Adaptive Prompt Structure Factorization: API-only framework using architect model to decompose and optimize compositional prompt programs for LLMs.
Studies multimodal LLM hallucinations, distinguishing obvious from elusive types; proposes steering hallucination verifiability.
CASE: recommendation system using cadence-aware encoding for next-basket repurchase prediction in retail.
CBM-Dual: silicon processor implementing chaotic Boltzmann machines for simulated annealing and reservoir computing at edge.
FedDetox: federated learning framework for small language model alignment with on-device data sanitization against toxic/poisoned data.
Uses inductive logic programming (ILASP) to approximate neural networks for preference learning; creates dataset of recipe preferences.
Energy-Regularized Spatial Masking (ERSM): feature selection framework improving robustness and interpretability of vision models via energy regularization.
Empirical study of how automotive industry practitioners perceive, detect, and manage data leakage between training and evaluation datasets.
NestPipe: decentralized embedding training framework for trillion-parameter recommendation models at 1,500+ accelerator scale with nested pipelining.
ELC: evidential lifelong classifier combining uncertainty quantification with continual learning for radar pulse classification with confidence estimation.
CAFP: post-processing fairness framework using counterfactual model averaging for group fairness without requiring access to model internals.
QNAS: neural architecture search framework for designing accurate and resource-efficient quantum neural networks on NISQ hardware.
ReDAct: uncertainty-aware deferral mechanism for LLM agents to mitigate hallucination errors in sequential decision-making by deferring to humans when uncertain.
Data-driven controller synthesis for nonlinear systems using Structured State-space Models as surrogates for long-term time-series dependencies.
EVGeoQA benchmark evaluates LLMs on dynamic multi-objective geo-spatial exploration with compound constraints and varying user locations.
DDP-SA: federated learning framework combining local differential privacy and secure aggregation for privacy-preserving distributed ML training.
Transformer-based network for multi-modal vehicle trajectory prediction without explicit graph structure or intention labels.
AegirJAX: fully differentiable hydrodynamic solver for coastal wave propagation and inverse problems using end-to-end differentiable computing.
Introduces Dynamic Context Evolution (DCE) to address cross-batch mode collapse in LLMs, where repeated prompting causes output repetitiveness. Proposes principled framework beyond ad hoc deduplication.
T-STAR: tree-structured reinforcement learning for multi-turn LLM agents with self-rectification and grafting for sparse reward optimization.
DINO-QPM: lightweight interpretability adapter converting DINOv2 visual foundation model features into human-interpretable representations.
Performance-energy trade-off analysis for real-time 3D Gaussian Splatting on edge GPUs with controlled computational budgets.
ATOM Report: comprehensive adoption analysis of ~1.5K open language models showing Chinese models surpassing US counterparts in 2025.
TraceSafe-Bench: first comprehensive benchmark evaluating LLM guardrails efficacy in multi-step tool-use agent trajectories.
Efficient learned data compression via dual-stream feature decoupling to balance probability modeling with system latency.
k-server-bench: automated mathematical discovery challenge for finding potential functions in the k-server conjecture.
Study on robustness of attribution heatmap explanations in deep networks trained to predict human image authenticity judgments.
Theory and practice of scalable Gaussian process regression using nearest neighbours for massive dataset applications.
Systematic study analyzing impact of retrieval pipeline components on RAG-based medical question answering systems with LLMs.
Integration of DeePMD-kit neural network potentials into GROMACS for multi-GPU accelerated molecular dynamics simulations.
CADENCE: adaptive depth estimation system for autonomous vehicles that dynamically scales neural network complexity for embedded processors.
AlignPrune: noise-robust dynamic data pruning method using loss trajectory alignment to preserve clean samples under label noise.
Convergence rate analysis for asynchronous Q-learning with polynomial stepsize under high-dimensional central limit theorem conditions.
Personalized RewardBench: novel benchmark for evaluating reward models' ability to capture individual user preferences in LLM alignment.
Methodology for measuring generative AI power consumption across data centers to address proprietary data gaps and infrastructure planning.
arXiv paper on motion-controlled video generation with disentangled control and motion causality for physical scene dynamics.
Elastic Test-Time Training method addressing catastrophic forgetting in long-context 3D reconstruction with plastic inference-time updates.
Comprehensive arXiv survey of generative AI covering LLM architectures, deployment protocols, and applications as of early 2026.
Framework for smooth optimization of explicitly regularized sparse objectives via Hadamard overparametrization, enabling gradient-descent compatible solvers.
Theoretical analysis showing single labels on more samples outperforms multiple labels per sample for binary classifier comparison under noisy label budgets.
Analytic federated learning (AFL) paradigm enabling closed-form solutions for federated learning with pre-trained models in single-round training.
SleepNet and DreamNet: deep learning models for visual classification via feature enrichment and reconstruction with pre-trained encoders.
DROP: distributional reinforcement learning framework with asymmetric learning rates modeling optimistic/pessimistic dopamine neuron behavior.
Pseudo-probability unlearning method for efficient privacy-preserving machine unlearning with reduced computational overhead and residual information.
Self-supervised physics-informed neural network for real-time human pose and dynamics estimation from sparse IMU sensor configurations.
Inference-time scaling method for discrete diffusion language models via trajectory refinement without retraining for reward optimization.
Method for quantitatively estimating target task performance from unsupervised pretext tasks in semi/self-supervised learning before full training.
In-context learning approach for AutoML pipeline optimization beyond hyperparameter tuning, incorporating fine-tuning and ensembling techniques.
PhISM: physics-informed deep learning architecture for unsupervised hyperspectral imaging using continuous basis functions for interpretable latent representations.
LoFT method for long-tailed semi-supervised learning using foundation models with parameter-efficient fine-tuning to improve pseudo-label quality.