Model Adaptation · Software component
Composite Reward Scorer
Software componentModel AdaptationModelsarc:CompositeRewardScorer
A reward scorer, often itself an agent, that synthesizes human-preference scores with factuality verification, instruction-following metrics and safety-filter signals into one reward.
Responsibility. Grounds preference rewards with objective correctness and safety signals.
Also known as: Agentic reward modeling, Reward agent, Hybrid reward function
Variant of Reward Scorer abstract
When to choose. Choose when human preference must be grounded by verifiable correctness, e.g., to prevent optimizing toward fluent hallucinations or annotator bias.
Relationships
invokes dependency
- Factuality Verifier abstract Ch3.5
alternative to variability
Design guidance
- SHOULD combine human preference data with factuality checks, logical-consistency validation and domain-expert verification.
Classification
- Patterns
- Agentic reward modelingMulti-signal reward synthesis
- Quality attributes
- Functional suitability: correctness and validity (ISO/IEC 25010 | NIST AI RMF: valid)Safety (ISO/IEC 25010 | NIST AI RMF: safe)
- Risks mitigated
- Optimizing toward human manipulationReward models exploiting annotator biasesFluent confident hallucinations scoring highly
Sources
- Ch3.5: T. Nguyen, "Prompt Optimization, Few-Shot Learning, Fine-Tuning," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 3.5. ISBN: 9798244538229.