🏢 University of Toronto

Contrastive learning is reframed as a distribution alignment problem, leading to a flexible framework (GCA) that improves representation learning with unbalanced optimal transport.

Artificial Dopamine (AD) algorithm achieves comparable performance to backpropagation methods in complex RL tasks by using only synchronously distributed per-layer TD errors, demonstrating the suffici…

This paper simplifies constraint inference in reinforcement learning, demonstrating that standard inverse RL methods can effectively infer constraints from expert data, surpassing complex, previously …

This paper introduces contextual Shtarkov sums, a new complexity measure characterizing minimax regret in sequential probability assignment with contexts, and derives the minimax optimal algorithm, co…

ExPerior leverages expert demonstrations to enhance online decision-making, even when experts use hidden contextual information unseen by the learner.

Curated synthetic data provably optimizes human preferences in iterative generative model training, maximizing expected reward while mitigating variance.

SCube: Instant large-scale 3D scene reconstruction from sparse images using VoxSplats, a novel 3D Gaussian splat representation.

Reaction-GFlowNet (RGFN) revolutionizes small molecule discovery by generating synthesizable molecules directly within the chemical reaction space, dramatically expanding the search space for drug dis…

Deep RL agents can now effectively learn complex tasks even with noisy, uncertain sensor readings by exploiting the structure of Reward Machines.

Random Cycle Coding (RCC) optimally compresses cluster assignments in large datasets, saving up to 70% storage in vector databases by eliminating the need for integer IDs.

Quantum Deep Equilibrium Models (QDEQs) achieve higher QML performance with shallower circuits by using a DEQ training paradigm, improving near-term quantum computation efficiency.

This paper introduces efficient algorithms that leverage social choice theory to aggregate multiple individual preferences, resulting in a desirable collective AI policy.

In n-player games, a satisficing path always exists leading from any initial strategy profile to a Nash equilibrium by allowing unsatisfied players to explore suboptimal strategies.

ERM’s efficiency in feature learning surprisingly remains high even with massive feature maps; its excess risk asymptotically matches an oracle procedure’s, implying potential for streamlined feature-…

Researchers predict language model performance by observing existing models, bypassing costly training, revealing surprising predictability in complex scaling phenomena.

NEUR2BILO: a neural network-based heuristic solves mixed-integer bilevel optimization problems extremely fast, achieving high-quality solutions for diverse applications.

A novel lossy compression framework, Minimum Entropy Coupling with Bottleneck (MEC-B), extends existing methods by integrating a bottleneck for controlled stochasticity, enhancing performance in scen…

Maia-2: A unified model for human-AI alignment in chess, coherently captures human play across skill levels, significantly improving AI-human alignment and paving the way for AI-guided teaching.

LLM Processes leverage LLMs to create probabilistic regression models guided by natural language, enabling seamless integration of expert knowledge and improving prediction accuracy.

Scaling causal language models reveals a connection between neural collapse properties, model size, and improved generalization, highlighting NC’s broader relevance to LLMs.