AI and Machine Learning Research Papers: 15 Must-Read Papers With Free PDFs (2026)

Download 15 landmark AI and machine learning papers as free PDFs, grouped by topic (deep learning, vision, NLP, LLMs, generative AI) with reading order.

Quick answer

Almost every landmark AI and machine learning research paper is free as a PDF on arXiv. Start with Attention Is All You Need, ResNet, Adam and BERT, then follow your subtopic: vision (U-Net, ViT), language models (GPT-3, InstructGPT), generative AI (GANs, diffusion) or reinforcement learning. Below are 15 verified free PDFs, grouped by topic, with a reading order.

Search machine learning papers free

Where to Get AI and Machine Learning Research Papers as Free PDFs

AI is one of the easiest fields to read for free. Since the early 2010s, nearly every important machine learning paper has been posted on arXiv, the open preprint server run by Cornell, before or alongside its conference version (NeurIPS, ICML, ICLR, CVPR, ACL). That means you can download "Attention Is All You Need" or the ResNet paper as a PDF in one click, legally, with no university login.

Updated October 2026.

The hard part is not access. It is choosing what to read. There are tens of thousands of new AI papers every year, and most students waste weeks on papers that are too new, too narrow or too hard. This page gives you a short list of 15 landmark papers, grouped by subtopic, each with a free PDF link we checked, plus a simple way to find more papers on any AI topic.

Jump to: How to find more · Reading order table · Foundations · Computer vision · Language models · Generative AI · Reinforcement learning and AI for science · Tips for choosing a topic · FAQ

Looking for sample papers in other fields? See our research paper examples by field and the companion list of education research papers.

How to Find More AI Research Papers in Ocean of Papers

Ocean of Papers searches 250M+ papers from six sources (OpenAlex, PubMed, arXiv, bioRxiv, medRxiv and Europe PMC) and shows a free, legal PDF link whenever one exists through arXiv, Unpaywall, PubMed Central or the preprint servers. For AI topics, arXiv coverage means most results have a PDF.

1. Search your topic. Try machine learning, deep learning, large language models, diffusion models or reinforcement learning. Sort by most cited to see the classics first, or by newest to see current work.

2. Paste a DOI or title for an exact paper. arXiv papers have DOIs of the form 10.48550/arXiv.1706.03762. Paste one into search to jump straight to that paper (see how to download a paper by DOI).

3. Build a citation map from a landmark paper. Open "Attention Is All You Need" from a search result and click Build map. You get about 30 connected papers (choose 15, 30 or 50). Switch modes to see:

Prior works: the papers it builds on (references), useful for background reading
Derivative works: papers that cite it, which is how you find BERT, GPT-style models and newer work
Similar: papers connected by shared references and co-citation

Use the Timeline view (year against citations) to see which follow-up papers took off. Bigger circles are more cited; darker ones are newer. Click any link to see why two papers are connected. You can export the map's papers as BibTeX, RIS or CSV for your reference manager. Read more in our citation network guide and literature mapping tool page.

Search, PDFs and exports are free and unlimited. The free plan includes 5 new citation maps per month; switching modes or expanding a map does not count. See pricing for unlimited maps.

Search the topic — e.g. /?q=machine learning, sort by most cited
Open the landmark paper — Free arXiv PDF in one click
Build map — Prior works, Derivative works or Similar
Switch to Timeline — See which follow-ups became influential
Export — BibTeX, RIS or CSV for Zotero or your thesis

AI Subtopics: What to Read First

Use this table to pick a starting point. "Read first" is the paper that makes the rest of the subtopic easier to follow. "Then" is the natural next paper. All papers in the table are on this page with free PDFs.

Reading order by subtopic
Feature	Read first	Then read	Good search to go further
Deep learning foundations	Adam (Kingma & Ba, 2014)	Dropout (Hinton et al., 2012)	optimization deep learning
Computer vision	ResNet (He et al., 2015)	Vision Transformer (Dosovitskiy et al., 2020)	image classification
Medical image analysis	U-Net (Ronneberger et al., 2015)	ResNet	medical image segmentation
NLP and transformers	word2vec (Mikolov et al., 2013)	Attention Is All You Need (2017)	natural language processing transformer
Large language models	BERT (Devlin et al., 2018)	GPT-3, then InstructGPT	large language models
Generative AI	GANs (Goodfellow et al., 2014)	DDPM diffusion (Ho et al., 2020)	diffusion models
Reinforcement learning	Atari DQN (Mnih et al., 2013)	InstructGPT (RL from human feedback)	deep reinforcement learning
AI for science	AlphaFold (Jumper et al., 2021)	U-Net (for imaging work)	protein structure prediction deep learning

Foundations: How Neural Networks Are Trained

These two papers explain tools that appear in almost every deep learning experiment section. They are short and readable, which makes them a good first step if you are new to research papers.

The Dropout idea was later expanded into a longer journal paper, Srivastava et al. (2014) in the Journal of Machine Learning Research, which is also free to read.

Adam: A Method for Stochastic Optimization — Kingma, D. P., Ba, J. (2014), ICLR 2015 / arXiv. Method. The optimizer most deep learning code still uses by default. Read it to understand learning rates and moment estimates; the one-page algorithm box is easy to follow.
Improving Neural Networks by Preventing Co-adaptation of Feature Detectors — Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., Salakhutdinov, R. R. (2012), arXiv. Method. The original dropout paper: randomly switching off units during training to reduce overfitting. A short, idea-first paper that shows how a simple trick is tested across many benchmarks.

Computer Vision

Vision papers are a good model for writing results sections: clear benchmark tables, ablations and architecture diagrams. ResNet made very deep networks trainable; U-Net became the standard for segmentation, especially in medicine; the Vision Transformer showed that transformers work on images too.

Deep Residual Learning for Image Recognition — He, K., Zhang, X., Ren, S., Sun, J. (2015), CVPR 2016 / arXiv. Empirical. Introduced residual (skip) connections, now used in nearly every deep architecture. Study how the authors use one small diagram and a few tables to make a big claim convincing.
U-Net: Convolutional Networks for Biomedical Image Segmentation — Ronneberger, O., Fischer, P., Brox, T. (2015), MICCAI 2015 / arXiv. Method. The go-to architecture for segmenting medical and microscopy images from few labelled examples. Essential if your thesis touches medical imaging, and short enough to read in one sitting.
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale — Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. (2020), ICLR 2021 / arXiv. Empirical. The Vision Transformer (ViT): split an image into patches and treat them like words. Read it to see how pre-training data size changes which architecture wins.

Natural Language Processing and Large Language Models

This is the most-searched part of AI research right now. Read these in order: word vectors, then attention, then the transformer, then pre-trained models, then instruction-tuned models. Each paper is easier once you have read the one before it.

Efficient Estimation of Word Representations in Vector Space — Mikolov, T., Chen, K., Corrado, G., Dean, J. (2013), arXiv (ICLR 2013 workshop). Method. The word2vec paper, home of the famous "king minus man plus woman is close to queen" example. The best starting point for understanding embeddings.
Neural Machine Translation by Jointly Learning to Align and Translate — Bahdanau, D., Cho, K., Bengio, Y. (2014), ICLR 2015 / arXiv. Method. Introduced the attention mechanism for translation. Reading it first makes the transformer paper much easier to follow.
Attention Is All You Need — Vaswani, A., Shazeer, N., Parmar, N., et al. (2017), NeurIPS 2017 / arXiv. Method. The transformer paper behind modern language models. If you read only one AI paper, read this one, and use it as the seed for a citation map of the whole LLM field.
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding — Devlin, J., Chang, M.-W., Lee, K., Toutanova, K. (2018), NAACL 2019 / arXiv. Empirical. Made "pre-train, then fine-tune" the standard recipe in NLP. Its ablation tables are a model for showing which part of a method matters.
Language Models are Few-Shot Learners — Brown, T. B., Mann, B., Ryder, N., et al. (2020), NeurIPS 2020 / arXiv. Empirical. The GPT-3 paper: scaling a language model lets it do new tasks from a few examples in the prompt. Long, so read the abstract, introduction and Section 3 first.
Training Language Models to Follow Instructions with Human Feedback — Ouyang, L., Wu, J., Jiang, X., et al. (2022), NeurIPS 2022 / arXiv. Empirical. The InstructGPT paper, which explains reinforcement learning from human feedback (RLHF). Key result: a 1.3B instruction-tuned model was preferred by human raters over the 175B GPT-3.

Generative AI: GANs and Diffusion Models

Image generators went from GANs to diffusion models in about six years. These two papers are the core of that story and are both readable with basic probability.

Generative Adversarial Networks — Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., et al. (2014), NeurIPS 2014 / arXiv. Method. Two networks, a generator and a discriminator, trained against each other as a game. A short paper with a clear theory section; ideal for learning how a new framework is argued.
Denoising Diffusion Probabilistic Models — Ho, J., Jain, A., Abbeel, P. (2020), NeurIPS 2020 / arXiv. Method. The DDPM paper that made diffusion models practical for high-quality image generation. Start here before reading any text-to-image paper.

Reinforcement Learning and AI for Science

Reinforcement learning (RL) trains agents through rewards. AI for science applies deep learning to research problems such as protein folding. AlphaFold is published in Nature, and its full text is free in PubMed Central.

Playing Atari with Deep Reinforcement Learning — Mnih, V., Kavukcuoglu, K., Silver, D., et al. (2013), NIPS Deep Learning Workshop 2013 / arXiv. Empirical. The first deep Q-network (DQN) to learn to play Atari games from raw pixels, using experience replay. The usual entry point to deep RL.
Highly Accurate Protein Structure Prediction with AlphaFold — Jumper, J., Evans, R., Pritzel, A., et al. (2021), Nature. Empirical. Deep learning that predicts 3D protein structures with near-experimental accuracy in CASP14. The best example of AI changing another science, and a model for writing for a general journal.

How to Read a Machine Learning Paper Without Getting Lost

ML papers follow a fairly fixed structure: abstract, introduction, related work, method, experiments (with ablations), conclusion and appendix. You do not need to read them front to back.

Read the abstract and the last paragraph of the introduction. Most papers list their contributions there as bullet points.
Look at the main figure and the main results table. Ask: what is being compared, on which dataset, with which metric?
Read the method section once without the maths, then again with a pen.
Read the ablation study. It tells you which parts of the method actually matter.
Skim related work last. It is a ready-made list of papers to read next, and a citation map shows the same list visually.

When you write your own literature review, group papers by idea, not by year. Our guide on how to write a literature review shows how, and AI tools for literature review compares tools that help with the reading.

Tips for Students Choosing an AI or Machine Learning Research Topic

Most student AI projects fail because the topic is too big ("improve ChatGPT") or needs computing power the student does not have. These tips keep your topic doable.

Pick an application you understand. Crop disease images, local-language text classification, traffic data from your city. Your domain knowledge is an advantage big labs do not have.
Use a pre-trained model. Fine-tuning ResNet, ViT or BERT on a small dataset runs on a free cloud GPU. Training from scratch usually does not.
Check the dataset exists first. No dataset, no project. Search your topic plus "dataset" and confirm the data is public and licensed for research.
Find a baseline paper. Choose one recent, well-cited paper on your task and aim to reproduce or slightly extend it. Build a citation map from it to see what has already been tried.
Keep the question narrow and measurable. "Does data augmentation improve U-Net accuracy on a small chest X-ray dataset?" is a good topic. "AI in healthcare" is not.
Plan an ablation. Remove one part of your method and measure the drop. It turns a demo into research.
Write the title last, but draft it early. See research title examples for patterns that work.

Illustrative example topics (not published papers): "Fine-tuning a Vision Transformer for rice leaf disease detection with fewer than 2,000 images"; "Comparing BERT and a multilingual model for sentiment analysis of Tagalog product reviews"; "Effect of class imbalance handling on U-Net segmentation of skin lesions."

For structure and formatting, see how to write a research paper and the IEEE paper format guide, since many CS courses use IEEE style.

Are These AI Papers Really Free and Legal to Download?

Yes. arXiv is an open-access repository where authors post their own papers; reading and downloading them is free and legal. Many conference proceedings (such as NeurIPS and JMLR) are also free on their official sites. AlphaFold is in PubMed Central, the US National Library of Medicine's free archive.

For papers that are not on arXiv, search the title or DOI in Ocean of Papers: it checks Unpaywall, PubMed Central and preprint servers for a legal free copy. If none exists, see how to read paywalled papers legally and our list of websites to download research papers free. We do not link to piracy sites.

FAQ

Where can I download machine learning research papers as PDF for free?

On arXiv, where almost all machine learning papers are posted. Ocean of Papers searches arXiv together with OpenAlex, PubMed, bioRxiv, medRxiv and Europe PMC, so you can search a topic like machine learning and open the free PDF from the results.

What are the most important AI research papers to read first?

A good first five: Attention Is All You Need (transformers), Deep Residual Learning (ResNet), Adam (optimization), BERT (pre-trained language models) and Generative Adversarial Networks. Then pick the subtopic list on this page that matches your interest.

Is there an artificial intelligence research papers free download site that is legal?

Yes: arXiv, PubMed Central and official proceedings sites like JMLR are free and legal. Ocean of Papers links to these legal copies and uses Unpaywall to find open-access versions of journal articles.

What are good deep learning research papers for beginners?

Start with short, idea-focused papers: Dropout, Adam, U-Net and word2vec. They have clear diagrams and limited maths. Leave GPT-3 and very long papers for later, or read only their introduction and main results.

How do I find the latest AI research papers on a topic?

Search the topic in Ocean of Papers and sort by newest, or build a citation map from a landmark paper and switch to Derivative works to see recent papers that cite it. The Timeline view shows which new papers are gaining citations.

Can I use arXiv papers in my thesis references?

Yes, but check whether the paper was later published at a conference or journal and cite that version if your department prefers peer-reviewed sources. Export the citation as BibTeX or RIS, or copy an APA, MLA, Chicago, Harvard or Vancouver citation from the paper's page.

How many papers should I read for an AI literature review?

For an undergraduate thesis, 20 to 40 well-chosen papers is common; for a master's thesis, more. Quality matters more than count: cover the landmark papers, the main baseline for your task, and recent work from the last two to three years.

Next Steps

Pick one subtopic, read its "read first" paper, then build a citation map from it to get your next ten papers. Search machine learning papers on Ocean of Papers to start; search, PDFs and exports are free.