CGM-JEPA self-supervised pretraining for consistent continuous glucose monitoring representations across multiple modalities.
Interpretable reinforcement learning model using state history and global feedback, evaluated on Atari environments.
Fisher subspace initialization method for LoRA fine-tuning of LLMs, improving adaptation performance through informed subspace selection.
LEAP enables efficient transformer inference by addressing incompatibility between layer-wise distillation and convergence-based early exit mechanisms.
GEODE framework for out-of-distribution detection with universal scorer compatibility, improving on Outlier Exposure method.
Bayesian symbolic regression using deep variational inference to generate probability distributions over interpretable expressions quantifying uncertainty.
RL approach using Infoprop Dyna for fast robot learning without physics simulators achieving sim-to-real transfer on wheeled robot.
DUET framework combining capable and lightweight models for collaborative two-stage inference reducing computational cost.
Continual reinforcement learning testbed (Forager) for studying plasticity and partial observability in non-stationary environments.
Approach using multi-perspective transformers and test-time training to solve ARC-AGI-2 visual reasoning puzzles achieving 96.1% training accuracy.
Method for controlling LLM behavior through activation steering while minimizing unintended changes to non-target feature directions.
Non-asymptotic generalization theory for deep learning explaining how neural tangent kernels partition output space into signal and noise channels.
Analysis of attention learning dynamics in transformers identifying focus-dilution cycles during training via gradient-flow analysis.
Identifies and analyzes attention sinks in multilingual neural machine translation where non-content tokens capture 83-91% of cross-attention in NLLB-200 model.
Proposes CombinationTS modular framework for time-series forecasting that decomposes models into components to identify true performance drivers.
Introduces Rhamba framework combining region-aware masking with hybrid Attention-Mamba architectures for self-supervised pretraining on resting-state fMRI data.
Proposes computationally efficient actor-critic algorithm for reinforcement learning with low-rank MDPs, establishing hierarchy of function approximation approaches.
Introduces S³-R1 using synthetic data and RL post-training for models to perform step-by-step retrieval and reasoning for question-answering with tool-use.
Develops theoretical framework and efficient algorithm for activation compression in LLM training, reducing memory requirements during backpropagation.
Proposes dynamic axonal delay mechanism for spiking neural networks using congestion-awareness to improve spike alignment in event-driven tasks.
Presents unified perspective on autonomous drift learning in non-stationary data streams beyond temporal concept drift in increasingly complex learning systems.
Proposes GA-VisAgent, a multi-agent application for code generation and visualization in Geometric Algebra learning, lowering adoption barrier through automated tooling.
Introduces GraphSculptor for constructing efficient pre-training coresets in graph self-supervised learning by exploiting dataset redundancy without additional training signals.
Analyzes confounding bias in offline language model evaluation from logs and proposes using randomized experiments to estimate true model performance differences.
Develops robust parameter learning method for uncertain MDPs that captures dependencies in transition probabilities from shared latent quantities.
Proposes model-based approach for learning safe offline policies with limited violation data in constraint-satisfying scenarios without risky online interaction.
Introduces PACE method for unsupervised environment design in reinforcement learning using parameter change as reliable environment evaluation signal.
Develops closed-form geometric characterization of regret gradients for Decision-Focused Learning, improving computational efficiency over existing surrogate loss methods.
Proposes physics-informed transformer architecture for foundational thermal modeling of buildings generalizable across diverse structures without building-specific calibration.
Studies sequential learning and catastrophic forgetting in differentiable resistor networks governed by physical constraints, analyzing gradient-based parameter adjustment.
Investigates whether carefully tuned baseline GNNs can match specialized methods for multi-label node classification, questioning the necessity of complex label-aware designs.
Theoretical generalization analysis for multimodal metric learning addressing how modality selection affects performance under incomplete data.
Theoretical analysis of counterfactual credit attribution for autoregressive generative models, formalizing credit to training data dependencies.
Benchmark evaluating LLM capabilities on graph property estimation tasks using random walk representations to test reasoning on structured data.
ProMORNA uses multi-objective RL and encoder-decoder models to generate full-length mRNA sequences from protein targets, balancing stability and translation.
Graph neural networks and transformers as ML surrogates for CFD simulations with improved training paradigms beyond node-wise supervision.
Hybrid quantum-classical reinforcement learning approach using QAOA for vehicle routing optimization, a combinatorial NP-hard problem.
Thesis on model merging paradigm for combining independently trained neural networks in weight space without optimization or original training data.
Analysis of transformer representations across languages, testing whether spectral properties enable cross-lingual concept transport.
Method for selecting singular-vector bases in low-rank LLM decomposition using importance-guided approach aligned with loss landscape geometry.
One-shot transfer learning method for Physics-Informed Neural Networks using Chebyshev augmentation to avoid retraining for different parameters.
Research on scaling laws for data-constrained training, moving beyond Chinchilla assumptions to optimize pretraining with limited high-quality data.
Adaptive Pluralistic Alignment proposes a pipeline to update aligned AI systems to track evolving values without retraining or large-scale data collection.
Paper applies law-and-economics deterrence models to AI alignment, treating misconduct in agentic systems as strategic responses to incentives.
Foundation model embeddings for geospatial analysis improve population estimation from satellite imagery in regions lacking census data.
Flow-Anchored Noise-conditioned Q-Learning algorithm for efficient offline reinforcement learning with expressive policies.
Large-scale benchmarking of AI-based molecular docking tools (DiffDock, AutoDock-GPU, GNINA) on LIT-PCBA library with 15 targets.
Method for surgically removing internal memorization traces from unlearned LLMs using cross-sequence probe alignment without capability loss.
Pareto set learning method that solves multiple multi-objective optimization tasks simultaneously using cross-task correlation.
Algorithm for linear dueling bandits under delayed feedback and adversarial corruption with learned context prediction.