Map

Controllable Thinking Style

Wiki conceptllmprompt-engineeringreasoning-models ↳ show in map Markdown
title
Controllable Thinking Style
type
concept
summary
Training-injected markers that toggle the style of a reasoning model's chain-of-thought without changing the answer pipeline
tags
llm, prompt-engineering, reasoning-models
sources
deepseek-v4-roleplay-instruct
created
2026-04-25
updated
2026-04-25

A pattern in modern reasoning models where specific marker phrases, injected at a fixed position in the conversation, switch the style of the model's hidden thinking trace (<think> block) without changing the answer-generation pipeline. The marker is trained — the model has seen training examples where the marker appears at a specific location and the assistant responds with the matching thinking style. At inference time, supplying the marker probabilistically reproduces that style.

The clearest documented example is DeepSeek-V4's roleplay markers, which toggle between two thinking styles for the same answer task: first-person immersive inner monologue versus third-person scene-and-strategy planning.

How it differs from prompt engineering

Ordinary prompt engineering ("think step by step", "be concise", "play the role of X") works at the level of natural-language instruction. The model interprets the instruction and adjusts as best it can, generalizing from broad training data.

A trained style marker works at the level of training-data injection. The model has seen many conversations where exactly this marker, in exactly this position, preceded exactly this output style. Triggering it is closer to invoking a stored mode than to giving an instruction the model has to interpret.

The properties that follow:

  • Position-sensitive. The marker only triggers reliably at the position it was injected during training. The DeepSeek guide specifies "end of the first user message"; putting the same text in a system prompt or a later turn loses reliability.
  • Wording-sensitive. Close paraphrases lose reliability. The literal string the model was trained on is what triggers the mode.
  • Probabilistic, not enforced. Unlike a control token (which deterministically gates behavior), a trained marker just biases generation. Re-rolling sometimes helps.
  • Persists in conversation history. Once the marker is in the conversation, the model continues to produce the matching style on later turns without needing the marker re-applied.

What this implies about training

If a model reliably switches thinking style on a specific marker injected at a specific position, the natural explanation is that the training data contained two distinct subsets — one with the marker present at that position and the matching output style, one without. The model learned the conditional. This makes a marker like the DeepSeek roleplay one a probe of training-data construction: it reveals that the lab built a dataset of paired (marker, thinking-style) examples for at least two distinct CoT modes.

This contrasts with the more common practice (visible in OpenAI o1/o3, Anthropic's extended-thinking modes) of having a single thinking style determined by model variant or by a model-side toggle, with no user-controllable style switch.

Why labs might build this

  • A/B testing in production data. Two thinking styles can be evaluated empirically on user satisfaction, with the marker as the assignment.
  • Domain-specific fine-tuning. Roleplay, code, math, and analysis benefit from very different thinking-trace styles. Having user-controllable style avoids forcing one mode globally.
  • Internal vs external personas. The character the model is playing wants one style of reasoning; the meta layer reasoning about the user wants another.
  • Format leakage management. If the thinking trace is exposed (DeepSeek shows it; OpenAI hides it), letting the user pick the style they want to see matters.

Why labs might not document it

These markers tend to surface organically from community reverse engineering rather than official docs:

  • Lab doesn't want users to rely on undocumented behavior they may change.
  • Markers may be a leftover artifact of training-data construction the lab doesn't consider a feature.
  • The capability competes with the lab's own product surface (Expert Mode toggle, etc.).
  • Probabilistic behavior is hard to support in an official API contract.

Adjacent patterns

  • Trained system-prompt format. Most chat models work best with system prompts in a specific shape because that's what they were fine-tuned on. The first-user-position marker is the same idea applied to user messages.
  • Anthropic's <thinking> tag prompt. Asking Claude to use <thinking> tags works because XML-tag-shaped CoT was prevalent in training data. Not the same as a deterministic control token, but the same flavor of "trained-in shape."
  • Cycle-conditioning markers in image models. Stable Diffusion's negative-prompt and quality-tag conventions emerged from training-data structure rather than designed controls.
Linked from