Self-correction Failure
A failure arising when a model attempts or is asked to detect, reconsider, or correct an error in its own answer or reasoning but does not correct it, changes something that was already correct, or makes the result worse.
Example
After being asked to review an arithmetic answer, the model repeats the same incorrect result without fixing it.
Distinctions and cautions
An initial reasoning error is related to Reasoning Failure; Self-correction Failure instead concerns what happens when the model attempts, or is asked, to detect, reconsider, or correct an error in its own answer or reasoning. Verification is the process of checking with additional evidence or tests, not self-correction itself. Self-critique or regeneration does not necessarily improve the result.
Evidence
View evidence →
Large Language Models Cannot Self-Correct Reasoning Yet
- Title
- Large Language Models Cannot Self-Correct Reasoning Yet
- Author / organization
- Huang et al.
- Year
- 2024
- Source type / publication status
- ICLR; peer reviewed
- Supported concepts / claim context
- Self-correction Failure; Intrinsic self-correction without external feedback can fail or degrade answers.
- Scope and limitations
- Reasoning tasks and models tested; does not rule out trained or externally verified correction systems.
- Access / version note
- Published ICLR.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
ProgCo: Program Helps Self-Correction of Large Language Models
- Title
- ProgCo: Program Helps Self-Correction of Large Language Models
- Author / organization
- Song et al.
- Year
- 2025
- Source type / publication status
- ACL; peer reviewed
- Supported concepts / claim context
- Self-correction Failure; Self-verification failure and program-supported mitigation.
- Scope and limitations
- Three instruction/math benchmarks; method-specific; pseudo-program verification changes the nature of the intervention.
- Access / version note
- Published ACL.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
S²R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning
- Title
- S²R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning
- Author / organization
- Ma et al.
- Year
- 2025
- Source type / publication status
- ACL; peer reviewed
- Supported concepts / claim context
- Self-correction Failure; Training for improved self-verification and self-correction.
- Scope and limitations
- Three base models and in/out-of-domain benchmarks; trained capability is not spontaneous intrinsic self-correction.
- Access / version note
- Published ACL.
- Recorded official URL / DOI / identifier
- Official URL / recorded link
LLM Self-Correction with DeCRIM
- Title
- LLM Self-Correction with DeCRIM
- Author / organization
- Ferraz et al.
- Year
- 2024
- Source type / publication status
- Findings EMNLP; peer reviewed
- Supported concepts / claim context
- Instruction-following Failure; Self-correction Failure; Multi-constraint failure; critic-assisted refinement; self-correction.
- Scope and limitations
- RealInstruct, IFEval, GPT-4/Mistral and related conditions; critic feedback means this is not pure intrinsic self-correction.
- Access / version note
- Published Findings EMNLP.
- Recorded official URL / DOI / identifier
- Official URL / recorded link