Theoretical analysis of finite-sample properties and identifiability bounds for nonlinear Independent Component Analysis algorithms.
Demonstrates power-law scaling of classification error with number of classes and how chain-of-thought decomposition reduces error through task splitting.
Practical analysis of chain-of-thought distillation from students to teachers, revisiting capacity gap assumptions and baseline comparisons.
Conformal prediction framework for transformers providing uncertainty quantification and calibration for trustworthy LLM deployment.
Analysis of causal inference applications in graph representation learning and risks of aggregating graph elements.
Adaptive Thompson sampling for high-dimensional Bayesian optimization addressing sparse candidate point grids.
Dynamic policy optimization bridging SFT and RL for LLMs, addressing bias-variance tradeoff in post-training through adaptive loss weighting.
Empirical study on effectiveness of advanced optimizers for multi-task learning, identifying overlooked factors in optimization approaches.
Analyzes calibration and paraphrase sensitivity in medical vision-language models using predictive entropy and uncertainty quantification.
SeqComm-DFL: Multi-agent coordination via sequential communication and decision-focused learning for value-aware message generation.
WOMBET: World model-based framework for experience transfer in robotics RL, generating and utilizing prior data for sample efficiency.
Hierarchical implicit flow Q-learning for offline goal-conditioned reinforcement learning with improved policy expressiveness.
SentryFuse: Framework for efficient multimodal model compression on edge devices with sensor dropout robustness via zero-shot pruning.
Graph neural network architecture addressing heterophilic graphs using switchable attention mechanism for monophily-aware learning.
Solution to NeurIPS 2023 LLM Efficiency Challenge: Fine-tuning LLaMA 70B on single A100 GPU within 24-hour constraint.
U-Cast: Efficient probabilistic weather forecasting model using standard U-Net architecture, simplifying state-of-art approaches.
PDYffusion: Diffusion model for long-horizon spatiotemporal prediction incorporating physics-based constraints and uncertainty quantification.
Proposes PML-MA method for partial multi-label learning using feature-label modal alignment to handle noisy labels.
arXiv paper proposing Temporal Patch Shuffle data augmentation for time series forecasting preserving temporal coherence and improving generalization.
arXiv paper integrating graph-based embeddings into event sequence models for user-item interactions in fraud and recommendation systems.
arXiv paper on GeoPAS geometric probing approach for automated algorithm selection in continuous black-box optimization.
arXiv paper on EquiformerV3, advancing SE(3)-equivariant graph attention Transformers for efficiency, expressivity, and 3D atomistic modeling.
arXiv paper on CORA, conformal risk-controlled GUI agents using vision language models with formal safety guarantees for mobile automation.
arXiv paper proposing truncated rectified flow policy for maximum entropy RL enabling one-step multimodal action distribution sampling.
arXiv paper augmenting distillation process dataset with simulations for deep learning-based anomaly detection in chemical batch processes.
arXiv paper developing generalization and scaling theory for Mixture-of-Experts Transformers with covering-number bounds and routing overhead analysis.
arXiv paper on GNN-based deep reinforcement learning scheduler for cloud workflow DAGs optimizing completion time and energy consumption.
arXiv paper proposing DiffHLS framework using GNNs and LLM code embeddings for high-level synthesis quality prediction via differential learning.
arXiv paper investigating LLM pretraining geometry and common minima to improve downstream generalization without changing loss function.
arXiv paper on distributed online convex optimization with compressed communication, establishing optimal regret bounds for large-scale applications.
Novel stability-enhanced Gaussian process VAE for training low-dimensional LTI systems from high-dimensional video data using probabilistic and physical models.
Controlled study of dataset scaling laws in attention-only decoder architecture across power-of-two subset sizes.
Machine unlearning approach using relearning convergence delay metric to remove contaminated data from pretrained models.
Online activation subspace learning (OASIS) to reduce memory requirements during LLM training through low-rank projections.
Scalable method for generating node embeddings on massive distributed graphs with millions to billions of nodes.
AdaCubic optimizer using adaptive cubic regularization with Hutchinson's method for approximating Hessian in deep learning.
One-step diffusion model for efficient chest X-ray report generation reducing inference latency compared to autoregressive models.
Method for safely updating deep reinforcement learning policies while preserving safety guarantees on previously encountered tasks.
Hardware optimization using electro-optic nonlinearities to replace softmax bottleneck in transformer attention mechanisms.
High-fidelity cyber operations simulator (NetForge_RL) using temporal graph networks for multi-agent reinforcement learning in cybersecurity.
OmniBehavior benchmark for evaluating LLMs as user simulators on long-horizon, cross-scenario behavior traces from real-world data.
Investigation of self-sovereign AI agents that can economically sustain themselves without human involvement using LLMs and agent frameworks.
Analysis of how bias mitigation reshapes embedding spaces in BERT and Llama2 through representational analysis of gender-occupation associations.
Systematic evaluation of chain-of-thought vs zero-shot prompting across temperature settings using Grok-4.1 for extended reasoning LLMs.
Research on attention-based sampling for diffusion language models enabling parallel decoding instead of sequential auto-regressive approach.
Tree-structured sparse feed-forward layers as drop-in MLP replacements in transformers enabling conditional computation via routing.
Theoretical framework for reward fine-tuning of diffusion models using stochastic optimal control and adjoint matching.
Protocol governing autonomous agent mutations with execution-bound safety checks and evidence chains for API-centric architectures.
Open-source browser extension for AI-assisted title and abstract screening in literature review with no-code, serverless architecture.
Attack method on LLM orchestration systems where single requests decompose into benign subtasks that jointly violate security policies.