Verification-First Self-Repair for LLM Agents
A Detect–Repair–Verify Benchmark, Runtime Wrapper, Cross-Model Replication, and Component Ablation
CRepair studies how LLM agents recover from task-state inconsistencies in long-horizon workflows. Instead of measuring only final task success, it asks whether an agent can detect a failure, repair it, explicitly verify that recovery succeeded, and avoid creating a new inconsistency.
Key findings
- Verification is the load-bearing step in structured self-repair.
- A lightweight CRepair wrapper improved repair-loop closure across Claude Sonnet, Gemini 2.5 Flash, and GPT-4o.
- Generic retry was much weaker than targeted repair/verification scaffolding.
- Detection-only prompting can make performance worse if there is no downstream repair or verification path.
- Results are pilot-scale and require human validation.
Preprint. Not peer-reviewed. Pilot studies; findings require human validation before use in production settings.