Glossary · Term

RLHF

← all terms

Definition

Plain language

Training a model with human feedback so it learns to answer the way humans prefer.

As stated in the literature

Reinforcement Learning from Human Feedback, the post-training pipeline that fits a reward model to human preferences and then optimizes a policy against it.

Why it matters: It's the post-training step that turned raw next-token predictors into usable chat assistants, and it remains the dominant recipe for aligning model outputs with human taste.

For example, human annotators rank pairs of chatbot responses, a reward model learns from those rankings, and the chatbot is then nudged to produce answers the reward model scores highly.

Heard on the show

“Frontier models are getting updated continuously — RLHF passes, post-training, knowledge injection.”
Episode 092 — When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks

Mentioned in 13 episodes

  1. 092
    When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks
  2. 082
    Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
  3. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  4. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
  5. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
  6. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy
  7. 058
    Why Upgrading Your AI Auditor to a Smarter Model Can Make Your System Less Safe
  8. 044
    How One Sentence and a Forged History Flip the Most Aligned Models
  9. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
  10. 020
    The Compliance Gap: Why AI Says Yes and Does No
  11. 018
    Language Models Compute the Rational Move, Then Override It
  12. 008
    Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps
  13. 004
    The Sycophancy Circuit That Survives Alignment Training

Related concepts

Related terms