Reasoning Traces

When a chat program receives a difficult task, it can answer immediately or first write down intermediate steps. These intermediate steps are called reasoning traces. The program formulates sentences like “First I’ll calculate the cost per unit” and works its way forward step by step. Only afterward does it output the actual answer. The comparison with a math notebook fits well: a student who writes down the calculation steps makes fewer mistakes than one who only writes down the result. It works the same way here – except that the thought trace consists of ordinary text that can be read along.

Why thinking before answering pays off

Language models generate text word by word. For each individual word, only a limited amount of computational work is available. A task with five reasoning steps therefore doesn’t fit into a single word. When the model writes out the steps, it distributes the work across many words. Each intermediate result already written then serves as the basis for the next step.

The effect is measurable. On math problems, logic puzzles, and programming tasks, success rates rise significantly when a model calculates first and answers afterward. This is exactly what the so-called reasoning models, which have been coming onto the market since 2024, are based on. They are specifically trained to produce long reasoning traces before stating a result.

This doesn’t come for free. Every word of the thought trace has to be computed and costs time and electricity. A model that thinks a thousand words before answering is correspondingly slower and more expensive to run. For simple questions, the effort isn’t worth it – nobody needs a calculation path for “What time is it in Tokyo?”

How models are taught to think

At first, reasoning traces were simply forced through an instruction. The sentence “Think step by step” in the prompt was often enough to make a model spell out its calculation path. This technique is called chain-of-thought prompting. It works because the training data contains countless solution paths from textbooks and forums.

Newer systems no longer need this prompt. They are trained using reinforcement learning: the model solves thousands of tasks for which the correct solution is known. If a reasoning trace leads to the correct result, this behavior is rewarded and reinforced. If it leads astray, it is weakened. Over many runs, the model thus learns on its own to check approaches, notice mistakes, and start over from scratch.

A common misconception should be cleared up here: the reasoning trace is not a record of what actually happens inside the model. It is self-generated text and can diverge from the actual computation. Researchers have shown cases where a model writes down a plausible path but actually derives its answer from a hidden hint in the question. A reasoning trace is thus a tool, not a lie detector.

Hidden reasoning traces in chatbots and the news

In common chat programs, the reasoning trace often appears as a collapsed box labeled “Thinking” or “Reasoning.” You can expand it and read along as the system breaks down the task. Some providers, however, only show a shortened summary. The reason is economic in nature: reasoning traces are valuable training material for the competition.

This was exactly what a dispute in early 2025 was about. After the success of the Chinese model DeepSeek R1, suspicion arose that its training had built on the reasoning traces of other systems. Such reports only make sense once you know that reasoning traces have become a tradable resource.

The term also comes up in pricing and benchmarks. Anyone using a model via a programming interface usually pays for the thought-out words too, even though they don’t appear in the answer. And when providers show results from math or programming exams, the top scores almost always come from models with long reasoning traces.