{
  "dataset": "glossary",
  "record": {
    "id": "rlhf",
    "term": "RLHF",
    "category": "knowledge-memory",
    "short_def": "Reinforcement Learning from Human Feedback — training a model to prefer responses humans rate higher, aligning it with human intent.",
    "long_def": "RLHF fine-tunes a model using human preference data: people rank model outputs, a reward model learns those preferences, and reinforcement learning nudges the model toward higher-rated behavior. It is a core step in turning a raw pretrained model into a helpful, instruction-following assistant.",
    "see_also": [
      "fine-tuning",
      "grounding",
      "transformer"
    ],
    "etymology_origin": {
      "value": null,
      "verify_status": "verify-against-primary-at-build",
      "source_hint": "https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback — popularized by InstructGPT (OpenAI, 2022)"
    },
    "related_to": [
      "fine-tuning",
      "grounding",
      "transformer"
    ],
    "contrast_with": "Unlike supervised fine-tuning on fixed target answers, RLHF optimizes against a learned reward model of human preference, shaping style and safety rather than copying exact outputs.",
    "example": "RLHF is a large part of why a modern assistant follows instructions and declines unsafe requests instead of merely completing text.",
    "source": "https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback",
    "status": "active",
    "why_it_matters": "RLHF shapes how an agent behaves once deployed on the agentic web — how it follows instructions, respects constraints and handles ambiguity.",
    "sameAs": [
      "https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback"
    ],
    "bridge_entity": "fine-tuning",
    "last_verified": "2026-07-06",
    "md_twin": "/glossary/rlhf.md"
  }
}