---
id: LIT-FOUNDATIONAL-001
title: "Foundational Papers Index"
alternative_names:
  - "Annotated Bibliography of Foundational AI Literature"
  - "Core Reading List"
category: "People & Organizations → Literature"
volume: X
chapter: 58
entry_type: "Statistical Entry"
level: 6
status: "Draft"
last_reviewed: 2026-10-15
version: "0.1.0"
prerequisites: []
related_entries:
  - "LIT-SOURCES-001"
  - "TECH-ARCH-001"
  - "TECH-LLM-001"
  - "TECH-DL-001"
  - "TECH-RL-001"
  - "SAFE-ALIGN-001"
  - "HIST-005"
  - "LIB-001"
---

# Foundational Papers Index

## 1. Concise Definition
This is an annotated bibliography of 47 works that the encyclopedia treats as foundational for modern machine learning, large language models, reinforcement learning and AI safety, grouped by topic. Every bibliographic record was checked against Crossref or DataCite (which carries arXiv records) on 2026-10-15; annotations state only what the title or abstract of the work supports.

## 2. Overview & Significance
The index serves readers who want to move from encyclopedia articles to the literature. It is selective, not exhaustive, and inclusion is neither an endorsement of a work's claims nor a ranking of importance. Pre-1986 primary documents and policy instruments are in [LIT-SOURCES-001]. Citation counts are not used as a criterion because they were not collected.

Conventions:
- Preprints on arXiv are cited as preprints, because the arXiv record is the one verified; many have later conference or journal versions that were not checked and are therefore not listed. By the Source Policy, an arXiv preprint is Tier 3 evidence unless a peer-reviewed version is cited.
- Where a work has more than eight authors, the first three are given followed by "et al." and the author count from the verified record.
- Laboratory-authored reports establish what the laboratory claims; they are not independent replications.

## 3. Historical Development
The literature falls into recognizable waves, which also organize the groups in section 4.

1. **Representation learning (1986-2016).** Papers on backpropagation, recurrent and convolutional networks, deep belief networks and large image datasets (Rumelhart et al., 1986; Hochreiter & Schmidhuber, 1997; LeCun et al., 1998; Hinton et al., 2006; Deng et al., 2009; Krizhevsky et al., 2017). For the narrative, see [HIST-005].
2. **Sequence models and attention (2014-2017).** Encoder-decoder and attention papers (Sutskever et al., 2014; Bahdanau et al., 2014), culminating in the Transformer (Vaswani et al., 2017).
3. **Pre-training and scaling (2018-2022).** Language-model pre-training, scaling studies and instruction tuning (Devlin et al., 2018; Brown et al., 2020; Kaplan et al., 2020; Hoffmann et al., 2022; Ouyang et al., 2022).
4. **Learning from human feedback and safety (2016-2025).** Safety problem statements and preference-based training methods (Amodei et al., 2016; Christiano et al., 2017; Bai et al., 2022; Casper et al., 2023).
5. **Reasoning-model reports (2025).** A report on reinforcement-learning-trained reasoning models, later published in *Nature* (Guo et al., 2025).

## 4. Technical / Conceptual Explanation: The Annotated List
Each entry gives author(s) and year, matching section 11, followed by a one- or two-sentence annotation. The annotation restates the work's own stated contribution; it is not an assessment of later impact.

### 4.1 Neural-network foundations and image recognition
- **Rumelhart, Hinton & Williams (1986).** Describes a learning procedure for multi-layer networks that adjusts weights by propagating errors backward. See [TECH-DL-001].
- **Hochreiter & Schmidhuber (1997).** Introduces Long Short-Term Memory, a recurrent architecture for learning over long time lags.
- **LeCun, Bottou, Bengio & Haffner (1998).** Applies gradient-based learning with convolutional networks to document recognition.
- **Hinton, Osindero & Teh (2006).** Presents a fast learning algorithm for deep belief nets. It is frequently cited in accounts of the renewed interest in deep networks.
- **Deng et al. (2009).** Introduces the ImageNet hierarchical image database.
- **Russakovsky et al. (2015).** Describes the ImageNet Large Scale Visual Recognition Challenge and its results through 2014.
- **Krizhevsky, Sutskever & Hinton (2017).** Reports a deep convolutional network for ImageNet classification. The cited record is the Communications of the ACM reprint; the original conference version was not checked.
- **LeCun, Bengio & Hinton (2015).** A review article on deep learning in *Nature*, useful as a broad entry point.
- **He, Zhang, Ren & Sun (2015).** Introduces residual learning for very deep image-recognition networks.
- **Kingma & Ba (2014).** Proposes Adam, a stochastic optimization method.
- **Goodfellow et al. (2014).** Introduces generative adversarial networks.

### 4.2 Sequence models, attention and the Transformer
- **Sutskever, Vinyals & Le (2014).** Sequence-to-sequence learning with neural networks.
- **Bahdanau, Cho & Bengio (2014).** Neural machine translation that jointly learns to align and translate, an attention mechanism.
- **Vaswani et al. (2017).** Introduces the Transformer, based on attention. See [TECH-ARCH-001].
- **Devlin, Chang, Lee & Toutanova (2018).** BERT, bidirectional pre-training for language understanding.

### 4.3 Large language models and scaling
- **Brown et al. (2020).** Reports GPT-3 and few-shot learning with language models.
- **Kaplan et al. (2020).** An empirical study stating that language-model loss scales as a power law with model size, dataset size and training compute. Scaling is discussed in [ECON-COMP-001].
- **Hoffmann et al. (2022).** Studies compute-optimal training and states that model size and training tokens should be scaled equally.
- **Wei, Wang et al. (2022), chain of thought.** Shows that prompting with intermediate reasoning steps improves performance on reasoning tasks for sufficiently large models.
- **Wei, Tay et al. (2022), emergent abilities.** Defines emergent abilities as those present in larger but not smaller models. Whether these are genuine discontinuities is **Contested**.
- **Schaeffer, Miranda & Koyejo (2023).** Argues that apparent emergence can arise from the choice of metric. Read alongside the preceding entry.
- **Ouyang et al. (2022).** Fine-tuning language models with human feedback to follow instructions (InstructGPT).
- **Christiano et al. (2017).** Reinforcement learning from human preferences over trajectory segments, the technique underlying later feedback-based training.
- **Bai et al. (2022).** Constitutional AI: training a harmless assistant using a list of principles and AI-generated feedback.
- **Bommasani et al. (2021).** A 114-author report on the opportunities and risks of foundation models.
- **OpenAI (2023).** The GPT-4 Technical Report. It states that architecture, size, hardware, training compute, dataset construction and training method are not disclosed.
- **Guo et al. (2025).** Reports reasoning capabilities incentivized through reinforcement learning (DeepSeek-R1); the *Nature* version is cited here, and the preprint is listed in [ORG-LABS-001].

### 4.4 Generative and multimodal models
- **Ho, Jain & Abbeel (2020).** Denoising diffusion probabilistic models. See [TECH-DIFF-001].
- **Radford et al. (2021).** Learns visual models from natural-language supervision (CLIP).
- **Dosovitskiy et al. (2020).** Vision Transformer: transformers for image recognition at scale.

### 4.5 Reinforcement learning, games and science
- **Sutton & Barto (2018).** The standard textbook on reinforcement learning. See [TECH-RL-001].
- **Mnih et al. (2015).** Human-level control on Atari games with deep reinforcement learning.
- **Silver et al. (2016).** Mastering Go with deep neural networks and tree search.
- **Silver et al. (2017).** A Go program trained without human game knowledge (AlphaGo Zero).
- **Campbell, Hoane & Hsu (2002).** Describes the Deep Blue chess system.
- **Jumper et al. (2021).** AlphaFold: protein structure prediction. See [APP-SCI-001].

### 4.6 Textbooks
- **Russell & Norvig (2020).** The standard general AI textbook (4th edition).
- **Goodfellow, Bengio & Courville (2016).** A textbook on deep learning.

### 4.7 Safety, risk and societal impact
- **Amodei et al. (2016).** Frames five practical research problems on accident risk in machine learning systems.
- **Hubinger et al. (2019).** Analyzes mesa-optimization, where a learned model is itself an optimizer. This is a **Hypothetical** risk analysis, not an observed failure.
- **Casper et al. (2023).** Surveys open problems and limitations of reinforcement learning from human feedback.
- **Bostrom (2014).** *Superintelligence*, a book-length argument on paths and risks. Its arguments are philosophical and **Contested**; see [ASI-001].
- **Russell (2019).** *Human Compatible*, a book on the control problem and a proposed approach to it. See [SAFE-ALIGN-001].
- **Bender, Gebru, McMillan-Major & Shmitchell (2021).** Discusses risks of large language models, the "stochastic parrots" paper.
- **Grace et al. (2024).** A survey of 2,778 researchers about the pace and impacts of AI progress. The survey reports opinion, not measured fact; see [AGI-TIME-001].
- **Bengio et al. (2025).** The International AI Safety Report, a multi-author assessment (96 authors in the DataCite record).
- **Butlin et al. (2023).** A 19-author preprint assessing AI systems against scientific theories of consciousness. See [PHIL-CONSC-001].

## 5. Types, Components & Classifications
| Group | Works | Typical venue types |
| :--- | :--- | :--- |
| Neural-network foundations | 11 | Journals, conference proceedings, preprints |
| Sequence models and Transformer | 4 | Preprints |
| LLMs and scaling | 12 | Preprints, one journal, one technical report |
| Generative and multimodal | 3 | Preprints |
| RL, games and science | 6 | *Nature*, books, journals |
| Textbooks | 2 | Books |
| Safety, risk and society | 9 | Preprints, books, conference proceedings |

The group counts sum to 47. Where the same work bears on two groups it is listed once.

## 6. Applications & Concrete Examples
Suggested reading paths (all items are in section 4):
- **Language models:** Vaswani et al. (2017), Devlin et al. (2018), Brown et al. (2020), Kaplan et al. (2020), Hoffmann et al. (2022), Ouyang et al. (2022).
- **Alignment methods:** Christiano et al. (2017), Ouyang et al. (2022), Bai et al. (2022), Casper et al. (2023).
- **Safety problem statements:** Amodei et al. (2016), Hubinger et al. (2019), Russell (2019), Bostrom (2014).
- **Game-playing systems:** Campbell et al. (2002), Mnih et al. (2015), Silver et al. (2016, 2017).

## 7. Advantages, Limitations & Unresolved Questions
**Strengths.** All records are verified; all items include identifiers.

**Limitations.**
- **Selection bias.** The list favors English-language work by large laboratories and universities in North America and Europe; important work elsewhere is under-represented.
- **Preprint status.** Many items are listed as preprints even though final versions may exist.
- **Annotation scope.** Annotations are drawn from titles and abstracts, not from a full re-reading of every paper.
- **Contested claims.** The inclusion of Wei, Tay et al. (2022) and Bostrom (2014) records influential claims whose interpretation is disputed; inclusion is not endorsement.
- **Missing areas.** Robotics, speech, optimization theory, interpretability, evaluation methodology and AI law are not covered and are planned for separate lists.

**Unresolved question.** What criteria should define "foundational" for a field changing this quickly; the current criterion (appearing in the encyclopedia's own technical articles or as a standard reference) is editorial.

## 8. Current State of Knowledge
As of 2026-10-15, the 47 entries are verified as existing with the cited authors, years, titles and identifiers. The list has not been reviewed by subject experts. For the fast-moving frontier (models, benchmark results, hardware), see the Living References `MODEL-001`, `FACT-001`, `TIME-001` and the Research Library index `LIB-001`.

## 9. Future Trajectory & Strategic Outlook
This is a living record. Update protocol: (1) additions need verified bibliographic records and a one-sentence annotation limited to the abstract; (2) works whose central claims are later refuted or retracted stay listed with a dated note; (3) the list is reviewed at least annually, next review 2027-10-15. No forecasts are made.

## 10. Related Entries & Knowledge Graph
- **Companion Literature**:
  - [[LIT-SOURCES-001] Primary Sources and Historical Records](./LIT-SOURCES-001_primary_sources_and_historical_records.md): Pre-1986 documents and policy instruments.
- **Related Methods & Technologies**:
  - [[TECH-ARCH-001] Transformer Architecture](../vol03_technical/TECH-ARCH-001_transformer_architecture.md): Context for sections 4.2 and 4.3.
  - [[TECH-LLM-001] Large Language Models](../vol03_technical/TECH-LLM-001_large_language_models.md): Context for section 4.3.
  - [[TECH-DL-001] Deep Learning](../vol03_technical/TECH-DL-001_deep_learning.md): Context for section 4.1.
  - [[TECH-RL-001] Reinforcement Learning](../vol03_technical/TECH-RL-001_reinforcement_learning.md): Context for section 4.5.
  - [[TECH-DIFF-001] Diffusion Models](../vol03_technical/TECH-DIFF-001_diffusion_models.md): Context for section 4.4.
- **Advanced Implications**:
  - [[SAFE-ALIGN-001] AI Alignment](../vol08_safety_governance/SAFE-ALIGN-001_ai_alignment.md): Context for section 4.7.
  - [[ASI-001] Artificial Superintelligence](../vol06_asi/ASI-001_artificial_superintelligence.md): Context for Bostrom (2014).
  - [[HIST-005] The Deep Learning Breakthrough](../vol02_history/HIST-005_deep_learning_breakthrough.md): Narrative history of section 4.1.
- **Living References**: `LIB-001` (Research Library index), `MODEL-001`.

## 11. References & Foundational Literature
Identifiers: DOI where available, otherwise arXiv ID or ISBN. arXiv records were checked on DataCite; DOI records on Crossref.

1. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). *Concrete Problems in AI Safety*. arXiv preprint. arXiv:1606.06565. [Five research problems on accident risk.]
2. Bahdanau, D., Cho, K., & Bengio, Y. (2014). *Neural Machine Translation by Jointly Learning to Align and Translate*. arXiv preprint. arXiv:1409.0473. [Attention for translation.]
3. Bai, Y., Kadavath, S., Kundu, S., et al. (2022). *Constitutional AI: Harmlessness from AI Feedback* (51 authors). arXiv preprint. arXiv:2212.08073. [Principle-guided training of a harmless assistant.]
4. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). *On the Dangers of Stochastic Parrots*. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623. https://doi.org/10.1145/3442188.3445922. [Risks of large language models.]
5. Bengio, Y., Mindermann, S., Privitera, D., et al. (2025). *International AI Safety Report* (96 authors). arXiv preprint. arXiv:2501.17805. [Consolidated multi-author assessment.]
6. Bommasani, R., Hudson, D. A., Adeli, E., et al. (2021). *On the Opportunities and Risks of Foundation Models* (114 authors). arXiv preprint. arXiv:2108.07258. [Report on foundation models.]
7. Bostrom, N. (2014). *Superintelligence: Paths, Dangers, Strategies*. Oxford University Press. ISBN 978-0199678112. [Book-length argument on superintelligence risks.]
8. Brown, T. B., Mann, B., Ryder, N., et al. (2020). *Language Models are Few-Shot Learners* (31 authors). arXiv preprint. arXiv:2005.14165. [GPT-3 and few-shot learning.]
9. Butlin, P., Long, R., Elmoznino, E., et al. (2023). *Consciousness in Artificial Intelligence: Insights from the Science of Consciousness* (19 authors). arXiv preprint. arXiv:2308.08708. [Assesses AI systems against theories of consciousness.]
10. Campbell, M., Hoane, A. J., & Hsu, F. (2002). *Deep Blue*. Artificial Intelligence, 134(1-2), 57-83. https://doi.org/10.1016/S0004-3702(01)00129-1. [The Deep Blue chess system.]
11. Casper, S., Davies, X., Shi, C., et al. (2023). *Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback* (32 authors). arXiv preprint. arXiv:2307.15217. [Survey of RLHF limitations.]
12. Christiano, P., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2017). *Deep reinforcement learning from human preferences*. arXiv preprint. arXiv:1706.03741. [Learning from human preferences.]
13. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). *ImageNet: A large-scale hierarchical image database*. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248-255. https://doi.org/10.1109/CVPR.2009.5206848. [The ImageNet dataset.]
14. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). *BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding*. arXiv preprint. arXiv:1810.04805. [Bidirectional pre-training.]
15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. (2020). *An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale* (12 authors). arXiv preprint. arXiv:2010.11929. [Vision Transformer.]
16. Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). *Generative Adversarial Networks*. arXiv preprint. arXiv:1406.2661. [GANs.]
17. Goodfellow, I., Bengio, Y., & Courville, A. (2016). *Deep Learning*. MIT Press. ISBN 978-0262035613. [Deep learning textbook.]
18. Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., Brauner, J., & Korzekwa, R. C. (2024). *Thousands of AI Authors on the Future of AI*. arXiv preprint. arXiv:2401.02843. [Survey of 2,778 researchers.]
19. Guo, D., Yang, D., Zhang, H., et al. (2025). *DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning* (194 authors in the Crossref record). Nature, 645(8081), 633-638. https://doi.org/10.1038/s41586-025-09422-z. [Reasoning via reinforcement learning.]
20. He, K., Zhang, X., Ren, S., & Sun, J. (2015). *Deep Residual Learning for Image Recognition*. arXiv preprint. arXiv:1512.03385. [Residual networks.]
21. Hinton, G. E., Osindero, S., & Teh, Y.-W. (2006). *A Fast Learning Algorithm for Deep Belief Nets*. Neural Computation, 18(7), 1527-1554. https://doi.org/10.1162/neco.2006.18.7.1527. [Deep belief nets.]
22. Ho, J., Jain, A., & Abbeel, P. (2020). *Denoising Diffusion Probabilistic Models*. arXiv preprint. arXiv:2006.11239. [Diffusion models.]
23. Hochreiter, S., & Schmidhuber, J. (1997). *Long Short-Term Memory*. Neural Computation, 9(8), 1735-1780. https://doi.org/10.1162/neco.1997.9.8.1735. [LSTM.]
24. Hoffmann, J., Borgeaud, S., Mensch, A., et al. (2022). *Training Compute-Optimal Large Language Models* (22 authors). arXiv preprint. arXiv:2203.15556. [Compute-optimal scaling.]
25. Hubinger, E., van Merwijk, C., Mikulik, V., Skalse, J., & Garrabrant, S. (2019). *Risks from Learned Optimization in Advanced Machine Learning Systems*. arXiv preprint. arXiv:1906.01820. [Mesa-optimization.]
26. Jumper, J., Evans, R., Pritzel, A., et al. (2021). *Highly accurate protein structure prediction with AlphaFold* (34 authors). Nature, 596(7873), 583-589. https://doi.org/10.1038/s41586-021-03819-2. [AlphaFold.]
27. Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). *Scaling Laws for Neural Language Models*. arXiv preprint. arXiv:2001.08361. [Power-law loss scaling.]
28. Kingma, D. P., & Ba, J. (2014). *Adam: A Method for Stochastic Optimization*. arXiv preprint. arXiv:1412.6980. [Adam optimizer.]
29. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). *ImageNet classification with deep convolutional neural networks*. Communications of the ACM, 60(6), 84-90. https://doi.org/10.1145/3065386. [Deep CNN for ImageNet; reprint of the original conference paper.]
30. LeCun, Y., Bengio, Y., & Hinton, G. (2015). *Deep learning*. Nature, 521(7553), 436-444. https://doi.org/10.1038/nature14539. [Review of deep learning.]
31. LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). *Gradient-based learning applied to document recognition*. Proceedings of the IEEE, 86(11), 2278-2324. https://doi.org/10.1109/5.726791. [Convolutional networks for document recognition.]
32. Mnih, V., Kavukcuoglu, K., Silver, D., et al. (2015). *Human-level control through deep reinforcement learning* (19 authors). Nature, 518(7540), 529-533. https://doi.org/10.1038/nature14236. [Deep Q-networks on Atari.]
33. OpenAI. (2023). *GPT-4 Technical Report*. arXiv preprint. arXiv:2303.08774. [Technical report with undisclosed internals.]
34. Ouyang, L., Wu, J., Jiang, X., et al. (2022). *Training language models to follow instructions with human feedback* (20 authors). arXiv preprint. arXiv:2203.02155. [InstructGPT.]
35. Radford, A., Kim, J. W., Hallacy, C., et al. (2021). *Learning Transferable Visual Models From Natural Language Supervision* (12 authors). arXiv preprint. arXiv:2103.00020. [CLIP.]
36. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). *Learning representations by back-propagating errors*. Nature, 323(6088), 533-536. https://doi.org/10.1038/323533a0. [Backpropagation learning procedure.]
37. Russakovsky, O., Deng, J., Su, H., et al. (2015). *ImageNet Large Scale Visual Recognition Challenge* (12 authors). International Journal of Computer Vision, 115(3), 211-252. https://doi.org/10.1007/s11263-015-0816-y. [The ILSVRC benchmark.]
38. Russell, S. (2019). *Human Compatible: Artificial Intelligence and the Problem of Control*. Viking. ISBN 978-0525558613. [Control problem and proposed approach.]
39. Russell, S., & Norvig, P. (2020). *Artificial Intelligence: A Modern Approach* (4th ed.). Pearson. ISBN 978-0134610993. [General AI textbook.]
40. Schaeffer, R., Miranda, B., & Koyejo, S. (2023). *Are Emergent Abilities of Large Language Models a Mirage?* arXiv preprint. arXiv:2304.15004. [Metric-choice explanation of emergence.]
41. Silver, D., Huang, A., Maddison, C. J., et al. (2016). *Mastering the game of Go with deep neural networks and tree search* (20 authors). Nature, 529(7587), 484-489. https://doi.org/10.1038/nature16961. [AlphaGo.]
42. Silver, D., Schrittwieser, J., Simonyan, K., et al. (2017). *Mastering the game of Go without human knowledge* (17 authors). Nature, 550(7676), 354-359. https://doi.org/10.1038/nature24270. [AlphaGo Zero.]
43. Sutskever, I., Vinyals, O., & Le, Q. V. (2014). *Sequence to Sequence Learning with Neural Networks*. arXiv preprint. arXiv:1409.3215. [Seq2seq.]
44. Sutton, R. S., & Barto, A. G. (2018). *Reinforcement Learning: An Introduction* (2nd ed.). MIT Press. ISBN 978-0262039246. [RL textbook.]
45. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). *Attention Is All You Need*. arXiv preprint. arXiv:1706.03762. [Transformer.]
46. Wei, J., Tay, Y., Bommasani, R., et al. (2022). *Emergent Abilities of Large Language Models* (16 authors). arXiv preprint. arXiv:2206.07682 (arXiv comments indicate publication in Transactions on Machine Learning Research). [Emergent-abilities claim.]
47. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). *Chain-of-Thought Prompting Elicits Reasoning in Large Language Models*. arXiv preprint. arXiv:2201.11903. [Chain-of-thought prompting.]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Existence, authorship, year, title of each entry | Established (bibliographic record) | Crossref / DataCite checks, 2026-10-15 | High |
| Neural-network foundations (4.1) | Established (published methods) | Rumelhart et al. (1986); Hochreiter & Schmidhuber (1997); LeCun et al. (1998) | High |
| Loss scales as power law with size, data, compute | Well Supported (empirical; originates in a preprint) | Kaplan et al. (2020) | Moderate to high within tested ranges |
| Model size and tokens should scale equally | Preliminary / Well Supported (single laboratory study with later discussion) | Hoffmann et al. (2022) | Moderate |
| Emergent abilities are genuine discontinuities | Contested | Wei, Tay et al. (2022) vs. Schaeffer et al. (2023) | Active debate |
| Mesa-optimization as a risk | Hypothetical | Hubinger et al. (2019) | Low-moderate |
| Researcher survey on AI progress | Preliminary (opinion survey) | Grace et al. (2024) | Moderate; reports beliefs |
| Annotation fidelity to abstracts | Well Supported (checked for 13 preprints' abstracts) | arXiv abstract pages | Moderate; remaining annotations rest on titles |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.