Comparisons
Post-training method comparisons
Side-by-side comparisons of post-training methods and data sources: RLVR vs RLHF, rubric rewards vs reward models, expert vs crowd annotation, synthetic vs human preference data, and process vs outcome rewards.
RLVR vs RLHF: Verifiable Rewards or Human Preference Rewards?
How reinforcement learning with verifiable rewards differs from RLHF in reward source, cost, failure modes, and the tasks each suits, with primary sources.
Rubric Rewards vs Reward Models: Which Should Score Your RL Policy?
Explicit, per-prompt rubrics graded by an LLM judge versus a learned scalar reward model: interpretability, cost, hacking risk, and when to use each.
Expert vs Crowd Annotation for AI Training Data
When credentialed domain experts are worth the cost over crowdsourced annotators, and how to measure label quality with agreement statistics.
Synthetic vs Human Preference Data (RLAIF vs RLHF Labels)
AI-generated preference labels versus human labels: cost, quality evidence, bias risks, and how labs combine them.
Process vs Outcome Rewards (PRM vs ORM)
Step-level process reward models versus final-answer outcome rewards: labeling cost, credit assignment, and evidence from math reasoning.