Develops and analyzes algorithms and guarantees: convex and non-convex optimization, stochastic gradient methods and adaptive optimizers, convergence rates, min-max and bilevel optimization, statistical learning theory, PAC/online learning, kernel methods, high-dimensional statistics, conformal prediction and uncertainty quantification, Bayesian inference, and causal inference.
Includes generation and editing of images and video, text-to-image and text-to-video models, video understanding and long-video modeling, 3D representation and generation (Gaussian splatting, NeRF, meshes, point clouds), scene reconstruction, pose estimation, and visual world simulators.
Covers the construction of new benchmarks, datasets, and evaluation protocols for assessing model capabilities, reliability, and safety across domains, as well as meta-evaluation of evaluation itself: benchmark contamination, LLM-as-a-judge, leaderboards, and measurement validity.
Covers how LLMs are turned into reasoners and aligned after pre-training: reinforcement learning with verifiable rewards (GRPO/RLVR), RLHF and DPO-style preference optimization, reward and process reward models, chain-of-thought and long thinking, test-time scaling and search, efficient reasoning, and the theory of post-training dynamics.
Covers the theory and practice of modern generative models: diffusion and score-based models, flow matching and stochastic interpolants, consistency and few-step distillation, discrete and masked diffusion (including diffusion language models), guidance and controllable sampling, and reward-driven fine-tuning of generative models.
Applies machine learning to scientific discovery and modeling: protein and molecular design, drug discovery, genomics and single-cell biology, materials and chemistry, physics-informed learning and PDE solving, neural operators, climate and weather, neuroscience and brain signals, and automated scientific research.
Encompasses core RL and decision-making research: online and offline RL, model-based RL, exploration, multi-agent RL and game theory, bandits, policy gradient and actor-critic methods, safe/constrained RL, inverse RL, and theory of RL and control.
Focuses on models that jointly process vision, language, audio, and other modalities: large multimodal/vision-language models, visual reasoning and perception, multimodal alignment and hallucination, audio and speech models, omni-modal and unified understanding-generation models.
Focuses on LLM-based agents that plan, use tools, and act over long horizons: tool calling, multi-agent collaboration, agent memory and experience, GUI/web/software-engineering agents, deep research agents, and agentic RL training.
Includes techniques for making large models cheaper to train and serve: quantization, pruning and sparsity, KV-cache compression, sparse/linear attention, speculative decoding, mixture-of-experts, state-space models and other sub-quadratic architectures, long-context handling, and distributed/efficient training systems.
Advances the theory and practice of learning on graphs and structured data: graph neural networks and graph transformers, expressivity and oversmoothing, knowledge graphs, graph generation, combinatorial and relational reasoning, hypergraphs, geometric and topological deep learning, and learning on manifolds and non-Euclidean spaces.
Focuses on learning useful representations and the foundations of neural network design: self-supervised and contrastive learning, masked modeling, disentanglement and identifiability, multi-view and cross-modal representation alignment, embedding geometry, neural network architectures and initialization, normalization, and the theory of representation.
Addresses trustworthy ML beyond LLM-specific safety: adversarial robustness and certified defenses, backdoor and poisoning attacks, out-of-distribution detection and generalization, differential privacy, federated and decentralized learning, machine unlearning, fairness and bias mitigation, and model/data auditing.
Focuses on agents acting in physical or simulated environments: vision-language-action models, robot manipulation and navigation, imitation learning and policy learning from demonstrations, world models and latent dynamics for planning, sim-to-real transfer, autonomous driving, and humanoid/legged control.
Aims to understand what models learn and how: mechanistic interpretability, sparse autoencoders and feature discovery, circuits and probing, attribution and explanation methods, in-context learning theory, training dynamics, grokking, emergence, memorization vs. generalization, and the inner workings of transformers.
Focuses on modeling temporally ordered data: time-series forecasting and foundation models, anomaly detection, irregularly sampled and multivariate series, spatio-temporal modeling, dynamical systems and sequence modeling for sensors, finance, energy, traffic, and healthcare signals.
Covers the data side of model building: pre-training corpora and data mixtures, data selection, filtering and pruning, synthetic data generation and model collapse, data augmentation, curriculum and active learning, scaling laws for data and compute, tokenization and vocabularies, and the economics of data and compute.
Addresses risks of LLMs and foundation models: jailbreaks and prompt injection, safety alignment and refusal, red teaming and guardrails, hallucination and factuality, machine unlearning, watermarking, privacy and memorization, and misuse detection.
Examines the societal dimensions of machine learning and the field's own practices: AI governance and regulation, economic and labor impacts, human-AI collaboration and user well-being, sociotechnical risk, participatory and democratic uses of AI, sustainability, and position papers arguing for changes in research norms, peer review, and evaluation standards.