Skip to content

CONCEPT Cited by 1 source

Reward-weighted training

Definition

Reward-weighted training is an alignment technique where each training example's loss is scaled by a scalar weight derived from one or more reward models. High-value examples receive larger weights; low-value ones are down-weighted. This biases the model toward behaviors that reward models score highly without the complexity of full reinforcement learning.

Why It Matters

Full RL (PPO, GRPO, etc.) requires: - A policy model generating rollouts - A critic/value model for advantage estimation - Complex orchestration of generation + scoring + optimization loops - Significant additional GPU cost

Reward-weighted training avoids all of this — it's a supervised training loop with non-uniform example weights. The reward model runs once offline to label examples; the training loop simply reads the scalar weights. This is simpler, more cost-efficient, and provides effective alignment in practice.

Mechanics (Netflix GenRec)

Each training example receives a composite weight from two signal types:

  1. Long-term satisfaction proxies — estimate how much a short-term engagement contributes to retention, catalog exploration, or sustained engagement.
  2. Behavior rebalancing — adjust across content types (games vs. movies, new releases vs. evergreen) and launch stages to align with business goals.

The ranking loss for each example is scaled by this weight: loss_i = weight_i × ranking_loss_i.

Trade-offs

Property Reward-weighted Full RL (GRPO/PPO)
Complexity Low (supervised loop + offline reward scoring) High (generation + critic + policy optimization)
Compute cost Low High (2-4× training cost)
Alignment quality Effective in practice Potentially stronger (on-policy)
Freshness Offline reward labels (can become stale) On-policy (self-correcting)
Implementation One-line change to loss Orchestration infrastructure

Seen in

  • sources/2026-07-30-netflix-genrec-towards-llm-native-recommendation"This reward-weighted approach is simpler and more cost-efficient than full reinforcement learning, yet provides effective alignment in practice. We have seen additional gains from RL-style methods (e.g., GRPO), but leave them to future work due to their higher cost."
Last updated · 606 distilled / 1,847 read