At a glance The ocean of papers is vast and deep, and diving in blind will only drive you away — this page lays out five thematic reading routes (LLMs, generative AI, application engineering, alignment and safety, plus a less-but-better fallback), each with a reading order, a self-check for when you finish, and a suggested time budget.

Reading Paths

In one sentence: reading papers is not about "how many you've read" but about "taking the right path" — first pick your main line, then decide which papers to read and how deep to go. This page arranges the papers covered in this section into five thematic routes, each with a reading order and a self-check of what you should be able to answer when you're done.

Start with a few numbers: the cs.LG (machine learning) category alone on arXiv gains more than 20,000 new papers a year; add in cs.CL (NLP), cs.CV (vision), and cs.AI, and hundreds of new papers go online every day. At this scale, two brutal facts hold:

No one can read them all. Even top researchers close-read at most three to five papers a week, filtering the rest by title and abstract alone.
Not filtering is a disaster. Paper quality is wildly uneven: there are paradigm-shifting foundational works, engineering-tuning reports, and padded filler. Read indiscriminately, and three months later you'll be holding a pile of fragments, unable to describe the skeleton of the field.

That is what a "path" is for. A path = a goal + an order + a check standard:

No path:    skim a diffusion paper today → read an agent paper tomorrow → three months later:
            you've seen everything and can explain none of it

With path:  pick your main line first (LLMs? generation? engineering? alignment?)
            → read in the planned order (each paper builds on the previous one)
            → test yourself with the checks (can you answer those questions?)
            → three months later: a structured thematic map has taken shape

Read along a path, and no paper is an isolated file — each is a node on a map, where the question one paper raises is precisely the one the next paper answers. This interlocking of knowledge nodes makes your understanding compound. The site's companion pages support every step: The Paper Map provides the macro coordinate system, Classic Paper Deep-Dives the paragraph-by-paragraph microscope, and Frontier Developments the always-on radar.

2. The Five Routes at a Glance

Four main tracks follow today's hottest AI concepts (LLMs, generative AI, application engineering, alignment and safety), plus one "less but better" route as a fallback:

Route	Sequence	Papers	Time budget	Best for
Route 1 The LLM Track	Transformer → BERT → GPT-3 → InstructGPT → Llama	4 + 1 optional	1 week skimmed / 2 weeks close-read	Everyone — understand "where language intelligence comes from"
Route 2 The Generative AI Track	GAN → DDPM → LDM → DiT / the Sora report	3 + 2 optional	1 week skimmed	Anyone interested in image/video generation
Route 3 The Application Engineering Track	RAG → ReAct → LoRA → GraphRAG	3 + 1 optional	4–5 days skimmed	Engineers and anyone putting AI into production
Route 4 The Alignment and Safety Track	InstructGPT → Constitutional AI → DPO	3	3–4 days skimmed	Anyone who cares about trustworthy AI or is job-hunting for LLM roles
Route 5 Less but Better	Attention + GPT-3 + RAG + LoRA + DDPM	5	1–2 hours per paper	Anyone short on time who wants one minimal, complete set

Routes Stack — You Don't Have to Pick Just One

The five routes are not an either/or choice. Route 5 is the foundation for everyone — those five papers alone give you a feel for the whole landscape; Routes 1 and 3 are the engineer's standard kit; Route 4 is the main line for anyone going into alignment research. Suggested order: walk Route 5 first to build confidence, then enter the other tracks according to your goal.

3. Route 1: The LLM Track

Goal: understand the full causal chain of "why large language models work" — from the attention mechanism and the pretraining paradigm to scaling laws, alignment tuning, and open-source reproduction.

The papers and the logic of the sequence

#	Paper (authors, year)	Contribution in one sentence	What problem from the previous paper it solved
1	Attention Is All You Need (Vaswani et al., 2017)	Replaced recurrence with self-attention and introduced the Transformer	Solves the RNN's inability to parallelize and its fading long-range dependencies
2	BERT (Devlin et al., 2018)	Bidirectional Transformer + masked language model (MLM)	Solves "how to pretrain on vast amounts of unlabeled text"
3	GPT-3 (Brown et al., 2020)	175 billion parameters; demonstrated in-context learning	Solves "downstream tasks still need fine-tuning" — at scale, few-shot examples are enough
4	InstructGPT (Ouyang et al., 2022)	Used RLHF to make the model "understand people and speak like one"	Solves "powerful pretrained models that don't listen" — that is, alignment
5 (optional)	Llama (Touvron et al., 2023)	Openly reproducible foundation models at 7B–65B scale	Solves "frontier models are closed off and can't be studied" — the starting point of the open-source ecosystem

4. Route 2: The Generative AI Track

Goal: understand "how machines create images and video from scratch" — from adversarial games to denoising diffusion, then into latent space and video generation.

The papers and the logic of the sequence

#	Paper (authors, year)	Contribution in one sentence	What problem from the previous paper it solved
1	GAN (Goodfellow et al., 2014)	Generator and discriminator trained against each other to produce realistic images	Solves "generative models lacked a differentiable objective" — use the discriminator as the loss function
2	DDPM (Ho et al., 2020)	Add noise forward step by step, denoise backward — the diffusion model	Solves GAN's unstable training and mode collapse — a simple, stable objective
3	LDM / Stable Diffusion (Rombach et al., 2022)	Moves the diffusion process into latent space	Solves "pixel-space diffusion is too expensive" — a low-dimensional latent space is fast and good
4 (optional)	DiT (Peebles et al., 2023)	Diffusion on a Transformer backbone (Diffusion Transformer)	Solves "U-Net's limited scalability" — the Transformer architecture improves smoothly with scale
5 (optional)	The Sora technical report (OpenAI, 2024)	A video generation model framed as a "world simulator"	From generating static images to temporally consistent video generation

5. Route 3: The Application Engineering Track

Goal: understand "how methods from papers become shippable product capabilities" — retrieval, agents, efficient fine-tuning; this is the set most often asked about when landing AI in production.

The papers and the logic of the sequence

#	Paper (authors, year)	Contribution in one sentence	What problem from the previous paper it solved
1	RAG (Lewis et al., 2020)	Combines retrieval with generation so the model can cite external knowledge	Solves "a model's knowledge is frozen and it hallucinates" — bring in an external knowledge base
2	ReAct (Yao et al., 2022)	An alternating loop of reasoning (Thought) and acting (Action)	Solves "reasoning from parameters alone, with no contact with the outside world" — lets the model call tools
3	LoRA (Hu et al., 2021)	Efficient fine-tuning via low-rank decomposition, training only a small set of parameters	Solves "full fine-tuning is too expensive to afford or to store" — freeze the backbone and learn only the delta
4 (optional)	GraphRAG (Microsoft, 2024)	Indexes with a knowledge graph before retrieving	Solves "vector retrieval can't answer global questions" — graph structure adds relational reasoning

6. Route 4: The Alignment and Safety Track

Goal: understand "how to make models both capable and obedient" — from RLHF to self-alignment, then to DPO, which gets by without reinforcement learning. This is a high-frequency topic in LLM-role interviews and in AI Safety and Governance discussions.

The papers and the logic of the sequence

#	Paper (authors, year)	Contribution in one sentence	What problem from the previous paper it solved
1	InstructGPT (Ouyang et al., 2022)	Three-stage RLHF: SFT → reward model → reinforcement learning	Solves "powerful pretrained models that don't follow instructions"
2	Constitutional AI (Bai et al., 2022)	A set of principles by which the model critiques and corrects itself	Solves "RLHF depends on massive human labeling" — rules replace the labelers
3	DPO (Rafailov et al., 2023)	Reduces RLHF to a preference-classification loss	Solves "reinforcement learning is engineering-heavy and unstable" — direct optimization without RL

7. Route 5: Less but Better

Goal: when time is extremely scarce, build a minimal yet complete mental model of "AI's hot concepts" from just 5 papers. Each of the 5 maps onto a pillar: mechanism (Attention), scale (GPT-3), engineering (RAG and LoRA), generation (DDPM).

#	Paper	Which pillar it represents	Skim time
1	Attention Is All You Need (2017)	The foundational mechanism of modern LLMs	1–2 hours
2	GPT-3 (2020)	Scaling laws and in-context learning	1–2 hours
3	RAG (2020)	Retrieval augmentation: the engineering answer to hallucination	1 hour
4	LoRA (2021)	Efficient fine-tuning: adapting large models on a small budget	1 hour
5	DDPM (2020)	Diffusion models: the technical foundation of generative AI	1–2 hours

Reading order and method

For each paper, read only the abstract → introduction → figures and tables → conclusion, and don't grind through the formulas. When you finish each one, write down in a single sentence of your own what problem it solved. Once all five are done, you can already answer the three trunk questions of today's hot AI topics: why large models work, how to make them more useful, and how images get generated.

The Post-Reading Check (Baseline Version)

After the five papers, you should be able to explain each to a non-technical friend in one plain sentence: Attention is "letting the model look at the whole context"; GPT-3 is "feed it enough and it can do a bit of everything"; RAG is "when it doesn't know, it looks it up first"; LoRA is "changing a little is enough"; DDPM is "walking the noise back to an image, step by step". If you can put it in plain words, you've truly understood it.

8. How to Budget Your Reading Time

The scarcest resource in paper reading is time. Set the budget first, then the depth — don't let perfectionism wreck your rhythm:

Reading mode	Time per paper	What to read	What you get out of it
Skim	1–2 hours	Abstract + introduction + figures + conclusion	Can state what problem it solves, the core method in one sentence, and how well it works
Close read	4–6 hours	Full text + method + ablations + limitations + notes	Can explain the mechanism, the ablation evidence, and the boundaries of applicability — even judge whether it can be reused
Study	8+ hours	Close read + derivations + reproduction	Can spot flaws in the paper and propose improvements (for researchers)

Two pieces of practical advice:

The 80/20 rule: about 80% of papers deserve only a skim (1–2 hours), 15% deserve a close read (4–6 hours), and 5% deserve full study. Close reading is an investment; skimming is an expense — use the expense to maintain breadth and the investment to buy depth.
One hour a day beats a seven-hour weekend cram: understanding papers takes overnight fermentation; small, steady doses are far more solid than a single blitz. For methods and note-taking in detail, see Reading Discipline and FAQ.

The Most Common Time Trap

"Close-read every paper" is the most expensive mistake — it caps you at three papers a month, after which you quit over the slow progress. The right posture: skim first to quickly shortlist the candidates worth a close read, then put 4–6 hours into just a few of them. To judge whether a paper is worth a close read, look up where it sits in the genealogy on The Paper Map.

References

Everything below is a real, publicly available resource for going deeper on your own:

The arXiv preprint library — the first-release home of the vast majority of AI papers
Vaswani et al. Attention Is All You Need (NeurIPS 2017) — the original Transformer paper
Devlin et al. BERT: Pre-training of Deep Bidirectional Transformers (2018)
Brown et al. Language Models are Few-Shot Learners (NeurIPS 2020) — GPT-3
Ouyang et al. Training language models to follow instructions with human feedback (2022) — InstructGPT
Touvron et al. LLaMA: Open and Efficient Foundation Language Models (2023)
Goodfellow et al. Generative Adversarial Nets (NeurIPS 2014) — GAN
Ho et al. Denoising Diffusion Probabilistic Models (NeurIPS 2020) — DDPM
Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2022) — LDM / Stable Diffusion
Peebles et al. Scalable Diffusion Models with Transformers (ICCV 2023) — DiT
Lewis et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) — RAG
Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models (ICLR 2023)
Hu et al. LoRA: Low-Rank Adaptation of Large Language Models (ICLR 2022)
Microsoft. From Local to Global: A Graph RAG Approach to Query-Focused Summarization (2024) — GraphRAG
Bai et al. Constitutional AI: Harmlessness from AI Feedback (2022)
Rafailov et al. Direct Preference Optimization: Your Language Model is Secretly a Reward Model (NeurIPS 2023) — DPO
OpenAI. Video generation models as world simulators (the Sora technical report, 2024)
Jay Alammar. The Illustrated Transformer (2018) — the classic example of an illustrated Transformer close read
The Annotated Transformer (Harvard NLP) — a close reading of the Transformer with line-by-line annotations and runnable code

Reading Paths

Reading Paths

1. Why Read Along Paths: An Ocean of Papers Needs Navigation

2. The Five Routes at a Glance

3. Route 1: The LLM Track

The papers and the logic of the sequence

Suggested reading order

4. Route 2: The Generative AI Track

The papers and the logic of the sequence

Suggested reading order

5. Route 3: The Application Engineering Track

The papers and the logic of the sequence

Suggested reading order

6. Route 4: The Alignment and Safety Track

The papers and the logic of the sequence

Suggested reading order

7. Route 5: Less but Better

Reading order and method

8. How to Budget Your Reading Time

Further Reading

References

Reading Paths

Reading Paths ​

1. Why Read Along Paths: An Ocean of Papers Needs Navigation ​

2. The Five Routes at a Glance ​

3. Route 1: The LLM Track ​

The papers and the logic of the sequence ​

Suggested reading order ​

4. Route 2: The Generative AI Track ​

The papers and the logic of the sequence ​

Suggested reading order ​

5. Route 3: The Application Engineering Track ​

The papers and the logic of the sequence ​

Suggested reading order ​

6. Route 4: The Alignment and Safety Track ​

The papers and the logic of the sequence ​

Suggested reading order ​

7. Route 5: Less but Better ​

Reading order and method ​

8. How to Budget Your Reading Time ​

Further Reading ​

References ​

Reading Paths

1. Why Read Along Paths: An Ocean of Papers Needs Navigation

2. The Five Routes at a Glance

3. Route 1: The LLM Track

The papers and the logic of the sequence

Suggested reading order

4. Route 2: The Generative AI Track

The papers and the logic of the sequence

Suggested reading order

5. Route 3: The Application Engineering Track

The papers and the logic of the sequence

Suggested reading order

6. Route 4: The Alignment and Safety Track

The papers and the logic of the sequence

Suggested reading order

7. Route 5: Less but Better

Reading order and method

8. How to Budget Your Reading Time

Further Reading

References