Models, methods and papers mentioned

Consolidated list of models, techniques, and papers referenced in the DL4SCI 2026 lectures. Links are added only where verified; others are marked (look up).

Day 1

Generative AI for Science — Arash Vahdat (Nvidia)

Molecules

• SMILES — string notation for molecules (look up).
• SAFE — Sequential Attachment-based Fragment Embedding (fragment-based strings) (look up).
• SAFE-GPT — autoregressive fragment model (look up).
• GenMol — fragment-based discrete diffusion model (look up).
• ReaSyn (Reaction Synthesis) — encoder–decoder transformer predicting synthetic pathway (look up).

Proteins

• La-Proteina — fully-atomistic protein design (encoder–decoder) (look up).
• Complexa — protein binder design (look up).

Concepts

• Reward-tilted / RL-as-inference distribution $q^*(x)\propto p_\theta(x)\exp(R(x;c)/\lambda)$.
• Three paradigms: generative training, post-training alignment w/ RL, test-time search.

Accelerated Diffusion Models — Julius Berner (Nvidia)

• Flow matching — interpolation + velocity field.
• DiT (Diffusion Transformer) with adaLN time-conditioning (look up).
• Latent diffusion (compress to latent, e.g. VAE / representation encoder).
• Rectified flow / Reflow — straighten trajectories (look up).
• Optimal transport coupling (mini-batch OT).
• FlashAttention — hardware-optimized attention (look up).
• FLUX.2-dev — image model used as inference-speedup case study (look up).
• State-space models (SSMs) — sub-quadratic sequence models (look up).
• Progressive distillation — learn large jumps from smaller ones (look up).
• Consistency models / Eulerian flow maps — self-consistency along trajectory (look up).
• Lagrangian flow maps — differentiate flow map w.r.t. target time $t$.
• Pi-Flow (π-Flow) — parallel decoding distillation (look up).
• Adversarial distillation — Projected GANs on teacher features, diffusion init (look up).
• Variational Score Distillation (VSD) — divergence via teacher score (look up).
• Schrödinger bridges / stochastic interpolants (look up).

Transformers, Scaling and Hardware — Jacob Austin (Anthropic)

Large Physics Models (Science Application) — Dar Mehta (Casual Labs)

• Themes: counterfactual, credit assignment, generalization.
• LPM for weather (recover physical states from raw noisy inputs; causal structure to predict/control).

Atomic AI Foundation Models (Science Application) — Alex Morehead (Berkeley Lab)

• Ways to build FMs: autoregressive vs bidirectional.
• Zatorm-1 — atomic AI foundation architecture (name to verify against slide).

Day 2

Discrete Diffusion — Ferdinando Fioretto (University of Virginia)

• Constraint-based / constrained diffusion — sampling as sequential optimization.
• Augmented Lagrangian sampling — for non-convex constraints / residual-only (verifiers).
• CADM — constraint-aware diffusion model (protein design) (look up).
• RFdiffusion — protein-design baseline (look up).
• Applications: microstructure/porosity, physics-informed motion, safe multi-robot autonomy, protein design, inverse design of mechanical metamaterials.

Day 3

Introduction to RL — Elynn Chen (NYU)

Classic RL

• Q-learning — off-policy TD control (Watkins, look up).
• DQN — deep Q-network (replay buffer, target network); Mnih et al. 2015 (look up).
• Policy gradient / REINFORCE — Williams 1992; actor–critic (look up).
• TRPO — Trust Region Policy Optimization; Schulman et al. 2015 (look up).
• PPO — Proximal Policy Optimization; Schulman et al. 2017, arXiv:1707.06347.
• GAE — Generalized Advantage Estimation (look up).

LLM post-training

• RLHF — RL from human feedback; Christiano et al. 2017 (preferences), InstructGPT Ouyang et al. 2022, arXiv:2203.02155.
• Bradley–Terry — pairwise preference model used to fit the reward model (look up).
• DPO — Direct Preference Optimization; Rafailov et al. 2023, arXiv:2305.18290.
• RLVR — RL with verifiable (rule-based) rewards; popularized by Tulu 3 (AI2) (look up).
• GRPO — Group Relative Policy Optimization; DeepSeekMath Shao et al. 2024, arXiv:2402.03300.
• DeepSeek-R1 / R1-Zero — RLVR-trained reasoning models; arXiv:2501.12948.

Reasoning — Aditya Grover (UCLA)

Understanding the Visual World Through Physical Intrinsics — Joseph Redmon (Ai2)

Day 4

Background only: these items come from the Day 4 background primers, not from the lectures (no live notes for this day).

The Rise and Ramifications of Artificial Mathematical Intelligence — Tony Feng (UC Berkeley / Google DeepMind)

• Lean 4 — interactive theorem prover / proof assistant (look up).
• Mathlib — community library of formalized mathematics for Lean (look up).
• Autoformalization — natural-language math → formal (Lean) statements/proofs (look up).
• AlphaGeometry — DeepMind; neuro-symbolic olympiad geometry, synthetic data (look up).
• AlphaProof — DeepMind; RL over formal Lean statements (look up).

Scaling 4D World Models for Scientific AI — Peter Yichen Chen (University of British Columbia)

• NeRF — Neural Radiance Fields; Mildenhall et al. 2020 (look up).
• 3D Gaussian Splatting — Kerbl et al. 2023 (look up).
• Differentiable simulation — gradients through physics simulators (look up).
• MPM / FEM — material point method / finite element method (look up).
• PINNs — physics-informed neural networks (look up).
• Neural operators / learned graph-network simulators (look up).

Reasoning / Time series / Tabular Data — Bernie Wang (MIT)

• DeepAR — probabilistic autoregressive RNN forecaster (look up).
• Chronos — time-series foundation model (quantized tokens, T5-style LM) (look up).
• TimesFM, Moirai, Lag-Llama — time-series foundation models (look up).
• TabPFN — prior-fitted transformer for tabular in-context prediction (look up).
• XGBoost / LightGBM / CatBoost — gradient-boosted trees baselines (look up).

World Models — Jiajun Wu (Stanford University)

• World Models — Ha & Schmidhuber 2018 (look up).
• Dreamer — Hafner et al.; latent imagination for model-based RL (look up).
• MuZero — DeepMind; planning with a learned model (look up).
• Genie — DeepMind; action-controllable generated environments (look up).

Day 5

Trustworthy AI Agents for Science — Sheng Liu (Stanford University)

• GPT-RadPlan — "Automated radiotherapy treatment planning guided by GPT-4Vision"; evaluator agent (vision + DVH experts, toolboxes, aggregation) ↔ planner agent (memory → new objective weights); framed as bi-level meta-optimization. Best in Medical Physics, AAPM 2024. https://iopscience.iop.org/article/10.1088/1361-6560/adf02c
• DVH — dose-volume histogram (plan-quality summary) (look up).
• Cerebra — "A Multidisciplinary AI Board for Multimodal Dementia Characterization and Risk Assessment"; https://cerebra-health.com.
• TextGrad — Python library: natural-language feedback as an optimization signal; LLM as optimizer (backprop analogy). Zou group, Stanford. https://github.com/zou-group/textgrad.
• Toolbox tools mentioned: MedSAM2 (medical-image segmentation), FreeSurfer (brain-MRI processing), PyTorch, scikit-learn (look up).

Note: papers specific to my own longitudinal-brain research (BrLP, Temporal Flow Matching, Anatomically Guided Latent Diffusion, single-cell trajectory inference, etc.) are kept in my private ideas/ notes, not here.

These are my own notes and may contain errors. Not official course material. Source on GitHub.