Model Adaptation · Software component

Ensemble Reward Scorer

Software componentModel AdaptationModelsarc:EnsembleRewardScorer

A reward scorer that combines scores from multiple independent reward models trained on different data or with different architectures into one reward signal.

Responsibility. Produces a reward that a policy can only exploit by fooling several reward models simultaneously.

Also known as: Ensemble reward models

Variant of Reward Scorer abstract

When to choose. Choose when reward hacking via single-pattern exploitation is a concern and the cost of training multiple reward models is acceptable.

hostsspecializesalternative toalternative toReward Model: hostsReward ModelReward Scorer: specializesReward ScorerPreference Reward Scorer: alternative toPreference Reward ScorerSegment-Routed Reward Scorer: alternative toSegment-Routed Reward Sc…
Direct neighbourhood (hover for relationship types)

Relationships

hosts structural

alternative to variability

Design guidance

Classification

Risks mitigated
Reward hackingSpecification gaming

Sources

  1. Ch10.3: T. Nguyen, "RLHF Methodology," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 10.3. ISBN: 9798244538229.