Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!

Models such as DeepSeek-R1 are trained to write a long sequence of tokens before their final answer. Training usually checks only the answer: if it is right, the whole sequence is made more likely, whatever the tokens in between said. These tokens are commonly called a chain of thought, a reasoning trace or thinking.

This position paper argues the names are wishful, have little evidence behind them, create false confidence and send research in unproductive directions. It separates three things: the unfiltered intermediate tokens (the subject), summaries written afterward for users (not the subject), and tool calls (which have meaning outside the model and are excluded).

The evidence is mostly from the authors' group. Transformers trained on A* search traces for mazes produced traces whose validity was only loosely related to whether the final path was correct, especially on unfamiliar mazes. Models trained on traces swapped between unrelated problems solved problems as well or better, and accuracy dipped only when correct and swapped traces were mixed. Reinforcement learning raised accuracy without raising trace validity. In question answering, fine-tuning on incorrect traces paired with correct answers beat fine-tuning on correct traces. A model trained on hard mazes wrote extremely long traces for empty grids. In a user study, showing traces or summaries made people more likely to accept wrong answers.

The recommendation: get trust from verifying the answer, stop treating trace length as effort, and think of intermediate tokens as prompt augmentations the model learned to give itself.

Customers ask to show the model's reasoning to their users, to store it as an audit record, or to use it to decide which answers to trust. This paper is the argument against each, with sources.

A trace is not an audit record. If the training signal only rewarded the final answer, nothing forced the intermediate tokens to describe how the answer was produced. Log the inputs, retrieved documents, tool calls and outputs, which do have meaning outside the model, and do not present the trace to a regulator as the decision rationale.

Showing traces can raise trust in wrong answers. The cited user study found traces and summaries increased acceptance of answers regardless of correctness. In workflows where users cannot check the answer, a fluent trace is a risk.

Verification has to be outside the model. Schema checks, calculations, retrieval citations, unit tests, solvers or a human reviewer. The paper's own framing comes from its authors' generate-and-test work: the model proposes, something with guarantees checks.

Hidden tokens are billed. Several providers charge for intermediate tokens they do not show, so length is a cost line, not a quality signal.