C.W.K.
Stream
Lesson 03 of 05 · published

Visible, Hidden, and Summarized Thinking

~9 min · reasoning, ux, transparency

Level 0Scout
0 XP0/41 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete

Three ways the model's internal thoughts can reach the user

Visible chain-of-thought

The user sees the full reasoning trace — every intermediate token. DeepSeek-R1 emits a literal <think>...</think> block before the final answer, and you can read all of it. Early Claude extended thinking also exposed thinking content directly. Visible CoT is the most transparent option and the most expensive in terms of "look at this giant wall of internal monologue".

Hidden reasoning

The model generates thinking tokens internally but the user only sees the final answer. OpenAI o1 and o3 use this approach — the reasoning tokens count toward billing but are not exposed to the API consumer (except as a summarized stream in some cases). Hidden CoT is cheaper from a UX standpoint (no scrolling through internal monologue) but trades transparency for tidiness.

Summarized reasoning

A condensed, human-readable summary of the reasoning is shown alongside the final answer. Anthropic's Claude provides summarized thinking blocks for some configurations — you see the gist of how the model approached the problem without the full token-level trace. Best of both worlds for many production UIs.

Why the visibility choice is consequential

  • Trust: visible CoT lets users sanity-check the reasoning. Hidden CoT requires trusting the final answer with no audit trail.
  • Faithfulness: research shows thinking tokens are only partially faithful to the model's actual internal computation. The model may reach the right answer for reasons not reflected in the visible thinking.
  • UX: visible CoT is overwhelming for short interactions. Summarized is gentler. Hidden is cleanest but opaque.
  • Cost transparency: visible CoT makes thinking-token cost obvious. Hidden CoT can surprise developers who didn't realize how many thinking tokens were billed.

The faithfulness caveat

You cannot fully trust visible thinking as a window into the model's actual reasoning process. Anthropic and others have shown that models sometimes produce confident-sounding thinking that does not match the true cause of the answer. Visible CoT is a useful but imperfect lens — treat it as a hint, not a guarantee.

External links

Exercise

Run the same hard math question on a visible-CoT model (DeepSeek-R1) and a hidden-CoT model (OpenAI o3 or similar). Compare the experiences. Which did you trust more, and why? Now check if the answers were both correct — does your trust correlate with correctness? This is calibration on faithfulness.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.