Safety & Alignment · since 2022

RLHF & Alignment

Reinforcement learning from human feedback — the training technique that made raw language models follow instructions, refuse harm, and become shippable. Conceived in 2017 (Christiano et al.) but it is the 2022 applied moment (InstructGPT) that turned it into the alignment substrate of every deployed assistant.

5

events traced

5

source records

30 Jun 2026

first signal

11 Jul 2026

last activity

Who drove it

OpenAI34%
Anthropic30%
Google DeepMind20%
academia16%

Key movements

2017origin

conceived (Christiano)

2022

applied at scale (InstructGPT)

2024

evolved into RL-for-reasoning (RLVR)

The story, event by event

Every point below is traced to a real source — nothing on this page is invented.

  1. 12 Jun 2017impact 72

    Deep RL from human preferences

    value: origin; metric: method

  2. 4 Mar 2022impact 90

    InstructGPT — RLHF at scale

    value: InstructGPT; metric: application

  3. 29 May 2023impact 66

    DPO simplifies alignment

    value: DPO; metric: method

  4. 22 Nov 2024impact 80

    RLVR — RL from verifiable rewards

    value: verifiable; metric: reward

  5. 19 Dec 2025impact 60

    Release of Gemma Scope 2

    data storage: 110 Petabytes of data; total parameters: Over 1 trillion parameters; model sizes supported: All sizes of Gemma 3 models, from 270M to 27B parameters

Lineage

Descended from

Led to

This is one node. The map holds the whole field.

Watch RLHF & Alignment — and everything it connects to — grow as real news threads onto the map every day.