Stanford CS336 Language Modeling from Scratch | Spring 2026 | Lecture 16: Post-Training - RLVR

| Podcasts | May 27, 2026 | 458 views | 1:15:51

TL;DR

This lecture explains why RLHF hits overoptimization limits with learned reward models, and how RLVR (Reinforcement Learning from Verifiable Rewards) enables unlimited compute scaling on verifiable tasks like math and coding through simpler algorithms like GRPO.

🎯 The Shift to Verifiable Rewards 3 insights

RLHF hits an overoptimization wall

Training against learned reward models inevitably overfits to preference data, creating a hard annotation bottleneck that limits how much compute can be applied.

Verifiable rewards unlock unlimited scaling

Hard verification signals like math correctness or code execution provide ground-truth objectives similar to AlphaGo, allowing RL to scale indefinitely without overfitting.

Breakthroughs in thinking models

OpenAI's recent solutions to open math problems demonstrate how RLVR enables extended reasoning chains through long-context training on verifiable objectives.

⚠️ PPO Implementation Challenges 3 insights

Complexity and sensitivity

PPO requires navigating dozens of implementation details and hyperparameters, with small engineering choices drastically altering optimization outcomes.

The value network burden

PPO requires training a separate value model as large as the policy itself, consuming significant memory that could otherwise support larger models or inference.

Common degenerate configurations

Many practitioners unknowingly reduce PPO to a bandit algorithm by setting gamma=lambda=1, destroying the temporal structure the algorithm was designed to capture.

🚀 GRPO: The Simpler Alternative 3 insights

Eliminating the value network

GRPO removes PPO's most complex component by estimating advantages as z-scores across groups of outputs sampled from the same prompt.

Group-based relative advantage

Instead of comparing rewards to a learned value function, GRPO computes relative performance by comparing each output against the mean and standard deviation of its peer group.

The open-source standard

Originally introduced by DeepSeek, GRPO has become the dominant algorithm for post-training on verifiable tasks due to its simplicity and reduced memory requirements.

Bottom Line

For verifiable reasoning tasks like mathematics and coding, use GRPO instead of PPO to eliminate complex value networks and enable scalable reinforcement learning through group-based relative advantage estimation.

More from Stanford Online

View all
Stanford CS25: Transformers United V6 I Advancing Science and Medicine with Collaborative AI Agents
1:06:33
Stanford Online Stanford Online

Stanford CS25: Transformers United V6 I Advancing Science and Medicine with Collaborative AI Agents

Google DeepMind researcher Vivek Natarajan discusses the development of Co-Scientist, an AI system designed to act as a collaborative partner for scientific discovery by moving beyond fast System 1 thinking to rigorous System 2 reasoning, emphasizing that true scientific AI requires the generality of human cognition rather than narrow specialization.

about 10 hours ago · 7 points