A curated map of papers I think are worth reading across AI4Math, AI4Science, world models, machine learning theory, and reinforcement learning.

2025 / Formal theorem proving

DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition

A major Lean 4 prover built around recursive subgoal decomposition and reinforcement learning.

Paper
2025 / Formal theorem proving

Goedel-Prover: A Frontier Model for Open-Source Automated Theorem Proving

Open-source prover training with synthetic formal statements and iterative proof generation.

Paper
2025 / Geometry

Gold-Medalist Performance in Solving Olympiad Geometry with AlphaGeometry2

A strong example of language models, symbolic engines, and search combining for Olympiad geometry.

Paper
2024 / Formal theorem proving

DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search

Proof-assistant feedback plus tree search for Lean theorem proving.

Paper
2024 / Lean tools

Large Language Models as Copilots for Theorem Proving in Lean

Lean Copilot integrates LLM inference directly into Lean workflows for proof assistance.

Paper
2023 / NeurIPS

LeanDojo: Theorem Proving with Retrieval-Augmented Language Models

A benchmark and retrieval-augmented environment that helped standardize neural theorem proving in Lean.

Paper
2023 / Models for mathematics

Llemma: An Open Language Model for Mathematics

A math-specialized language model line that became a useful baseline for mathematical reasoning.

Paper
2023 / Benchmark

ProofNet: Autoformalizing and Formally Proving Undergraduate-Level Mathematics

A benchmark for natural-language-to-Lean formalization and undergraduate theorem proving.

Paper
2021 / Benchmark

miniF2F: A Cross-System Benchmark for Formal Olympiad-Level Mathematics

A small but influential benchmark for comparing formal math systems across proof assistants.

Paper
2025 / Nature

A Foundation Model for the Earth System

Aurora shows how one model can transfer across weather, air quality, waves, and other Earth-system tasks.

Paper
2025 / Nature

A Generative Model for Inorganic Materials Design

MatterGen uses diffusion-style generation for controllable discovery of stable inorganic materials.

Paper
2025 / Atomistic simulation

Orb-v3: Atomistic Simulation at Scale

A scalable universal interatomic potential focused on accuracy, latency, memory, and larger atomistic systems.

Paper
2024 / Nature

Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3

A landmark structure model for proteins, nucleic acids, small molecules, ions, and modified residues.

Paper
2024 / Nature

Probabilistic Weather Forecasting with Machine Learning

GenCast is a diffusion-based ensemble weather model for calibrated medium-range forecasts.

Paper
2024 / Materials

MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures

A broadly trained atomistic model for efficient simulation and property prediction under realistic conditions.

Paper
2023 / Nature

Scaling Deep Learning for Materials Discovery

GNoME pushed graph-network based materials discovery to a very large design space.

Paper
2023 / Science

Learning Skillful Medium-Range Global Weather Forecasting

GraphCast is a reference point for fast, data-driven global forecasting at operational scales.

Paper
2023 / Scientific foundation models

Towards Foundation Models for Scientific Machine Learning

A useful paper for thinking about pretraining, transfer, and scaling for PDE-centered AI4Science.

Paper
2023 / JMLR

Neural Operator: Learning Maps Between Function Spaces

The conceptual base for learning operators rather than finite-dimensional vector maps.

Paper
2021 / ICLR

Fourier Neural Operator for Parametric Partial Differential Equations

A core architecture for PDE surrogate modeling using spectral convolution.

Paper
2024 / Generative environments

Genie: Generative Interactive Environments

A generative world model that turns internet videos into controllable interactive environments.

Paper
2023 / Dreamer

Mastering Diverse Domains through World Models

DreamerV3 is a strong general-purpose model-based RL baseline across many domains.

Paper
2023 / Model-based control

TD-MPC2: Scalable, Robust World Models for Continuous Control

A practical world-model control stack for scalable continuous-control learning.

Paper
2022 / Transformers

Transformers are Sample-Efficient World Models

IRIS uses a transformer world model for sample-efficient Atari learning.

Paper
2020 / Nature

Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model

MuZero remains one of the cleanest examples of planning with a learned model.

Paper
2018 / Classic

World Models

The compact latent-dynamics paper that gave the area its modern name and aesthetic.

Paper
2023 / Representation learning

Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture

I-JEPA is a clean predictive-representation paper, useful for thinking about non-generative world modeling too.

Paper
2020 / Optimization

Sharpness-Aware Minimization for Efficiently Improving Generalization

A clean optimization lens on flat minima and generalization.

Paper
2019 / PNAS

Reconciling Modern Machine-Learning Practice and the Classical Bias-Variance Trade-Off

The double-descent paper that reframed overparameterized generalization.

Paper
2017 / ICLR

Understanding Deep Learning Requires Rethinking Generalization

A classic stress test for simple explanations of why deep networks generalize.

Paper
2023 / Robotics

RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

A key paper for grounding web-scale vision-language representations in robot actions.

Paper
2021 / Sequence modeling

Decision Transformer: Reinforcement Learning via Sequence Modeling

A simple and influential framing of offline RL as conditional sequence modeling.

Paper
2020 / Textbook

Bandit Algorithms

A precise modern treatment of bandits, regret, optimism, and exploration.

Book
2018 / Textbook

Reinforcement Learning: An Introduction

Sutton and Barto remain the backbone reference for RL concepts.

Book