mlsanityData to 14 Oct 2026

Transcendence: Generative Models Can Outperform The Experts That Train Them

arXiv 2406.11741 ↗First posted 17 Jun 2024cs.LG cs.AIcited by 20

Not on the map. Why?

Abstract (arXiv)

Generative models are trained with the simple objective of imitating the conditional probability distribution induced by the data they are trained on. Therefore, when trained on data generated by humans, we may not expect the artificial model to outperform the humans on their original objectives. In this work, we study the phenomenon of transcendence: when a generative model achieves capabilities that surpass the abilities of the experts generating its data. We demonstrate transcendence by training an autoregressive transformer to play chess from game transcripts, and show that the trained model can sometimes achieve better performance than all players in the dataset. We theoretically prove that transcendence can be enabled by low-temperature sampling, and rigorously assess this claim experimentally. Finally, we discuss other sources of transcendence, laying the groundwork for future investigation of this phenomenon in a broader setting.

Key facts. Cited by 20 papers in the map (data through 14 Oct 2026). Builds on 19 of its 40 references that are in the map. First posted on arXiv 17 Jun 2024.

Builds on

19 of 40 references.

References · 19

  1. 2017Attention Is All You Needcited by 60,361
  2. 2016Communication-Efficient Learning of Deep Networks from Decentralized Datacited by 7,669
  3. 2022OPT: Open Pre-trained Transformer Language Modelscited by 3,512
  4. 2020Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problemscited by 1,920
  5. 2021Decision Transformer: Reinforcement Learning via Sequence Modelingcited by 1,546
  6. 2019Self-training with Noisy Student improves ImageNet classificationcited by 1,311
  7. 2017Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithmcited by 734
  8. 2020Ensemble Distillation for Robust Model Fusion in Federated Learningcited by 682
  9. 2021Offline Reinforcement Learning as One Big Sequence Modeling Problemcited by 644
  10. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 467
  11. 2016Safe and Efficient Off-Policy Reinforcement Learningcited by 264
  12. 2020Aligning Superhuman AI with Human Behavior: Chess as a Model Systemcited by 79
7 more
  1. 2023ChessGPT: Bridging Policy Learning and Language Modelingcited by 75
  2. 2021Chess as a Testbed for Language Model State Trackingcited by 72
  3. 2024Emergent World Models and Latent Variable Estimation in Chess-Playing Language Modelscited by 62
  4. 2020The Chess Transformer: Mastering Play using Generative Language Modelscited by 37
  5. 2020Stable Policy Optimization via Off-Policy Divergence Regularizationcited by 16
  6. 2022Offline Reinforcement Learning with Closed-Form Policy Improvement Operatorscited by 14
  7. 2023Leveraging Ensemble Diversity for Robust Self-Training in the Presence of Sample Selection Biascited by 7

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2024: 6; 2025: 9; 2026: 5.

20 papers.

Sort:

By area:

  1. 2024Generative Verifiers: Reward Modeling as Next-Token Predictioncited by 480
  2. 2024AI Safety in Generative AI Large Language Models: A Surveycited by 33
  3. 2025Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challengescited by 31
  4. 2024Provable Weak-to-Strong Generalization via Benign Overfittingcited by 16
  5. 2024Human-aligned Chess with a Bit of Searchcited by 11
  6. 2026Understanding Reasoning from Pretraining to Post-Trainingcited by 5
  7. 2025EnsemW2S: Enhancing Weak-to-Strong Generalization with Large Language Model Ensemblescited by 4
  8. 2025Weak-to-Strong Generalization under Distribution Shiftscited by 4
  9. 2025Data-Efficient Multi-Agent Spatial Planning with LLMscited by 3
  10. 2025A Taxonomy of Transcendencecited by 2
  11. 2025Out-of-distribution Tests Reveal Compositionality in Chess Transformerscited by 2
  12. 2026Weak-to-Strong Generalization is Nearly Inevitable (in Linear Models)cited by 2
8 more
  1. 2024EnsemW2S: Enhancing Weak-to-Strong Generalization with Large Language Model Ensemblescited by 1
  2. 2025Limitations of refinement methods for weak to strong generalizationcited by 1
  3. 2025Understanding temperature tuning in energy-based modelscited by 1
  4. 2026Mixture of Masters: Sparse Chess Language Models with Player Routingcited by 1
  5. 2024SimpleStrat: Diversifying Language Model Generation with Stratificationcited by 0
  6. 2025Iterative Deployment Improves Planning Skills in LLMscited by 0
  7. 2026UniMaia: Steering Chess Policies with Language for Human-like Playcited by 0
  8. 2026Language Models that Play Chess and Explain Their Movescited by 0