← Back to Library

Controlling reasoning effort in LLMs

Sebastian Raschka cuts through the marketing hype surrounding the latest wave of artificial intelligence models to reveal a fundamental shift in how these systems operate: the ability to dial up or down their own cognitive effort. While the industry fixates on model size, Raschka argues that the real breakthrough lies in inference scaling—spending more compute time during the actual usage of the model rather than just during its training. This distinction is critical for busy professionals who need to understand not just what these tools can do, but how to deploy them efficiently without burning through resources on simple tasks.

The Illusion of "Thinking"

Raschka begins by dismantling a pervasive misconception. He writes, "When talking about 'reasoning models', we shouldn't expect that these models literally reason like us humans." This is a vital correction. The term "reasoning" in this context is technical shorthand for a specific output pattern, not a replication of human consciousness. He explains that these models generate an intermediate "reasoning trace," a step-by-step breakdown of a problem that occurs before the final answer is delivered.

Controlling reasoning effort in LLMs

The author is careful to separate the mechanism from the mystique. "These '' tags are cosmetic with respect to reasoning ability," Raschka notes, pointing out that the specific delimiters used to hide this internal monologue are arbitrary. They are merely formatting tools to help user interfaces distinguish between the model's internal deliberation and its final response. The real magic isn't in the tags; it's in the training process that encourages the model to backtrack and self-correct.

This framing is particularly effective because it demystifies the "black box" nature of modern AI. By focusing on the output structure rather than the internal "thought," Raschka grounds the discussion in observable engineering realities. However, one might argue that by dismissing the human analogy so quickly, the piece risks underplaying the emergent behaviors that make these traces feel genuinely intelligent to the end user, even if they are technically just statistical predictions.

"One could train the same model without these delimiters and likely reach similar benchmark performance."

The Mechanics of Effort

The article's core contribution lies in its explanation of how models now toggle between different levels of cognitive intensity. Raschka traces the lineage from early, rigid reasoning models that were verbose by default to the new hybrid architectures that can adapt on the fly. He highlights the recent GPT-5.6 family, which offers multiple reasoning-effort settings, as the culmination of this trend.

"In short, training a model with reinforcement learning with verifiable rewards is already implicitly leading to a form of inference scaling," Raschka writes. This connects the current trend to the broader history of the field, echoing the principles found in deep dives on neural scaling laws where increasing compute often yields diminishing returns unless applied strategically. The key innovation here is the ability to control that compute at inference time.

Raschka details how this is achieved through techniques like "Thinking Mode Fusion," where models are trained to recognize signals to either engage in deep deliberation or provide a quick, direct answer. He describes how a simple system prompt can alter the model's behavior: "The effort level seems directly correlated to token usage, which in turn seems correlated to accuracy." This suggests a future where users can explicitly balance speed against precision, a feature that transforms these tools from static calculators into dynamic partners.

Critics might note that the lack of transparency from major developers regarding the exact implementation of these effort settings leaves room for skepticism. Without open benchmarks, it is difficult to verify if a "low" effort setting is truly safe for complex tasks or merely a cost-saving measure that sacrifices reliability.

The Trade-off Between Speed and Depth

The most practical insight Raschka offers is the recognition that reasoning is a resource, not a binary state. He illustrates this with data showing how response length and accuracy scale together under different effort settings. "By the way, note how different effort settings scale the response length in the figure above," he observes, emphasizing the direct link between the time a model spends "thinking" and the quality of the result.

This reframes the user's relationship with AI. Instead of hoping a model is smart enough to get it right on the first try, users can now instruct the system to spend more time on critical problems. Raschka points to the evolution from models like DeepSeek-R1, which required a separate architecture for reasoning, to systems that integrate these capabilities seamlessly. "The first generation of reasoning models was dedicated reasoning models... Later models... experimented with hybrid approaches," he writes, marking a clear inflection point in the industry.

The implication for enterprise users is profound. It suggests that the future of AI deployment will be less about buying the biggest model and more about configuring the right "effort" profile for each specific workflow. This aligns with the broader trend of optimizing for efficiency, a lesson learned from the early days of reinforcement learning from human feedback where the cost of training had to be balanced against the quality of the reward signal.

"The effort level seems directly correlated to token usage, which in turn seems correlated to accuracy."

Bottom Line

Sebastian Raschka's analysis succeeds in shifting the conversation from the mystique of "AI thinking" to the engineering reality of compute allocation. The strongest part of his argument is the clear delineation between cosmetic formatting and the actual mechanics of inference scaling, providing a practical framework for understanding model behavior. The biggest vulnerability remains the opacity of the major players; while the theory of adjustable reasoning effort is sound, the lack of public verification on how these settings impact safety and accuracy in high-stakes scenarios is a gap that needs closing. Readers should watch for how these effort controls are standardized across the industry, as this will likely become the primary interface for interacting with advanced AI systems.

Deep Dives

Explore these related deep dives:

  • Reinforcement learning from human feedback

    Understanding this foundational method is essential to grasp how the article's discussed 'verifiable rewards' system evolved from human preference tuning to automated mathematical verification.

  • Neural scaling law

    This concept directly illuminates the article's central argument about 'inference scaling,' where models trade increased computational cost during the answering phase for higher accuracy.

  • Prompt engineering

    This technique explains the specific mechanism of generating the intermediate 'reasoning traces' that the article identifies as the defining feature of modern reasoning models, distinguishing them from conventional LLMs.

Sources

Controlling reasoning effort in LLMs

by Sebastian Raschka · Ahead of AI · Read full article

It has been almost two years since OpenAI released o1, a model that popularized the idea of LLM-based reasoning models. DeepSeek-R1 followed about four months later, together with details of a reinforcement learning with verifiable rewards (RLVR) recipe to train such reasoning models.

Last week, OpenAI released the GPT-5.6 model family. It comes in three sizes, each with roughly five or six reasoning-effort settings.

So yes, reasoning models are here to stay. They have become a standard part of modern model releases.

In the past, I covered the methodology of reasoning models (Understanding Reasoning LLMs) as well as relevant research papers (The State of Reinforcement Learning for LLM Reasoning and The State of LLM Reasoning Model Inference). And I even wrote a whole new 440-page book on how to develop reasoning models, Build A Reasoning Model (From Scratch).

These resources have focused on turning a conventional LLM into a reasoning model. Now, in this article, I want to focus on and explain how to develop a reasoning model that has multiple effort modes, similar to what’s shown in the figure at the beginning of this article.

No worries, this article can be read as a standalone article. However, the aforementioned resources may be interesting and useful.

1. A brief definition of reasoning models.

When talking about pretty much any machine learning or AI technique or subfield, the one lesson is that we usually shouldn’t take technical terms “literally”. For example, an (artificial) neural network in machine learning and AI doesn’t literally work like a biological neural network like the human brain.

Similarly, when talking about “reasoning models”, we shouldn’t expect that these models literally reason like us humans. In the context of AI and LLM research, “reasoning model” means a model that outputs an intermediate reasoning trace, which is like an intermediate response that works through a question or task step by step.

It’s probably easiest to explain this by showing an example.

2. A brief overview of training and inference scaling reasoning models.

There are essentially two ways to improve (reasoning) task performance: training scaling and inference scaling.

Let’s briefly talk about training first.

2.1 Training reasoning models.

In a nutshell, DeepSeek-R1 proposed training an LLM using reinforcement learning with verifiable rewards (RLVR) to turn it into a reasoning model. RLVR is a technique to provide a reward signal (0=incorrect and 1=correct) for verifiable data domains. These ...