← All notes

Summary

Deep reinforcement learning from human preferences (2017) introduced the method of training a reward model from human comparisons, then optimizing a policy against that reward. It was first demonstrated on simulated control tasks, years before it became the standard way to align language-model assistants.

What it introduced: preference-based reward learning plus reinforcement learning on the resulting reward model, so that humans can steer a system by comparing behaviors instead of writing a scalar reward by hand.

What it led to: InstructGPT and the later RLHF stack for chat models, and then alternatives such as direct preference optimization. Live notes: 2204-05862-helpful-harmless-rlhf and 2305-18290-direct-preference-optimization. See also the PPO pointer in this batch, which became the usual policy-gradient algorithm in that stack.

Where to read it: https://arxiv.org/abs/1706.03741 (v4, last revised 17 Feb 2023).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

@misc{iil_note_hsmww27j,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Deep reinforcement learning from human preferences}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/1706-03741-rl-from-human-preferences/},
  note = {Permanent ID iil:note:hsmww27j. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/1706-03741-rl-from-human-preferences.md.