Essential AI Reading Queue
This page is a placeholder queue for papers that may deserve future notes. Keep entries short until a paper becomes important enough for a dedicated note.
Architecture
| Topic | Candidate Papers | Route |
|---|---|---|
| Transformer and attention | Attention Is All You Need, BERT, GPT-style decoder-only models, Hopfield Networks is All You Need | Architecture papers |
| Structured outputs and routing | Pointer Networks, Dynamic Routing Between Capsules, Set Transformer, DETR | Architecture papers |
| Memory and recurrence | LSTM, Neural Turing Machines, Memory Networks, Differentiable Neural Computer, modern Hopfield networks, Transformer-XL | Architecture papers |
| Vision and convolution | AlexNet, ResNet, U-Net, Vision Transformer | Architecture papers |
| Graph and geometry | GCN, GAT, SchNet, EGNN, SE(3)-Transformer, AlphaFold-style structure modules | Architecture papers, Computational Biology papers |
| Sequence alternatives | S4, Mamba, linear attention variants | Architecture papers |
| Conditional compute | Mixture of Experts, Switch Transformer, routed sparse models | Architecture papers, Systems papers |
| Agent and control architectures | Deep Q-Network, AlphaZero, World Models, Dreamer, MuZero, model-based control, value-based control, tree search, latent imagination | Architecture papers, Agents |
| Iterative and adaptive computation | Universal Transformers, recurrent depth, early exit, adaptive halting | Architecture papers, Adaptive computation |
| Shared-weight representations | Siamese Networks, FaceNet, Siamese/triplet embeddings, CLIP-style dual encoders | Architecture papers, Learning method papers |
| Architecture design automation | Neural Architecture Search with Reinforcement Learning, EfficientNet, search-space and budget design | Architecture papers |
Learning Methods
| Topic | Candidate Papers | Route |
|---|---|---|
| Self-supervised learning | SimCLR, MoCo, BYOL, MAE, DINO, JEPA-style objectives | Learning method papers |
| Language-model pretraining | GPT, BERT, T5, instruction tuning, RLHF-style alignment | Learning method papers |
| Representation learning | contrastive learning, masked modeling, metric learning, probing papers | Learning method papers |
Generative Models
| Topic | Candidate Papers | Route |
|---|---|---|
| Diffusion and score models | DDPM, score-based generative modeling, classifier-free guidance | Generative model papers |
| Flow and likelihood models | Normalizing flows, CNF, flow matching, rectified flow | Generative model papers |
| Structured generation | molecular generation, protein design, structure generation | Generative model papers, Computational Biology papers |
Systems and Evaluation
| Topic | Candidate Papers | Route |
|---|---|---|
| Scaling and efficiency | scaling laws, Chinchilla-style compute/data tradeoff, efficient inference | Systems papers |
| Evaluation and reproducibility | benchmark design, leakage audits, artifact papers | Paper reproducibility, Paper analysis |
| Agents and tools | tool use, ReAct-style reasoning, retrieval, agent evaluation | Agents |
Computational Biology
| Topic | Candidate Papers | Route |
|---|---|---|
| Protein structure and design | AlphaFold family, RFdiffusion, protein language models | Computational Biology papers, Protein modeling papers |
| Structure-based modeling | DiffDock-style docking, PoseBusters-style evaluation, docking benchmarks | Structure-based modeling papers |
| Molecular modeling | molecular property prediction, molecular generation, assay-aware evaluation | Computational Biology papers |
Promotion Rule
Create a dedicated paper note only when at least one of these is true:
- The paper defines a concept used across several wiki pages.
- The paper is needed to interpret a current project or research direction.
- The paper is a baseline or evaluation reference for computational biology work.
- The paper changes how architecture, learning method, generation, systems, or evaluation should be explained.