Determine whether large language models can self-correct rule violations without fine-tuning

Determine whether large language models can autonomously self-correct violations of formal rules without specific fine-tuning, as assessed across reasoning tasks.

Background

The authors synthesize findings on LLM reasoning limits, noting uncertainty about self-correction capabilities in the absence of fine-tuning. This unresolved question bears on the reliability of LLMs in formal domains and their potential for robust reasoning.

References

Models are unlikely to know when they are violating formal rules and it is unclear whether they can self-correct~\, but with specific fine-tuning they might self-correct against harmful text~\, and that training on generated data might not be the best approach to preserve reasoning about outlier cases~.

— Functional Benchmarks for Robust Evaluation of Reasoning Performance, and the Reasoning Gap  (2402.19450 - Srivastava et al., 2024) in Related Work, Understanding the bounds of reasoning, generalization, and memorization in large language models

Whether self-correction and oscillation recur outside IaC is unexamined. These are hypotheses: confirming them requires replications that vary domain and validator while holding the repair loop fixed.

— Does Fixing Break Security? An Empirical Study of Security Degradation in Iterative LLM-Driven Infrastructure-as-Code Repair  (2608.13404 - Agyekum et al., 13 Aug 2026) in Section 7, Discussion, subsection “What Generalizes Beyond Terraform and Checkov?”

We conjecture that tasks gated by private conventions will remain highly resistant to this loop. Because nothing in the task's instruction or environment reveals what the planted thresholds or internal normalisation tables are, self-critique alone is unlikely to recover them. However, this durability strictly applies to non-derivable construction-time knowledge; operator gates may remain susceptible to advanced tool-use, search, or scaled test-time compute.

— Ignorance or Incompetence? Constructing Knowledge-Gated, Verifiable Tasks for LLM Agents  (2608.30322 - Tian et al., 31 Aug 2026) in Section 5, subsection “Durability under test-time self-improvement”