLM-5	Introduces 'interleaved', 'preserved', and 'turn-level' thinking behaviors during multi-task SFT, later reinforced with several RL stages and on-policy distillation.	A turn-level on/off switch in the API (e.g., `thinking: {"type": "disabled"}`) or chat template.
Qwen3	A 'Thinking Mode Fusion' SFT stage trains the model on a mix of thinking (`/think`) and non-thinking (`/no_think`) examples.	A boolean flag (`enable_thinking=False`) in the tokenizer prefills an empty `<think></think>` block to disable reasoning.
Inkling	Conditions RL on a continuous effort value (0.0 to 1.0) specified in the system prompt. The token penalty in the reward function is adjusted based on this value.	A system message like `Thinking effort level: 0.8` controls the model's token usage.

Performance and Cost Trade-off:

Data from the Artificial Analysis Coding Agent Index v1.1 for the GPT-5.6 family (Luna, Terra, Sol models) shows that increasing reasoning effort (inference scaling) and using a larger model (training scaling) both improve performance but also increase API cost. The performance curves show that a smaller model at a high effort setting can sometimes match or exceed the performance of a larger model at a low effort setting, giving users multiple options to hit a desired cost/performance target.

This article provides a comprehensive synthesis of the various training techniques used by state-of-the-art LLMs to implement adjustable reasoning effort. While the concept of reasoning models is not new, this is a detailed breakdown of how user-facing controls (e.g., 'low', 'medium', 'high' effort) are created during post-training. It contrasts multiple distinct recipes—from training separate specialists (DeepSeek V4) and alternating RL budgets (Kimi K2.5) to conditioning RL on continuous values (Inkling)—drawing on the technical reports of the latest proprietary and open-weight models.

The analysis is based on publicly available technical reports and blog posts from model developers. These documents are not peer-reviewed and often omit critical implementation details, such as exact reward formulas, data mixtures, and hyperparameters, for competitive reasons. Therefore, the described mechanisms are reconstructions based on incomplete information rather than fully reproducible recipes.

Controlling Reasoning Effort in LLMs

Don't read this site daily. Get it in your inbox.