~/writing/math-foundations-for-reasoning-models $ ls
Reasoning Models as Representation Transformation
Scratchpad
A scratchpad is an external workspace where intermediate steps can be represented, inspected, and reused.
~/writing/math-foundations-for-reasoning-models $ cat topic.md
Scratchpad
Definition
A scratchpad is an external workspace for intermediate work.
Simple form:
task -> intermediate notes -> final answer
In a reasoning model, a scratchpad may be written text, structured data, or tool-visible state.
A scratchpad externalizes intermediate representation.
Why A Scratchpad Helps
Some information is difficult to keep only in hidden state.
A scratchpad lets the system store intermediate results in a visible form.
Examples:
partial calculation
list of constraints
candidate plan
tool result
draft answer
This can reduce the need to recompute or remember everything internally.
A scratchpad gives reasoning a reusable workspace.
Scratchpad vs Memory
Memory stores information across steps.
A scratchpad is one possible memory medium.
memory:
information preserved for later use
scratchpad:
external place where preserved information is written
The scratchpad makes some intermediate state explicit.
Scratchpad is externalized memory for a task.
Scratchpad vs Hidden State
Hidden state is inside the model's computation.
Scratchpad content is outside the model as visible or retrievable text/state.
hidden state:
not directly inspectable
scratchpad:
can be read again as input
This distinction matters for tool use and verification.
A scratchpad can become input to later steps.
Scratchpad Is Not Automatically Correct
Writing intermediate steps does not guarantee correctness.
The scratchpad can contain:
wrong assumptions
bad calculations
irrelevant notes
incomplete plans
It is useful because it can be inspected and revised.
A scratchpad helps structure reasoning, but it still needs checking.
Final Summary
- A scratchpad is an external workspace.
- It stores intermediate representations.
- It can act as task-level memory.
- It differs from hidden state because it can be read again.
- It is useful but not automatically correct.
~/writing/math-foundations-for-reasoning-models $ open
Continue the series.
Chain of Thought