contrastive-reflection

Introduces Contrastive Reflection, an iterative prompt-optimization framework for agentic IR workflows that uses structured traces to identify error-anchored behavioral slices and applies contrastive repair via a Teacher LLM, achieving significant improvements on HotpotQA.

Contrastive Reflection (CORE) is a non-parametric algorithm that generates concise, interpretable insights from comparing successful and unsuccessful reasoning traces, enabling faster and more efficient self-improvement for language models with fewer samples and rollouts than existing methods.