Article · 2026-07-18 · 6 moments

Controlling Reasoning Effort in LLMs

How LLMs Learn Low-, Medium-, and High-Effort Reasoning Modes ✦ AI generated

Definition

A 'reasoning model' is one that outputs an intermediate reasoning trace, working through a task step by step — it does not literally reason the way humans do, just as artificial neural networks don't literally work like biological ones.

Raschka defines 'reasoning model' as a model that produces an intermediate reasoning trace, cautioning against taking the term literally as human-like reasoning.

transcript

Sebastian Raschka: 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.

extends · 1

Mechanism

During DeepSeek-R1's RLVR training, the reasoning trace itself was not used as a training signal — the reward was based only on the final answer's correctness and format, since including the trace in training was reported to not help.

Raschka explains that RLVR rewards only the final answer and format, not the intermediate reasoning trace, per the DeepSeek-R1 paper's findings.

transcript

Sebastian Raschka: Notably, the reasoning trace itself was not used for training or updating the model. Although they tried to use this intermediate response information for training, the DeepSeek-R1 paper reported that it wasn't helpful for the model training, so it was ultimately not used.

Anecdote◆

GPT-5's Auto mode, which tried to automatically select reasoning effort, largely failed in practice and was quietly removed from the UI.

Raschka notes that automatic effort selection is still an unsolved problem, citing GPT-5's Auto mode as an example that was mostly a miss and got pulled from the interface.

transcript

Sebastian Raschka: We saw this a while back with GPT 5's Auto mode. It's a tricky problem to solve, and in the end, the implementation was probably more miss than hit, which is why it got removed from the UI (at least, I can't find it anymore).

Data

Model size and reasoning effort are two separate scaling knobs whose performance curves overlap — a smaller model run at higher reasoning effort can sometimes match a larger model run at lower effort.

Analyzing GPT-5.6 benchmark curves, Raschka finds that training-scale (model size) and inference-scale (reasoning effort) trade off against each other, with smaller high-effort models rivaling larger low-effort ones.

transcript

Sebastian Raschka: As expected, both approaches can improve the benchmark score, but they also increase the cost. More interestingly, the curves overlap. For instance, a smaller model at a higher reasoning effort can sometimes reach a similar score as a larger model at a lower reasoning effort.

gives example · 1supports · 1

Mechanism

Reasoning effort levels are typically implemented one of two ways: applying different length penalties during RLVR keyed to the system-prompt effort label, or fine-tuning the model after RLVR via supervised fine-tuning on effort-labeled target responses.

Raschka lays out the two plausible training mechanisms behind reasoning-effort toggles: RLVR with effort-dependent length penalties, and SFT on effort-labeled data.

transcript

Sebastian Raschka: First, we can implement it as part of the RLVR process and apply a different length penalty when different system prompts are used. For example, a high length penalty when "Reasoning effort: low" and a mild or no penalty when "Reasoning effort: high". Second, we can fine-tune the model after RLVR to follow different effort instructions via supervised fine-tuning (SFT).

Claim◆

The <think></think> tags are purely cosmetic markers for separating the reasoning trace from the final answer; they do not confer or improve reasoning ability, and a model trained without them would likely perform just as well.

Raschka argues that think tokens are cosmetic formatting devices, not the source of a reasoning model's ability to reason.

transcript

Sebastian Raschka: 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.

explains mechanism · 1rebuts · 1

Highlight slides

<think> Tags Are Cosmetic, Not Causal✦ from: The <think></think> tags are purely cosmetic markers for separating the reasoning trace from the final answer; they do not confer or improve reasoning ability, and a model trained without them would likely perform just as well.GPT-5's Auto Mode: A Cautionary Tale✦ from: GPT-5's Auto mode, which tried to automatically select reasoning effort, largely failed in practice and was quietly removed from the UI.Same Model, No Tags: Similar Results Expected✦ from: The <think></think> tags are purely cosmetic markers for separating the reasoning trace from the final answer; they do not confer or improve reasoning ability, and a model trained without them would likely perform just as well.Why It Failed✦ from: GPT-5's Auto mode, which tried to automatically select reasoning effort, largely failed in practice and was quietly removed from the UI.

Related episodes

Article · Latent Space

[AINews] How to steal a Reasoning Trace

→

Video · Syntax

I Built an LLM From Scratch

→

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→