AI Systems · advanced
RL Alignment (GRPO & Policy Gradient)
Policy gradient, advantages, GRPO/CISPO, on- vs off-policy RLHF.
Mental model
Post-training RL treats the LM as a policy: sample completions, score them with a reward model, then nudge token probabilities toward high-reward outputs. GRPO compares completions within a group (relative advantages) and reuses rollouts with importance sampling + clipping so training stays stable off-policy.
How to study RL Alignment (GRPO & Policy Gradient)
Begin by restating the mental model in your own words, then connect it to a concrete system you have built or operated. Name the mechanism, the constraint it addresses, and the trade-off it introduces. Use Lightweight GRPO & RL intro (Murali), Policy Gradient for LMs — RL101 notes (Hamish Ivison), Stanford CS336 — L15 Alignment: SFT & RLHF (Spring 2025) to check details, but close the source before writing your explanation. Retrieval is the learning step; rereading is only preparation.
Next, compare RL Alignment (GRPO & Policy Gradient) with Model Evaluation, Training & Debugging. Ask what changes in correctness, latency, resource use, operability, and failure recovery. Complete GRPO group-relative advantage and preserve the command, input, output, and one failed attempt as evidence. Finish by explaining the idea without jargon to someone who has not studied the track.
Proof of understanding
- Explain the mechanism from first principles and identify the state it reads or changes.
- Give one situation where the concept is the right choice and one where it is not.
- Predict a realistic failure mode before running the drill, then compare the prediction with evidence.
- Connect the result to a roadmap or build artifact instead of treating the concept as isolated trivia.
Learn from primary sources
Practice and explain it back
GRPO group-relative advantage
Four completions with rewards [1.0, 0.5, 0.8, 0.3]. Mean μ=0.65, population std σ≈0.269. What is the normalized advantage for the top completion (1.0)?
Expected evidence: (1.0 − 0.65) / 0.269 ≈ 1.30 — GRPO upweights this rollout relative to the group.
Open the interactive drill →Review prompts
- In GRPO-style RL alignment, why are advantages computed relative to a group of completions for the same prompt, and why is importance sampling needed?
Build evidence
Synthesize: AI Models & Training
Move from transformer foundations through pre-training, fine-tuning, post-training, compression, and evaluation. Produce one working system, benchmark, or evidence-backed design that integrates the path.
- Implements or precisely models the core mechanisms from all three milestones
- Includes at least one injected failure or adversarial case and demonstrates recovery
- Reports quality, latency, resource, reliability, or usability measurements relevant to the domain
- Ships a concise architecture note explaining decisions, trade-offs, and remaining risks