Proposes conjugate learning theory to characterize trainability and generalization in deep neural networks using convex duality.
EnterpriseGym Corecraft: high-fidelity RL environment simulating customer support with 2,500+ entities and 23 tools for training generalizable agents.
Multi-agent bandit framework for submodular welfare problem maximizing agent utilities under bandit feedback conditions.
Design concepts for memory systems to support artificial superintelligence, exploring extraction and storage paradigms without novel methods.
Convex gated probing method for faithfully evaluating audio SSL embeddings, improving transformer ranking on AudioSet benchmark.
Lightweight hierarchical transformer for efficient 3D medical image segmentation balancing accuracy with computational efficiency.
Parameter-efficient implicit neural representation architecture using learnable periodic activations inspired by subtractive synthesis.
STING benchmark for measuring multi-turn, multilingual LLM agent misuse over sequential steps with automated red-teaming.
CAFE framework combining causal discovery with multi-agent reinforcement learning for automated feature engineering on tabular data.
RoboGene framework using diversity-driven agentic task generation to maximize robotic manipulation training data for VLA pre-training.
Mechanistic analysis showing looped and depth-grown LLM architectures exhibit convergent depth-wise signatures, unifying reasoning approaches.
Online conformal prediction method for non-stationary data streams with unknown distribution drift using training-conditional cumulative regret.
Zero-shot classifier editing method enabling fine-grained video understanding by splitting coarse categories without retraining on new annotations.
Contrastive learning framework with attention-based feature adaptation for street-view image classification using vision-language models like CLIP.
Theoretical analysis of error propagation when recursively training diffusion models on synthetic data, showing performance degradation from distribution drift.
Novel explainability method for transformer models using context-aware layer-wise integrated gradients to interpret predictions by capturing inter-token dependencies.
Research on using LLMs as comparative evaluators with reliability weighting. Analyzes bias and inconsistency in LLM judgment aggregation.
Research on multilingual safety alignment for LLMs using resource-efficient methods. Addresses cross-lingual safety consistency.
Research on AI agent reliability evaluation beyond accuracy metrics. Analyzes consistency, perturbation robustness, and operational failures in deployed agents.
Research on object-centric representations for compositional generalization in VQA tasks. Evaluates visual reasoning capabilities.
Research comparing parameter-free representations to foundation models on single-cell RNA-seq downstream tasks. Biology-focused ML research.
Research on Spiking Neural Networks for efficient temporal information extraction with neuromorphic deployment. Addresses temporal resolution domain adaptation.
Analysis of Transformer optimization through gradient heterogeneity lens to explain Adam's superiority over SGD.
Research on continual learning with abundant memory, challenging traditional memory minimization constraints.
Research on lookback window bias in long-term time series forecasting benchmarks and model evaluation validity.
Research on emergent capabilities in language models via scaling. Analyzes breakthrough performance vs metric thresholding.
FedEFC method for federated learning with noisy labels using enhanced forward correction. Privacy-preserving distributed ML.
FedMerge: Federated learning approach creating personalized models per client by merging multiple global models.
ReaCritic: DRL method with transformer-based critic for heterogeneous network management and wireless optimization.
PLAICraft: Large-scale time-aligned Minecraft dataset with vision, speech, and action for embodied AI agents.
WINA: Training-free sparse activation method for LLM inference efficiency via weight-informed neuron selection.
XENON: LLM-based agent for Minecraft that algorithmically corrects knowledge from experience for robust long-horizon planning.
Theoretical study of expressive power of mixture-of-experts networks for tasks with low-dimensionality and sparsity.
DiffusionBlocks: Framework for block-wise transformer training via diffusion interpretation to reduce memory overhead.
Security research on neural network model extraction attacks via black-box queries on deep networks.
Research on step ordering in chain-of-thought reasoning for transformer arithmetic tasks. Studies impact on reasoning difficulty.
Benchmark study of stochastic approximation algorithms for fairness-constrained DNN training using Census data.
Causal discovery algorithm for time series robust to noise using power-law frequency spectra analysis.
Model-agnostic dynamic feature selection method with uncertainty quantification for resource-constrained decision-making.
Neural network output layer using orthogonal projection for satisfying convex constraints in feasible-by-design optimization.
Systematization of knowledge on data minimization principles in ML with focus on GDPR/CPRA regulatory compliance.
LLM-based framework for generating synthetic healthcare tabular data with fairness constraints from limited samples.
Method for single-pass uncertainty estimation in edge ML using next-activation prediction for microcontroller deployment.
Research characterizing universal activation sparsity properties in modern LLMs with implications for efficiency and interpretability.
Study on LLM self-improvement through reinforcement learning without external labels, addressing convergence toward majority-favored solutions.
Analysis of classifier-free guidance dynamics in diffusion models showing three-stage sampling process under multimodal conditions.
Research on training re-evaluation curves to optimize data curriculum ordering during LLM training for improved model retention.
Comparison of RNN architectures vs modern models for irregular time series prediction in healthcare/sensor domains.
Safety mechanism for RL agents balancing exploration and constraint satisfaction using uncertainty-aware modulation. Relevant to agent training.
Theoretical and empirical analysis of Transformers learning graph algorithms with proper training data. Explains generalization failures.