GAI-empowered intelligent transportation digital twin using UAVs with diffusion models for processing roadside sensor data.
Tree-of-Evidence inference algorithm for faithful multimodal model grounding with interpretable reasoning in healthcare and high-stakes domains.
CausalVAE as plug-in module for world models improving counterfactual dynamics prediction and robustness under distribution shift.
Theoretical analysis of single hidden-layer neural networks with ReLU, fixed biases, proving convergence and spectral bias properties.
MIPT-SSM sequence architecture using measurement-induced phase transitions to achieve O(1) inference cache for language models.
Agent-as-Annotators framework distilling web navigation capabilities from Gemini 3 Pro into smaller models via structured trajectory generation.
PolicyLong method for extending LLM context windows using on-policy data synthesis to align with model capabilities during training.
Information-theoretic framework for predicting task affinity in multi-task learning, addressing gradient-based task relationship estimation.
QaRL method for LLM RL training addressing training-inference mismatch by aligning quantized rollouts with learning updates.
Progressive quantization-aware training framework for ultra-low-bit LLMs using outlier channel splitting to stabilize convergence.
Neuro-inspired architecture using Kuramoto oscillatory phase encoding to improve learning efficiency by incorporating phase dynamics.
Theoretical analysis of doubly stochastic attention in Transformers, examining rank decay and signal degradation across layers.
Analysis of output-length prediction for efficient LLM serving, examining prompt-conditioned length distributions for batching and memory optimization.
Framework for tabular data disentanglement transforming complex attribute relationships into latent variables with reduced interdependencies.
ML-based fraud detection framework for banking transactions using PaySim synthetic dataset to address imbalanced classification.
Framework evaluating time series classification methods across performance and energy efficiency, exploring pruning and resource consumption trade-offs.
Analysis of weight compensation methods for LLM quantization, examining residual errors in techniques like GPTQ for reducing model precision.
Research clarifying distinction between machine unlearning and untraining—different approaches to removing data points or behaviors from trained models.
Deep learning benchmarking for liver remnant segmentation in surgical planning for colorectal metastases using medical imaging data.
Analysis of ML competition platforms (Kaggle, Zindi) examining workflows, evaluation methods, participant expertise, and impact on AI development.
Attack on Computer Use Agents targeting vision modality through attention concentration to redirect preferences.
Framework for personalized algorithmic recourse providing actionable recommendations with explicit user considerations.
RL method for partial observability using privileged planner guidance during training with MPC.
Framework automating aggregation strategy selection in federated learning across heterogeneous settings.
Pearl framework for multimodal reasoning using predictive embeddings to reduce tool-use overhead in VLMs.
Studies bias redistribution when vision models selectively unlearn demographic groups.
Efficient MoE inference through budget-aware expert activation allocation reducing latency bottlenecks.
Bandit algorithm for contextual decision-making with latent hidden Markov chain dynamics.
Flow-based method for offline multi-agent reinforcement learning using value guidance.
Recommender system using long-term embeddings to balance recency bias and stable user preferences.
Method for assessing model generalization in Vision Transformers via internal representations under distribution shift.
Examines vulnerabilities in machine unlearning methods by analyzing internal representations and concept reintroduction.
DMax enables efficient parallel decoding in diffusion language models through progressive self-refinement.
Architecture combining frozen LLMs as nodes communicating through learned projections in a shared latent space.
Continual learning method using complementary self-supervised embeddings to improve replay buffer sample selection.
Benchmark for egocentric video understanding in AR using long-context reasoning over temporal activities.
Bias-constrained diffusion models for PDE emulation with improved accuracy and training efficiency.
Data selection framework for autonomous driving models balancing multiple evaluation metrics.
Parameter-efficient fine-tuning compression framework reducing communication costs for model adaptation.
Self-supervised pre-training method for time series classification with adaptive input handling.
Data augmentation method using adversarial training for out-of-distribution generalization on graphs.
KV cache offloading technique to reduce memory and latency overhead for long-context LLM inference.
Hybrid post-training combining reinforcement learning and distillation to improve LLM confidence calibration.
Machine learning framework for estimating turbofan engine health from sensor data.
Test-time variational synthesis method for reinforcement learning in domains without verifiable rewards.
Impact of quantization on federated learning accuracy-efficiency trade-offs for aerospace predictive maintenance.
Analysis of how embedding dimensionality affects stability of graph node embeddings.
Mechanistic study of how steering vectors modify LLM behavior for alignment and refusal control.
Meta-learning approach for brain signal decoding without per-subject training.
Multi-agent system for language-agnostic code translation and validation across programming languages.