~/writing/math-foundations-for-reasoning-models $ ls
Reasoning Models as Representation Transformation
Self-Correction
Self-correction is the process of using feedback or verification to revise a model's previous output or intermediate state.
~/writing/math-foundations-for-reasoning-models $ cat topic.md
Self-Correction
Definition
Self-correction is revising a previous output or intermediate state after detecting a problem.
Simple form:
candidate -> check -> identify issue -> revise
It is a feedback loop.
Self-correction uses evaluation to update behavior.
Why Self-Correction Matters
Multi-step reasoning can produce errors.
If the system can detect an error, it can try to repair it.
Examples:
wrong calculation -> recompute
bad plan -> adjust steps
failed test -> edit code
missing constraint -> revise answer
Self-correction helps recover from intermediate mistakes.
Correction Requires A Signal
The model needs some signal that something is wrong.
Signals may come from:
validation check
tool output
test failure
user feedback
contradiction
constraint mismatch
Without a signal, revision may be arbitrary.
Correction depends on useful feedback.
Self-Correction Is Not Guaranteed
A model can revise in the wrong direction.
Examples:
misdiagnose the issue
change a correct part
ignore the real constraint
become more verbose without becoming more accurate
Correction needs verification, not just another response.
Self-correction works best when grounded by checks.
Correction Loop
A simple reasoning loop:
generate
verify
revise
verify again
finalize
This loop treats reasoning as iterative state improvement.
Each pass updates the representation of the problem and solution.
Self-correction is iterative representation refinement.
Final Summary
- Self-correction revises earlier output or state.
- It depends on feedback or verification.
- It can repair errors in multi-step tasks.
- It can also fail if the correction signal is weak.
- Strong correction loops combine generation, verification, and revision.
~/writing/math-foundations-for-reasoning-models $ open
Continue the series.
Symbolic Reasoning vs Learned Reasoning