NLP Lineage

31 May 2023 · Scale and data

Process supervision

Let's Verify Step by Step

H. Lightman, V. Kosaraju, Y. Burda, et al. (10 authors)

Date
31 May 2023
What the date marks
arXiv submission
Kind
Research paper
Organisation
OpenAI
Access
MIT (dataset)
Other dates
ICLR 2024; PRM800K released
Threads
In the guided tour
Chapter 12. Thinking longer
Learning phases
2. How LLMs are trained

A milestone: 1 later milestone descends directly from it, and its descendants reach 5 of the 10 threads. How levels are decided

Read the primary source (arxiv.org)

Full summary in progress. Until then, this page shows the milestone's record from the research data.

What it changed

Before it

Reasoning was rewarded by whether the final answer was right.

The problem

A model can reach the right answer through faulty reasoning, and outcome rewards teach it to keep doing that.

What it introduced

Reward each reasoning step individually, with 800,000 human step-level labels released.

Why it matters

Process supervision beat outcome supervision decisively, and is a visible ancestor of the RL-trained reasoning models.

What modern models inherit

Judging the reasoning, not only the answer.

What came next

Reward models over reasoning steps, then o1 and DeepSeek-R1.

Before and after

The milestones it descends from, and the milestones that descend from it. Work developed alongside it is not counted. Weak links are left out unless you include them.

Before: where it came from

Process supervision has 19 documented ancestors across 6 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 10 turning points among them.

Process supervision has 19 documented ancestors across 6 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

Process supervision has 29 recorded ancestors, including indirect ones, across 6 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 10 turning points among them.

Process supervision has 29 recorded ancestors, including indirect ones, across 6 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2023Process supervisionThis milestone.
  2. 2022Chain of thought Turning pointProcess supervision built directly on thisChapter 12. Thinking longer
  3. 2020GPT-3 Turning pointChain of thought was influenced by thisChapter 5. Scale becomes a science
  4. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  5. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  6. 2019GPT-2 Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  7. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  8. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  9. 2017Transformer Turning pointGPT-2 built directly on this
    Sparse Transformer built directly on this
    Scaling laws built directly on this    Chapter 3. Attention, then only attention
  10. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  11. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  12. 2015LM pretraining Turning pointGPT-1 was influenced by thisChapter 2. Words become vectors
  13. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  14. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  15. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  16. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  17. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  18. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  19. 1997LSTM Turning pointSeq2seq built directly on thisChapter 1. Language as probability
  20. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2023Process supervisionThis milestone.
  2. 2022STaRProcess supervision was influenced by this (indirect influence)Chapter 12. Thinking longer
  3. 2022Chain of thought Turning pointProcess supervision built directly on thisChapter 12. Thinking longer
  4. 2020GPT-3 Turning pointChain of thought was influenced by thisChapter 5. Scale becomes a science
  5. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  6. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  7. 2019GPT-2 Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  8. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  9. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  10. 2018GPT-1GPT-2 built directly on thisChapter 4. Pretrain once, use everywhere
  11. 2018ELMoGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  12. 2018ULMFiTGPT-1 was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  13. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  14. 2017Transformer Turning pointGPT-2 built directly on this
    Sparse Transformer built directly on this
    Scaling laws built directly on this    Chapter 3. Attention, then only attention
  15. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  16. 2016Layer normGPT-2 built directly on thisChapter 3. Attention, then only attention
  17. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  18. 2015LM pretraining Turning pointGPT-1 was influenced by thisChapter 2. Words become vectors
  19. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  20. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  21. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  22. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  23. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  24. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  25. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  26. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  27. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  28. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  29. 1997LSTM Turning pointSeq2seq built directly on this
    ELMo built directly on this    Chapter 1. Language as probability
  30. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

Process supervision has 8 documented descendants across 2 chapters, reaching forward to Qwen3.8 in 2026. Showing its direct descendants.

Process supervision has 8 documented descendants across 2 chapters, reaching forward to Qwen3.8 in 2026. Showing all of them.

Process supervision has 42 recorded descendants, including indirect ones, across 5 chapters, reaching forward to Claude Opus 5.5 in 2026. Showing its direct descendants and the 1 turning point among them.

Process supervision has 42 recorded descendants, including indirect ones, across 5 chapters, reaching forward to Claude Opus 5.5 in 2026. Showing all of them.

  1. 2023Process supervisionThis milestone.
  2. 2024Test-time computewas influenced by Process supervisionChapter 12. Thinking longer
  3. 2025DeepSeek-R1was influenced by Test-time computeChapter 12. Thinking longer
  4. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  5. 2025IMO goldwas influenced by Test-time computeChapter 12. Thinking longer
  6. 2025IMO-ProofBenchwas influenced by IMO goldChapter 12. Thinking longer
  7. 2026Gemini 3 Deep Thinkwas influenced by IMO goldChapter 12. Thinking longer
  8. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  9. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out
  1. 2023Process supervisionThis milestone.
  2. 2024GRPOwas influenced by Process supervision (indirect influence)Chapter 12. Thinking longer
  3. 2024Test-time computewas influenced by Process supervisionChapter 12. Thinking longer
  4. 2024o1 Turning pointwas influenced by Process supervision (indirect influence)Chapter 12. Thinking longer
  5. 2024o3 announcedbuilt directly on o1
    was influenced by Test-time compute (indirect influence)    Chapter 12. Thinking longer
  6. 2024DeepSeek-V3built directly on GRPOChapter 8. The weights get out
  7. 2025DeepSeek-R1built directly on GRPO
    was influenced by o1
    was influenced by Test-time compute    Chapter 12. Thinking longer
  8. 2025Deep researchwas influenced by o3 releasedChapter 13. Models that act
  9. 2025Claude 3.7was influenced by o1Chapter 12. Thinking longer
  10. 2025Claude Codewas influenced by Claude 3.7Chapter 13. Models that act
  11. 2025Gemini 2.5was influenced by o1Chapter 12. Thinking longer
  12. 2025o3 releasedbuilt directly on o1Chapter 12. Thinking longer
  13. 2025Qwen3was influenced by DeepSeek-R1Chapter 8. The weights get out
  14. 2025Claude 4built directly on Claude 3.7Chapter 13. Models that act
  15. 2025Kimi K2was influenced by GRPOChapter 8. The weights get out
  16. 2025IMO goldwas influenced by Test-time computeChapter 12. Thinking longer
  17. 2025gpt-ossbuilt directly on o3 released
    was influenced by DeepSeek-R1 (indirect influence)    Chapter 8. The weights get out
  18. 2025GPT-5built directly on o3 releasedChapter 14. Where the threads meet now
  19. 2025GLM-4.5was influenced by DeepSeek-R1 (indirect influence)Chapter 8. The weights get out
  20. 2025DeepSeek-V3.2built directly on DeepSeek-V3Chapter 10. Reading more than you remember
  21. 2025Claude Sonnet 4.5built directly on Claude 4Chapter 13. Models that act
  22. 2025IMO-ProofBenchwas influenced by IMO goldChapter 12. Thinking longer
  23. 2025GPT-5.1built directly on GPT-5Chapter 14. Where the threads meet now
  24. 2025Gemini 3built directly on Gemini 2.5Chapter 14. Where the threads meet now
  25. 2025Claude Opus 4.5built directly on Claude Sonnet 4.5Chapter 13. Models that act
  26. 2025GPT-5.2built directly on GPT-5.1Chapter 14. Where the threads meet now
  27. 2026Claude Opus 4.6built directly on Claude Opus 4.5Chapter 14. Where the threads meet now
  28. 2026GPT-5.3-Codexbuilt directly on GPT-5.2Chapter 13. Models that act
  29. 2026Gemini 3 Deep Thinkwas influenced by IMO goldChapter 12. Thinking longer
  30. 2026Qwen3.5built directly on Qwen3Chapter 8. The weights get out
  31. 2026Gemini 3.1built directly on Gemini 3Chapter 14. Where the threads meet now
  32. 2026GPT-5.4built directly on GPT-5.3-CodexChapter 14. Where the threads meet now
  33. 2026Claude Opus 4.7built directly on Claude Opus 4.6Chapter 14. Where the threads meet now
  34. 2026GPT-5.5built directly on GPT-5.4Chapter 14. Where the threads meet now
  35. 2026DeepSeek-V4was influenced by DeepSeek-V3Chapter 8. The weights get out
  36. 2026Claude Opus 4.8built directly on Claude Opus 4.7Chapter 14. Where the threads meet now
  37. 2026Claude Sonnet 5succeeded Claude Opus 4.8Chapter 14. Where the threads meet now
  38. 2026GLM-5.2built directly on GLM-4.5Chapter 8. The weights get out
  39. 2026GPT-5.6 Solbuilt directly on GPT-5.5Chapter 14. Where the threads meet now
  40. 2026Kimi K3built directly on Kimi K2Chapter 8. The weights get out
  41. 2026Claude Opus 5succeeded Claude Opus 4.8Chapter 14. Where the threads meet now
  42. 2026Qwen3.8built directly on Qwen3.5Chapter 8. The weights get out
  43. 2026Claude Opus 5.5built directly on Claude Opus 5Chapter 14. Where the threads meet now

Direct links and their sources

Where it came from

Built directly on

Influenced by

What it led to

Influenced