LLM-powered workflow optimization for multidisciplinary software development in automotive industry bridging domain experts and developers.
KG-Hopper: Framework enabling compact open LLMs to perform multi-hop knowledge graph reasoning via reinforcement learning.
SmaAT-QMix-UNet: Parameter-efficient vector-quantized model for precipitation nowcasting weather prediction.
Sparse Feature Attention: Method to reduce transformer self-attention complexity via feature sparsity instead of sequence-level sparsity.
Code Review Agent Benchmark: Dataset for evaluating AI agents on code quality assurance and review tasks.
Synthetic Mixed Training method combining synthetic QA and document generation to improve LLM knowledge acquisition beyond RAG performance.
LLM-enabled automated threat hunting framework for SOC analysts integrating Splunk log analysis with policy guidance.
X-OPD: Cross-modal on-policy distillation method to align end-to-end speech LLMs with text-based performance.
Few-shot 3D reconstruction combining TensoRF tensor representation with frequency-driven regularization.
Hybrid memory mechanism for video world models handling dynamic objects that leave and re-enter view.
Vision-language-action model for autonomous driving with natural language instruction following capability.
Multi-speaker audio preprocessing framework for full-duplex speech language models with conversational data.
Physics-aware local conditioning scheme for generative video models enforcing physical constraints.
Efficient vision backbone architecture designed for low-parallelization CPU devices.
Optimization approaches for buffer storage and retrieval in automated production systems.
Open-source tendon-driven dexterous robot hand (Ruka-v2) with 11 DOF for robot learning applications.
Selective gradient projection method mitigating catastrophic forgetting in continual learning.
Domain adaptation technique for EEG emotion recognition across heterogeneous datasets.
Method for selecting optimal visual in-context demonstrations for multimodal LLMs using sequential selection.
Training-free distillation framework transferring multimodal reasoning knowledge via context-based selection.
Federated learning approach for pretraining multimodal large language models on distributed private data.
Batch-level query routing framework for LLMs optimizing model assignment under cost and capacity constraints.
Method detecting memorization in LLM-based financial forecasting using membership inference and cross-model disagreement.
Framework to explain and align semantic hierarchies in CLIP and other vision-language model embeddings.
Probabilistic self-supervised representation learning via Gaussian joint embeddings for multi-modal problems.
Neural operators for long-term fluid dynamics forecasting addressing stability and precision in PDE modeling.
Unified sparsification framework for cross-modality prediction across graphs, language, and tabular data.
Physics-informed contextual spectral reinforcement learning method for adaptive sensing in high-dimensional low-sample-size datasets using domain knowledge embeddings.
Patient-adaptive transformer framework for seizure prediction from EEG using two-stage training with self-supervised pretraining to handle inter-patient variability.
Analysis of throughput optimization as critical strategic lever in large-scale LLM training, synthesizing dataloader and memory profiling innovations to reduce bottlenecks.
Central-to-local adaptive generative diffusion framework for gene expression prediction in spatial transcriptomics addressing data scarcity through transfer learning.
Squish and Release activation-patching technique exposes hidden hallucinations in LLMs that models suppress via safety circuits after identifying false premises.
Statistical regression framework for analyzing impact of specific prompt components on LLM performance, extending XAI methods to understand LLM behavior.
MazeBench evaluates 16 multimodal AI models on visual maze solving, revealing models achieve high accuracy through token-space brute-force search rather than genuine visual planning.
Semi-end-to-end framework combining saliency-guided feature separation with Transformer for robust muscle fatigue recognition from surface electromyography signals.
Foundation model approach for time series anomaly detection using masked autoencoder and normalizing flow to improve generalization across datasets with limited training data.
Method detects LLM deception by exposing hidden hallucinations through activation patching, revealing safety circuit suppression of identified errors under conversational pressure.
Hierarchical sheaf spectral embedding framework for single-cell RNA-seq analysis capturing heterogeneous local structure across multiple scales.
MoltenFlow framework for inverse molecular design using latent flow models to optimize molecular structures for desired properties with improved validity and stability.
Analysis of strategic gaming in AI model ranking systems where producers submit multiple variants to artificially inflate rankings from noisy preference data.
Novel domain adaptation methods using unfolding approach to improve model generalization across domains with varying data distributions without separate per-domain training.
Probabilistic surrogate model for wildfire spread prediction using conditional flow matching to model fire progression as stochastic process.
Sequence-based predictor for T cell receptor specificity against peptide-MHC complexes using machine learning for immunology and personalized medicine applications.
Method compresses deep reinforcement learning policy parameter space into low-dimensional latent manifold to improve sample efficiency through state-occupancy matching.
Liquid neural networks with mixture density heads outperform diffusion policies in imitation learning with half the parameters and 2.4x lower prediction error.
Signal Temporal Logic inference learns interpretable rules for temporal behaviors using conformal prediction for uncertainty quantification under covariate shift.
arXiv: Deep reinforcement learning for dynamic manufacturing resource matching and allocation.
arXiv: Generative model for valid 3D molecular structures using hierarchy-guided topology flows.
arXiv: DRIFT framework for individualized treatment effect estimation across multiple health domains.
Text-to-time-series generation framework for meteorological data using spectral-aware architecture with billion-scale dataset.