🏢 MIT

NeuralFluid: Design & control complex fluidic systems with dynamic boundaries using differentiable simulation, achieving superior results in benchmark tasks.

Model-Based Transfer Learning (MBTL) boosts deep RL sample efficiency by strategically selecting training tasks, achieving up to 50x improvement over traditional methods.

MeMo: a novel framework for pretraining meaningful, modular robot controllers via noise injection, enabling efficient transfer learning across different robot morphologies and tasks.

New metrics and p-annealing improve sparse autoencoder training for better language model interpretability.

Researchers establish basis-free conditions for SGD learnability in two-layer neural networks learning subspace-sparse polynomials with Gaussian input, offering insights into training dynamics.

Optimizing against learning agents: New algorithms and computational limits revealed!

LLaMAR: LM-based planner for multi-agent robots excels in long-horizon, partially observable tasks, achieving 30% higher success than existing methods.

DDiffPG: A novel actor-critic algorithm learns multimodal policies from scratch using diffusion models, enabling agents to master versatile behaviors in complex tasks.

Researchers developed an efficient algorithm to uniquely identify causal relationships from mixed interventional and observational data with noisy interventions.

Problem (FAIR) framework and FORM algorithm achieve flexible multi-stakeholder fairness in online recommendation systems, balancing platform revenue with user and item fairness.

Data Debiasing with Datamodels (D3M) efficiently improves machine learning model robustness by identifying and removing specific training examples that disproportionately harm minority groups’ accurac…

This paper provides identifiability guarantees for causal disentanglement from purely observational data using nonlinear additive Gaussian noise models, addressing a major challenge in causal represen…

Low reward variance drastically improves contextual bandit regret, defying minimax assumptions and highlighting the crucial role of eluder dimension.

Diffusion models surprisingly learn factorized representations, enabling compositional generalization, but struggle with interpolation; training with independent factors drastically improves data effi…

Hamiltonian Generative Flows (HGFs) revolutionize generative modeling by leveraging Hamiltonian dynamics, offering enhanced score matching and generative capabilities.

MDGEN: Generative modeling unlocks MD data for diverse tasks, achieving significant speedups via flexible multi-task surrogate models.

Brain’s flexible mapping of abstract domains is achieved via self-supervised extraction and projection of generalized velocity signals by grid cells, enabling efficient map generation.

Biologically-inspired DCnet neural network flexibly modulates visual processing based on context, outperforming existing models on visual search and attention tasks.

Few-shot task learning through inverse generative modeling (FTL-IGM) enables AI agents to quickly master new tasks from minimal data by leveraging invertible generative models.

Fair Wasserstein Coresets (FWC) efficiently generates fair, representative subsets of large datasets for downstream machine learning tasks, improving fairness and utility.