Continuous FunctionStart learning

Foundation Lab

RLAIF: AI Feedback

Removes human labeling bottleneck

Concept 82 of 100Scaling & AlignmentPhase 7
Inspect equationAll Foundations
#82RLAIFScaling & Alignment
key equation
r(y)=LLM(y∣criteria)r(y) = \text{LLM}(y | \text{criteria})
Reading map and next steps

Selected Foundation Object

Keep the equation fixed; move through the evidence.

Concept 82 of 100RLAIFScaling & Alignment / Phase 7: Alignment & RLHF
Current question

AI feedback is consistent and follows complex rubrics

r(y)=LLM(y∣criteria)r(y) = \text{LLM}(y | \text{criteria})
PredictionCommit before tracing the equation.

Ask what should change under a concrete input, then trace that expectation through the equation.

EvidenceCompare the equation and source.

Use the key equation and canonical papers as the available witnesses, without implying that a runnable panel exists.

InvariantName what survives notation changes.

The useful learning product is the reusable mechanism you can carry into another model, paper, or engineering tradeoff.

Next moveContinue through the atlas.

Use prerequisites, dependents, and semantic links to repair the next gap without leaving the object behind.

Inspect equationBrowse atlasNext foundation

Why It Matters for Modern Models

  • Removes human labeling bottleneck
  • Powers Constitutional AI and production alignment
  • Quality approaching human at 10-100× lower cost

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • AI feedback is consistent and follows complex rubrics
  • Judge can be same model or stronger one
  • Works best with clear criteria; fails on subjective judgments

Visualization Status

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and cited sources. Come back here for the full walkthrough.

Key Equation
r(y)=LLM(y∣criteria)r(y) = \text{LLM}(y | \text{criteria})

Replace human with AI preferences:

RLHF: r(y)=Human(y)r(y) = \text{Human}(y) → expensive
RLAIF: r(y)=LLM(y∣criteria)r(y) = \text{LLM}(y | \text{criteria}) → scalable

AI judge: P(y1≻y2)=LLM(x,y1,y2,rubric)P(y_1 \succ y_2) = \text{LLM}(x, y_1, y_2, \text{rubric})

Correlates ~0.85 with human for many tasks.

Canonical Papers

RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback

Lee et al.2023arXiv
Read paper →

Connections

Prerequisites

⚖RLHF📜Constitutional

Next Moves

Choose the next question to carry this object forward.