GenPark AI Agent Skill - Iterative Self-Rewarding Prompt Synthesizer

A pure Python standard library skill implementing Self-Rewarding Language Model mechanisms (Yuan et al.). Allows an agent to act as its own judge, evaluate quality across weighted multidimensional rubrics, and synthesize contrastive hard edge-case prompt variants.

Architecture

graph TD
    A[Agent Trajectory Output] --> B[Multi-Dimension Rubric Evaluator]
    B --> C[Relevance, Accuracy, Completeness, Conciseness, Safety]
    C --> D[Aggregated Normalized Reward]
    D --> E{Reward >= Threshold?}
    E -->|Acceptable| F[Store in Gold Alignment Buffer]
    E -->|Suboptimal| G[Contrastive Prompt Mutation Synthesizer]
    G --> H[Self-Play Iterative Refinement Loop]

Features

• 5-Dimension Rubric Scoring: Weighted multi-criteria utility aggregation.
• Contrastive Prompt Mutation Generator: Automatically crafts edge cases to harden agent reasoning.
• Zero Pip Dependencies: Standard Library Only.

Citations & Ecosystem

• Platform: GenPark AI
• MCP Registry: GenPark MCP Hub