2026 edition · for learners
100 AI papers to read in your lifetime
Not the top of the Top 500: a reading list. Papers are chosen from the experts' must-read votes, at least 20 places are kept for papers from before 1990, and each one passed a readability check: it teaches something lasting, and a learner with the listed prerequisites can get through it. How the list is built.
0 of 100 papers read · saved in this browser only
ImageNet Classification with Deep Convolutional Neural Networks
Sparked the deep learning revolution through breakthrough ImageNet performance and GPU training.
AwardEditor's noteDeep Residual Learning for Image Recognition
Residual connections enabled extremely deep networks, transforming nearly every neural architecture.
AwardEditor's noteNeural Machine Translation by Jointly Learning to Align and Translate
Introduced soft attention, overcoming fixed-vector bottlenecks in neural machine translation.
AwardEditor's noteIntriguing properties of neural networks
Discovered imperceptible adversarial perturbations and their surprising transferability across neural networks.
AwardEditor's noteAuto-Encoding Variational Bayes
Introduced VAEs, amortized variational inference, and the reparameterization trick.
AwardEditor's noteAdam: A Method for Stochastic Optimization
Combined adaptive learning rates and momentum, becoming a dominant optimizer for deep learning.
AwardEditor's noteYou Only Look Once: Unified, Real-Time Object Detection
Revolutionized real-time detection through a unified, single-stage prediction framework.
AwardEditor's noteBatch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Made deep network optimization faster and more stable through mini-batch activation normalization.
AwardEditor's noteSequence to Sequence Learning with Neural Networks
Established end-to-end encoder-decoder learning for mapping variable-length sequences to sequences.
AwardEditor's noteDistributed Representations of Words and Phrases and their Compositionality
Popularized negative sampling, subsampling, and semantic relationships in word embeddings.
AwardEditor's noteAttention Is All You Need
Introduced the Transformer, the architecture underlying modern foundation models.
Editor's noteRich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
Brought deep CNN representations to object detection, launching the region-based detector family.
AwardEditor's noteLong Short-Term Memory
Introduced gated memory cells, enabling effective learning of long-range sequential dependencies.
Editor's noteHuman-level control through deep reinforcement learning
Made deep RL practical from pixels using experience replay and target networks.
Editor's noteDistilling the Knowledge in a Neural Network
Established teacher-student distillation, foundational to model compression and knowledge transfer.
Editor's noteEfficient Estimation of Word Representations in Vector Space
Introduced word2vec's efficient CBOW and skip-gram architectures, transforming NLP representation learning.
Editor's noteNeural Machine Translation of Rare Words with Subword Units
Established byte-pair encoding for neural NLP, enabling practical open-vocabulary modeling.
AwardEditor's noteMaximum Likelihood from Incomplete Data Via the EM Algorithm
Established the general EM framework for latent-variable estimation and incomplete data.
Editor's noteHighly accurate protein structure prediction with AlphaFold
Achieved near-experimental protein structure accuracy; transformed structural biology and computational drug discovery.
Editor's noteSemi-Supervised Classification with Graph Convolutional Networks
Established the canonical simple, scalable GCN architecture for learning on graphs.
Editor's noteLearning representations by back-propagating errors
Popularized efficient training of multilayer networks, enabling learned internal representations and modern deep learning.
Editor's noteThe Anatomy of a Large-Scale Hypertextual Web Search Engine
Introduced Google's search architecture and popularized link-based ranking at internet scale.
Editor's noteA Unified Approach to Interpreting Model Predictions
Unified feature-attribution methods using Shapley values; became a standard practical explanation framework.
Editor's noteNeural Message Passing for Quantum Chemistry
Unified graph neural architectures under the influential message-passing framework.
Editor's noteMastering the game of Go with deep neural networks and tree search
Demonstrated world-class Go through deep policy networks, value estimation, and tree search.
Editor's noteLearning Transferable Visual Models From Natural Language Supervision
Established web-scale image-text contrastive learning and open-vocabulary zero-shot vision.
Editor's noteLanguage Models are Few-Shot Learners
Established large-scale language models as general-purpose few-shot learners without parameter updates.
Editor's noteBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Established bidirectional Transformer pretraining and fine-tuning as the dominant NLP paradigm.
Editor's noteU-Net: Convolutional Networks for Biomedical Image Segmentation
Established the encoder-decoder architecture dominating biomedical segmentation and influencing segmentation across computer vision.
Editor's note“Why Should I Trust You?”: Explaining the Predictions of Any Classifier
Popularized model-agnostic local explanations, establishing a foundational approach to black-box interpretability.
Editor's noteDenoising Diffusion Probabilistic Models
Established practical diffusion training, launching the modern diffusion generation revolution.
Editor's noteSupport-Vector Networks
Established soft-margin support vector machines, transforming classification through maximum-margin learning and kernels.
Editor's noteLoRA: Low-Rank Adaptation of Large Language Models
Made low-rank adaptation the dominant practical approach to parameter-efficient LLM fine-tuning.
Editor's noteExplaining and Harnessing Adversarial Examples
Introduced FGSM and a foundational explanation of adversarial vulnerability in neural networks.
Editor's notePhysics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations
Popularized embedding governing physical equations into neural-network training for scientific computing.
Editor's noteGradient-Based Learning Applied to Document Recognition
Established LeNet-style CNNs as practical end-to-end trained systems for visual recognition.
Editor's noteGenerative Adversarial Nets
Introduced adversarial training, establishing an entirely new paradigm for generative modeling.
AwardEditor's noteTraining language models to follow instructions with human feedback
Established InstructGPT's instruction-tuning and RLHF recipe, transforming conversational AI deployment.
Editor's noteAn Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Demonstrated that pure transformers can rival CNNs, establishing the vision transformer paradigm.
Editor's noteInductive Representation Learning on Large Graphs
Introduced inductive neighborhood aggregation and sampling for large graphs.
Editor's noteA Simple Framework for Contrastive Learning of Visual Representations
SimCLR established augmentation, projection heads, and contrastive objectives as fundamental SSL ingredients.
Editor's noteRandom Forests
Established a powerful, robust tree ensemble that became a foundational practical machine-learning method.
Editor's noteHigh-Resolution Image Synthesis with Latent Diffusion Models
Made high-quality diffusion computationally accessible; foundation of Stable Diffusion.
Editor's noteA Formal Basis for the Heuristic Determination of Minimum Cost Paths
Established A*, the foundational optimal heuristic search algorithm.
Editor's noteEquation of State Calculations by Fast Computing Machines
Introduced Metropolis sampling, foundational to modern computational Bayesian inference.
Editor's noteDropout: A Simple Way to Prevent Neural Networks from Overfitting
Established dropout as a simple, widely adopted regularization technique for neural networks.
Editor's noteQ-learning
Established convergent model-free, off-policy control; foundation for DQN and countless descendants.
Editor's noteProximal Policy Optimization Algorithms
Introduced PPO, becoming a dominant practical policy-gradient method across RL and RLHF.
Editor's noteVisual Instruction Tuning
Popularized accessible visual instruction tuning and the modern open multimodal assistant architecture.
Editor's noteWaveNet: A Generative Model for Raw Audio
Established autoregressive raw-waveform generation and transformed neural speech synthesis.
Editor's noteComputing Machinery and Intelligence
Introduced the imitation game and foundational questions about machine intelligence.
Editor's noteHow Powerful are Graph Neural Networks?
Connected message-passing expressivity to Weisfeiler–Lehman tests and introduced GIN.
Editor's noteLatent Dirichlet Allocation
Made Bayesian topic modeling a central approach to discovering latent document structure.
Editor's noteA decision-theoretic generalization of on-line learning and an application to boosting
Introduced AdaBoost and generalized multiplicative weights, transforming ensemble learning and online prediction.
Editor's noteScaling Laws for Neural Language Models
Quantified predictable power-law relationships between model size, data, compute, and performance.
Editor's noteAxiomatic Attribution for Deep Networks
Introduced principled attribution axioms and Integrated Gradients, a widely adopted gradient-based explanation method.
Editor's noteStrips: A new approach to the application of theorem proving to problem solving
Introduced STRIPS, defining action representations used throughout automated planning.
Editor's notePolicy Gradient Methods for Reinforcement Learning with Function Approximation
Established the policy gradient theorem with function approximation and advantage-based estimators.
Editor's noteFlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
Revolutionized exact attention through IO-aware tiling, enabling faster training and longer contexts.
Editor's noteDistinctive Image Features from Scale-Invariant Keypoints
Established robust scale-invariant feature matching, foundational to recognition, registration, and 3D reconstruction.
Editor's noteChain-of-Thought Prompting Elicits Reasoning in Large Language Models
Demonstrated that intermediate reasoning steps dramatically improve complex problem-solving.
Editor's noteFourier Neural Operator for Parametric Partial Differential Equations
Established Fourier-domain operator learning for fast, resolution-flexible scientific simulation.
Editor's noteDeep reinforcement learning from human preferences
Demonstrated learning complex reward functions from human comparisons; foundational to modern RLHF.
Editor's noteReducing the Dimensionality of Data with Neural Networks
Demonstrated deep autoencoders learning nonlinear low-dimensional representations beyond traditional PCA.
Editor's noteUnderstanding the difficulty of training deep feedforward neural networks
Introduced Xavier initialization and analyzed activation saturation and gradient propagation through deep networks.
Editor's noteRobust Speech Recognition via Large-Scale Weak Supervision
Demonstrated robust multilingual ASR through massive weak supervision and zero-shot generalization.
Editor's noteA logical calculus of the ideas immanent in nervous activity
Establishes the computational neuron and connects neural activity to formal logic.
Editor's noteMasked Autoencoders Are Scalable Vision Learners
MAE demonstrated simple masked-pixel reconstruction as an effective, scalable vision pretraining objective.
Editor's noteRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Established RAG as a foundational architecture combining parametric knowledge with external retrieval.
Editor's noteScore-Based Generative Modeling through Stochastic Differential Equations
Unified score-based and diffusion models through continuous-time stochastic differential equations.
Editor's noteDiffusion Policy: Visuomotor Policy Learning via Action Diffusion
Established action diffusion as a powerful paradigm for learning multimodal visuomotor behavior.
Editor's noteDense Passage Retrieval for Open-Domain Question Answering
Established dense dual-encoder retrieval as a practical alternative to lexical passage search.
Editor's noteSegment Anything
Introduced promptable, general-purpose segmentation and large-scale mask data generation.
Editor's noteThe perceptron: A probabilistic model for information storage and organization in the brain
Established a trainable artificial neuron and foundational learning framework for neural networks.
Editor's noteConstitutional AI: Harmlessness from AI Feedback
Introduced principle-guided self-critique and AI feedback as scalable alignment methods.
Editor's noteBLEU: a Method for Automatic Evaluation of Machine Translation
Introduced BLEU, the benchmark metric that shaped machine translation research for decades.
Editor's noteAmazon.com Recommendations: Item-to-Item Collaborative Filtering
Demonstrated scalable, real-time item-to-item recommendation in a major production system.
Editor's notewav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
Established large-scale self-supervised speech pretraining, dramatically reducing labeled-data requirements.
Editor's noteMastering the game of Go without human knowledge
Demonstrated superhuman mastery through self-play without human demonstrations or expert features.
Editor's noteDeep contextualized word representations
Introduced ELMo, demonstrating the power of contextualized representations before BERT.
Editor's noteDirect Preference Optimization: Your Language Model is Secretly a Reward Model
Simplified preference alignment by eliminating explicit reward-model training and reinforcement-learning optimization.
Editor's noteMonte Carlo sampling methods using Markov chains and their applications
Generalized Metropolis sampling into the widely applicable Metropolis–Hastings algorithm.
Editor's noteConditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data
Introduced CRFs, transforming structured prediction, named entity recognition, and sequence tagging.
Editor's noteDecoupled Weight Decay Regularization
Separated weight decay from adaptive gradient updates, establishing the widely used AdamW optimizer.
Editor's noteA Machine-Oriented Logic Based on the Resolution Principle
Established resolution and unification as foundations of automated logical inference.
Editor's noteSimple statistical gradient-following algorithms for connectionist reinforcement learning
Introduced REINFORCE, establishing unbiased gradient estimation for stochastic policy optimization.
Editor's noteConcrete Problems in AI Safety
Defined five practical safety problems that shaped modern technical alignment research.
Editor's noteGroupLens: An Open Architecture for Collaborative Filtering of Netnews
Pioneered operational collaborative filtering using community ratings to predict individual preferences.
Editor's noteEquality of Opportunity in Supervised Learning
Introduced equal opportunity and equalized odds, foundational criteria for fair classification.
Editor's noteDomain Randomization for Transferring Deep Neural Networks from Simulation to the Real World
Popularized randomized simulation as a practical bridge to real-world robotic perception.
Editor's noteLearning to Predict by the Methods of Temporal Differences
Established TD learning, bootstrapping predictions from successive estimates rather than waiting for outcomes.
Editor's noteA logic for default reasoning
Formalized default reasoning for drawing revisable conclusions from incomplete knowledge.
Editor's noteNearest Neighbor Pattern Classification
Established foundational error bounds for nearest-neighbour classification and nonparametric pattern recognition.
Editor's noteA Theory of the Learnable
Founded computational learning theory through the probably approximately correct learning framework.
Editor's noteClassic quotaSome Studies in Machine Learning Using the Game of Checkers
Demonstrated self-improving game-playing programs and early learning from experience.
Editor's noteClassic quotaA machine program for theorem-proving
Established DPLL-style search, underpinning modern SAT solving.
Editor's noteClassic quotaNeural networks and physical systems with emergent collective computational abilities
Introduced associative-memory neural networks grounded in energy minimization and statistical physics.
Editor's noteClassic quotaInduction of decision trees
Popularized ID3 and information-gain splitting, establishing decision-tree induction as a mainstream learning approach.
Editor's noteClassic quotaStochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images
Popularized Gibbs sampling and Markov random fields for Bayesian image restoration.
Editor's noteClassic quotaA Computational Approach to Edge Detection
Defined optimal edge detection principles and an algorithm still widely used today.
Editor's noteClassic quota
Data sources
Catalogue data from OpenAlex (CC0), Crossref (CC0 metadata; no abstracts are shown), arXiv, and Papers With Code (CC BY-SA 4.0), and 
Semantic Scholar (used under its non-commercial API licence). Each count names its source. See the method.