A self-improving, human-centered agent framework for thematic analysis.
CentaurTA adopts an Actor–Critic architecture and performs prompt-level optimization guided by expert feedback, enabling structured human–AI collaboration while preserving human interpretive authority.
A constraint-based evaluation protocol for open coding and theme construction.
Our rubric framework generates fine-grained, actionable feedback signals that go beyond coarse LLM-as-Judge metrics, supporting reliable alignment and iterative improvement.
We demonstrate that:
- Iterative human feedback significantly improves alignment.
- Rubric-based early stopping helps prevent overfitting during self-improvement.
- Learned principles exhibit cross-platform transferability.
The full project source code, rubric library, and sample data will be released contingent upon paper acceptance.