← Study

The bibliography

The reading

Every paper on the 2026 H1 list, in the order it was published, section by section. Each row links out to the paper itself; a row shows a close-read only when one exists in the Atlas. Working through this is the point — the list is the queue, not the achievement.

Sebastian Raschka · Ahead of AIThe source list ↗

CAPTURED

165 On the list10 Starred0 In the Atlas165 To read
A brushed steel bookend with a copper edge holding two books, on paper.

The ten

One must-read per section, chosen by the source author

The sections

10 sections · 165 papers
§1 Architecture and Model Design 22§2 Efficient Training and Scaling 15§3 Inference Efficiency and KV Cache 8§4 Sparse Attention and Long Context 9§5 Reasoning and Test-Time Compute 15§6 Reinforcement Learning and RLVR 25§7 Agent Systems and Tool Use 33§8 Coding Agents and Software Engineering 16§9 Diffusion Language Models 9§10 Model Evaluation and Benchmarks 13

§1 · Architecture and Model Design

22 papers · ★1Section page →
Models

§2 · Efficient Training and Scaling

15 papers · ★1Section page →
ModelsData

§3 · Inference Efficiency and KV Cache

8 papers · ★1Section page →
ServingInference economicsInference Engineering · inference-bench

§4 · Sparse Attention and Long Context

9 papers · ★1Section page →
ServingModelsInference Engineering · inference-bench

§5 · Reasoning and Test-Time Compute

15 papers · ★1Section page →
ModelsInference economicsReasoning book ch4–5

§6 · Reinforcement Learning and RLVR

25 papers · ★1Section page →
ModelsReasoning book ch6–8 · RLHF book

§7 · Agent Systems and Tool Use

33 papers · ★1Section page →
Harnessesthe-machine

§8 · Coding Agents and Software Engineering

16 papers · ★1Section page →
Harnessesthe-machine

§9 · Diffusion Language Models

9 papers · ★1Section page →
Models

§10 · Model Evaluation and Benchmarks

13 papers · ★1Section page →
Evals
What counts as read

A paper is on this list because someone worth reading put it there. That is a pointer, not a claim: it counts as read only once it has a close-read file in kb/<topic>/raw/, which is what a close-read link on a row means. There is deliberately nothing to tick off here — the study desk is the only writer of study state, and a second way to mark something done is a second version of the truth.

The Atlas join

Graduation is computed on every render, never maintained by hand: a row links to its close-read when the paper’s arXiv id matches a source already in the Atlas. 0 of 165 have one. The two sets were selected on different criteria — this list adds to the shelf rather than confirming it — so that number is the queue’s progress, and it starts where every queue starts. The Atlas →