Coverage & Gaps¶

What this collection covers, where it is thin, and what is queued next - papers and explainers.

This page exists so the collection's boundaries are explicit. A curated list is only trustworthy if it says what it left out. Entries here are candidates, not promises - see CONTRIBUTING.md if you want to write one.

Last reviewed: 2026-10-16 · Papers at review time: 156

On 2026-09-30 the collection broadened from research papers to "everything AI": landmark essays joined as a paper category, robotics got its own category, and unnumbered explainers now cover what no single paper does - model families, benchmarks, compute and economics, policy, the ecosystem and the open questions.

Coverage map¶

Explainers are not numbered and so are not in this map; EXPLAINERS.md lists them all with their review dates.

Queued: recent work (added 2026-10-16)¶

2025 is the thinnest year in the collection (even after the 2026-10-16 additions) and 2026 has one paper. Candidates, none yet summarised (check the dates against the source before writing):

• Kimi k1.5 / K2 - the Moonshot line (Muon is already 144).
• LLaDA - diffusion language models, a line the collection does not cover at all.
• More 2026 papers - Mamba-3 is the only one so far; Raschka's 2026 list (architecture section: Nemotron 3 Super, Gated DeltaNet-2, Attention Residuals) is a starting point, rest of it is paywalled.

Queued: classics and cross-cutting¶

• Greshake et al. 2023 (indirect prompt injection) - the paper the queued jailbreak explainer needs.
• LLM watermarking (Kirchenbauer et al., 2023) and Concrete Problems in AI Safety (2016).
• Model merging (task arithmetic, TIES), Lost in the Middle, MemGPT, WebArena, DINOv2.
• AI 2027 (essay, forecast) and GPTs are GPTs (Eloundou et al., 2023) - the anchor for an AI-and-jobs explainer.

Queued: high priority¶

Papers whose absence is most likely to leave a reader with a hole in their mental model.

Code and software engineering¶

• SWE-agent (Yang et al., 2024) - the agent-computer interface paper, the other half of SWE-bench.
• AlphaCode (Li et al., 2022) - competition-level code generation by sampling and filtering at scale.

Interpretability & safety¶

• Toy Models of Superposition (Elhage et al., 2022) - why features share neurons; the setup that Sparse Autoencoders answer.
• Alignment Faking in Large Language Models (Greenblatt et al., 2024) - pairs with Sleeper Agents.
• Constitutional Classifiers (Anthropic, 2025) - the production defence against universal jailbreaks.

Long context & efficiency¶

• Ring Attention / context parallelism - the systems answer to million-token contexts.
• Medusa / EAGLE - the successors to Speculative Decoding.

Evaluation¶

• HELM (Liang et al., 2022) and BIG-Bench (Srivastava et al., 2022) - holistic evaluation and the crowd-sourced benchmark suite.
• GPQA and Humanity's Last Exam - covered in the knowledge benchmarks explainer, not yet summarised on their own.

Queued: medium priority¶

Generative modelling beyond images¶

• MusicLM / Stable Audio - music generation, next to AudioLM.
• Veo, Movie Gen and open video models - video generation after Sora.

Techniques¶

• Self-RAG, HyDE, and query rewriting - the retrieval techniques practitioners reach for after Dense Retrieval.
• Mixture-of-Depths and early-exit - conditional compute in the depth dimension.
• DPO variants (IPO, SimPO, ORPO) - the preference-optimisation family after DPO.

Robotics¶

• Diffusion Policy (Chi et al., 2023) and Gemini Robotics (2025) - the other major lines beside RT-2 and pi0.

Deep learning prerequisites¶

The collection covers roots selectively (Word2Vec, Seq2Seq, VAE, PPO, ResNet, U-Net, Adam). Candidates for completing that layer: AlexNet (2012), Batch/Layer Normalization, Dropout, LSTM (1997), and DQN / AlphaGo - the run-up to AlphaZero. Each is foundational; each is one step further from generative AI, so they stay optional rather than assumed.

Queued: explainers¶

• Model families: xAI Grok, Microsoft Phi (beyond phi-1), Moonshot Kimi.
• Concepts: mixture-of-experts routing in practice, multimodal tokenization (how images and audio become tokens), embeddings beyond retrieval.
• Applications: AI for code (the coding-assistant landscape), AI in science, AI search.
• Safety: jailbreaks and prompt injection as a threat class (linking AI threat modelling).
• Concepts: hallucination (pairs with 149) and multilingual / non-English models.
• Society: AI and jobs, copyright and training data litigation, energy use.

Deliberately out of scope¶

• Model cards and system cards that contain no methodological contribution - their facts go into the model-family explainers instead.
• Incremental version bumps where the previous entry already covers the technique.
• Papers with no public write-up - if there is nothing citable to link, there is nothing to summarize.
• Hands-on build guides and certification prep - those are the site's Build pillar and certification guides, not paper summaries.
• Anything the summary would have to invent details about. Every entry here is written from the published record; where a number is uncertain the summary says so rather than guessing.

How to close a gap¶

1. Pick an entry above (or propose one).
2. For a paper, follow "Adding a new paper summary" in CONTRIBUTING.md; for an explainer, follow "Adding an explainer" there.
3. Update the coverage table (papers only) and remove the entry from the queue above.