NLP Lineage

November 2025 · Reasoning

IMO-ProofBench

Towards Robust Mathematical Reasoning (IMO-ProofBench)

Luong et al., Google DeepMind

Date
November 2025
What the date marks
arXiv submission
Kind
Research paper
Organisation
Google DeepMind
Access
Not applicable
Threads
In the guided tour
Chapter 12. Thinking longer
Learning phases
2. How LLMs are trained

Too recent for its influence to be judged: no later milestone in the data descends from it yet. How levels are decided

The key result has been peer reviewed.

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

Gold-medal olympiad results were taken as evidence of general mathematical reasoning.

The problem

Competition problems and their solutions circulate widely, so scores may reflect familiarity rather than reasoning.

What it introduced

A proof benchmark with novel problems, on which most models fell far below their headline results.

Why it matters

The clearest independent check on the year's most-cited reasoning claims, and a caution for reading any vendor benchmark.

What modern models inherit

Scepticism as a standard part of reading capability claims.

What came next

More emphasis on contamination-resistant evaluation.

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

IMO-ProofBench has 41 documented ancestors across 9 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 14 turning points among them.

IMO-ProofBench has 41 documented ancestors across 9 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

IMO-ProofBench has 74 recorded ancestors, including indirect ones, across 12 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 20 turning points among them.

IMO-ProofBench has 74 recorded ancestors, including indirect ones, across 12 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2025IMO-ProofBenchThis milestone.
  2. 2025IMO goldIMO-ProofBench was influenced by thisChapter 12. Thinking longer
  3. 2025Gemini 2.5IMO gold built directly on thisChapter 12. Thinking longer
  4. 2024Gemini 2.0Gemini 2.5 built directly on thisChapter 14. Where the threads meet now
  5. 2024o1 Turning pointGemini 2.5 was influenced by thisChapter 12. Thinking longer
  6. 2024Test-time computeIMO gold was influenced by thisChapter 12. Thinking longer
  7. 2024Gemini 1.5Gemini 2.5 was influenced by thisChapter 10. Reading more than you remember
  8. 2023Gemini 1.0Gemini 1.5 built directly on thisChapter 11. Seeing and hearing
  9. 2023Process supervisionTest-time compute was influenced by thisChapter 12. Thinking longer
  10. 2022PaLMGemini 1.0 succeeded thisChapter 5. Scale becomes a science
  11. 2022Self-consistencyTest-time compute built directly on thisChapter 12. Thinking longer
  12. 2022Chain of thought Turning pointo1 was influenced by this
    Self-consistency built directly on this
    Process supervision built directly on this    Chapter 12. Thinking longer
  13. 2020Vision TransformerGemini 1.0 was influenced by thisChapter 11. Seeing and hearing
  14. 2020GShard Turning pointGemini 1.5 was influenced by thisChapter 5. Scale becomes a science
  15. 2020GPT-3 Turning pointChain of thought was influenced by thisChapter 5. Scale becomes a science
  16. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  17. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  18. 2019GPT-2 Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  19. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  20. 2018BERT Turning pointVision Transformer was influenced by thisChapter 4. Pretrain once, use everywhere
  21. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  22. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  23. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  24. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  25. 2017Transformer Turning pointVision Transformer built directly on thisChapter 3. Attention, then only attention
  26. 2017Sparse MoE Turning pointGShard built directly on thisChapter 3. Attention, then only attention
  27. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  28. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  29. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  30. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  31. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  32. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  33. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  34. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  35. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  36. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  37. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  38. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  39. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  40. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  41. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  42. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2025IMO-ProofBenchThis milestone.
  2. 2025IMO goldIMO-ProofBench was influenced by thisChapter 12. Thinking longer
  3. 2025Gemini 2.5IMO gold built directly on thisChapter 12. Thinking longer
  4. 2024Gemini 2.0Gemini 2.5 built directly on thisChapter 14. Where the threads meet now
  5. 2024Computer useGemini 2.0 was influenced by this (indirect influence)Chapter 13. Models that act
  6. 2024o1 Turning pointGemini 2.5 was influenced by thisChapter 12. Thinking longer
  7. 2024Test-time computeIMO gold was influenced by thisChapter 12. Thinking longer
  8. 2024Claude 3.5 SonnetComputer use built directly on thisChapter 14. Where the threads meet now
  9. 2024Claude 3Claude 3.5 Sonnet built directly on thisChapter 14. Where the threads meet now
  10. 2024Gemini 1.5Gemini 2.5 was influenced by thisChapter 10. Reading more than you remember
  11. 2023Gemini 1.0Gemini 1.5 built directly on thisChapter 11. Seeing and hearing
  12. 2023GPT-4VComputer use was influenced by thisChapter 11. Seeing and hearing
  13. 2023Lost in the middleGemini 1.5 was influenced by this (indirect influence)Chapter 10. Reading more than you remember
  14. 2023Function callingComputer use was influenced by this (indirect influence)Chapter 13. Models that act
  15. 2023Process supervisionTest-time compute was influenced by thisChapter 12. Thinking longer
  16. 2023Tree of thoughtsTest-time compute was influenced by this (indirect influence)Chapter 12. Thinking longer
  17. 2023100K contextLost in the middle was influenced by this (indirect influence)Chapter 10. Reading more than you remember
  18. 2023GPT-4 Turning pointGPT-4V built directly on thisChapter 7. The assistant goes public
  19. 2023Claude100K context built directly on thisChapter 7. The assistant goes public
  20. 2023ToolformerFunction calling was influenced by this (indirect influence)Chapter 13. Models that act
  21. 2022Constitutional AIClaude built directly on this
    Claude 3 built directly on this    Chapter 6. Teaching models what people want
  22. 2022GPT-3.5 Turning pointGPT-4 succeeded thisChapter 7. The assistant goes public
  23. 2022ReActFunction calling was influenced by this (indirect influence)Chapter 13. Models that act
  24. 2022HH-RLHFClaude was influenced by thisChapter 6. Teaching models what people want
  25. 2022PaLMGemini 1.0 succeeded thisChapter 5. Scale becomes a science
  26. 2022STaRTest-time compute was influenced by this (indirect influence)Chapter 12. Thinking longer
  27. 2022Self-consistencyTest-time compute built directly on thisChapter 12. Thinking longer
  28. 2022InstructGPT Turning pointGPT-4 built directly on thisChapter 6. Teaching models what people want
  29. 2022Chain of thought Turning pointo1 was influenced by this
    Self-consistency built directly on this
    STaR was influenced by this    Chapter 12. Thinking longer
  30. 2021WebGPTReAct was influenced by this
    Toolformer was influenced by this    Chapter 13. Models that act
  31. 2021HHH assistantHH-RLHF built directly on thisChapter 6. Teaching models what people want
  32. 2021FLANInstructGPT was influenced by thisChapter 6. Teaching models what people want
  33. 2021CLIPGPT-4V was influenced by this (indirect influence)Chapter 11. Seeing and hearing
  34. 2020Vision TransformerGemini 1.0 was influenced by thisChapter 11. Seeing and hearing
  35. 2020RLHF for summarisation Turning pointInstructGPT built directly on this
    WebGPT built directly on this
    HH-RLHF built directly on this    Chapter 6. Teaching models what people want
  36. 2020GShard Turning pointGemini 1.5 was influenced by thisChapter 5. Scale becomes a science
  37. 2020GPT-3 Turning pointChain of thought was influenced by thisChapter 5. Scale becomes a science
  38. 2020RAGLost in the middle was influenced by thisChapter 10. Reading more than you remember
  39. 2020Dense retrievalRAG built directly on thisChapter 10. Reading more than you remember
  40. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  41. 2019BARTRAG built directly on thisChapter 4. Pretrain once, use everywhere
  42. 2019T5 Turning pointFLAN was influenced by thisChapter 4. Pretrain once, use everywhere
  43. 2019C4 Turning pointT5 was influenced by thisChapter 4. Pretrain once, use everywhere
  44. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  45. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  46. 2019GPT-2 Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  47. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  48. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  49. 2018BERT Turning pointVision Transformer was influenced by thisChapter 4. Pretrain once, use everywhere
  50. 2018SentencePieceT5 built directly on thisChapter 4. Pretrain once, use everywhere
  51. 2018GPT-1BERT was influenced by this
    GPT-2 built directly on this    Chapter 4. Pretrain once, use everywhere
  52. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  53. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  54. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  55. 2017RLHF origino1 was influenced by this (indirect influence)Chapter 6. Teaching models what people want
  56. 2017Transformer Turning pointVision Transformer built directly on thisChapter 3. Attention, then only attention
  57. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  58. 2017FAISSRAG was influenced by thisChapter 10. Reading more than you remember
  59. 2017Sparse MoE Turning pointGShard built directly on thisChapter 3. Attention, then only attention
  60. 2016Layer normTransformer built directly on this
    GPT-2 built directly on this    Chapter 3. Attention, then only attention
  61. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  62. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  63. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  64. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  65. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  66. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  67. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  68. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  69. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  70. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  71. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  72. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  73. 1997LSTM Turning pointSparse MoE built directly on thisChapter 1. Language as probability
  74. 1991Mixture of expertsSparse MoE built directly on thisChapter 1. Language as probability
  75. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

No documented descendants are recorded for IMO-ProofBench without weak links.

No documented descendants are recorded for IMO-ProofBench without weak links.

No recorded descendants are recorded for IMO-ProofBench.

No recorded descendants are recorded for IMO-ProofBench.

Direct links and their sources

Where it came from

Influenced by

What it led to

None recorded in the sources used.