MineDraft: A Framework for Batch Parallel Speculative Decoding
MineDraft framework for batch parallel speculative decoding to accelerate LLM inference by hiding draft and verify stages.
MineDraft framework for batch parallel speculative decoding to accelerate LLM inference by hiding draft and verify stages.
Differentiable rendering technique for RF digital twins enabling gradient-based optimization of radio frequency systems.
Sensor fusion method combining UWB and inertial measurement for indoor localization under non-line-of-sight conditions.
Corpus poisoning attacks and defenses for RAG systems, demonstrating vulnerabilities in LLM-extended retrieval pipelines.
Training technique applying sharpness-aware minimization to spiking neural networks using surrogate gradient methods.
Medical imaging method combining intuitionistic fuzzy logic with U-Net architectures for MRI brain image segmentation.
FPGA-based SoC architecture for spiking neural networks using RISC-V controller and event-driven computation for edge AI.
Digital RTL architecture implementing predictive coding networks as alternative to backpropagation for distributed hardware learning.
Framework integrating deep generative models and normalizing flows to accelerate replica exchange molecular simulations.
Foundation diffusion model for computational pathology and histopathology image analysis with self-supervised learning.
Quantization-aware drift correction method for diffusion model sampling to reduce degradation from post-training quantization noise.
Few-shot learning adapter for CLIP using patch-level and text supervision without increasing inference costs.
Defense mechanism against backdoor attacks in audio/speech models using stability-based trigger detection at inference time.
Alternative training architecture for AI models using non-standard arithmetic and memory management for geometric and neuromorphic AI.
Transfer learning for pricing and assortment optimization across markets using multinomial logit choice models with bandit feedback.
Insight-V++ framework enables multi-agent visual reasoning for MLLMs with long-chain reasoning, addressing data scarcity and training optimization.
MAED detects activation errors in DNN inference to mitigate physical fault attacks on embedded neural networks.
Theoretical study on optimal sample complexity for learning unknown bosonic Gaussian quantum states in continuous-variable systems.
Learning-augmented algorithm framework using online learning to inform solutions for k-median clustering problems.
Theoretical analysis proving convergence of ResNet training dynamics in large-scale limits across depth, width, and embedding dimension.
Study of data preparation pitfalls in insurance modeling, highlighting instability of standard train-test splitting on imbalanced data.
Hybrid diffusion-DeepONet framework predicts stress fields in hyperelastic materials with improved handling of sharp gradients.
Study on how ASR quality impacts Alzheimer's disease detection from speech transcripts using lexical feature modeling.
ChoiceEval framework audits brand and cultural preference biases in LLMs to assess market fairness and information diversity risks.
Neural Architecture Search applied to NeRF models for efficient satellite scene 3D reconstruction with reduced training time.
MemArchitect adds governance layer for LLM agent memory management, handling contradictions, privacy, and outdated information in persistent RAG systems.
VCoT-Bench evaluates LLMs on Rust program verification via chain-of-thought reasoning, testing logical deduction abilities beyond binary pass/fail.
Method for reliable uncertainty quantification in Vision-Language-Action models by shifting focus to safety-critical moments in robotic control.
Dataset for detecting human situational awareness gaps in remote human-robot teaming through multimodal sensor data.
PowerFlow applies principled distribution matching to unsupervised reinforcement learning from LLM internal feedback without external supervision.
Theoretical study of computational and statistical hardness in computing calibration distance for probabilistic predictor evaluation.
Research on estimating causal representations from multi-domain data using empirical Bayes methods for causal representation learning.
TARo enables frozen LLMs to perform structured reasoning at inference time through token-level adaptive routing, avoiding expensive post-training alignment.
Statistical framework for quantifying reliability of results from data analysis pipelines using selective inference techniques.
Unsupervised discovery of transition-structure concepts in text via temporal co-occurrence patterns using contrastive learning on large corpus.
Adaptive context allocation method for LLM long-context inference using uncertainty-triggered token-level budgeting to address attention dilution.
Vision-language model method for temporal out-of-distribution detection and domain generalization in open-world settings using adaptive pattern matching.
Analysis of how standard LLM decoding strategies (top-k, nucleus sampling) exclude contextually appropriate but statistically rare tokens compared to human language production.
Theoretical analysis of linear denoisers for noisy data, studying performance in proportional regime without known covariance.
Analyzes optimal satisficing regret bounds for nonstationary K-armed bandits with piecewise-stationary segments.
ICE-Guard detects spurious feature reliance in LLM decision-making through intervention consistency testing on demographic, authority, and framing features.
Addresses sim-to-real transfer for vision-language-action models in robotics by generating diverse 3D simulation worlds for RL fine-tuning.
iSatCR optimizes onboard computing and routing for LEO satellite data processing using graph neural networks to reduce ground transmission bottlenecks.
Studies construction of compressed decision-sufficient datasets for linear programs with unknown cost vectors using decision-relevant dimension theory.
CausalVAD applies causal intervention to de-confound end-to-end autonomous driving models, addressing dataset bias and improving reliability.
ICE framework evaluates explanation faithfulness in LLMs via randomization tests with multiple intervention operators, distinguishing genuine faithfulness from chance.
WarPGNN applies physics-aware graph neural networks for efficient thermal warpage analysis in chiplet-package systems, replacing costly numerical simulations.
DRESS graph fingerprinting achieves unique fingerprints across 51,718 non-isomorphic strongly regular graphs using single-deletion vertex operations.
i-SDT combines predictive modelling and multi-class attack discrimination for detecting and responding to cyber-physical system attacks without full shutdowns.
SwiftGS enables rapid 3D satellite surface reconstruction via meta-learned Gaussian primitives predicted in a single forward pass for environmental monitoring.