Framework evaluating optimal automation intensity by modeling cost minimization across no automation to full AI automation spectrum.
Using GPT-4 to automatically generate finite state machine specifications from natural language requirements documents.
Graph coarsening method for large-scale graph node classification reducing computational overhead in Graph Convolutional Networks.
LatentPilot: vision-and-language navigation using latent visual reasoning to imagine future visual dynamics from actions.
Generative virtual brain model predicting neuromodulation outcomes for Parkinson's disease treatment selection.
SLVMEval: meta-evaluation benchmark for assessing text-to-video generation evaluation systems on videos up to 3 hours long.
Adaptive context compression framework for LLMs in long-running interactions using importance-aware selection, coherence filtering, and dynamic budget allocation.
Multi-Layer Memory Framework decomposing dialogue history into working, episodic, semantic layers for stable long-horizon LLM agent interactions.
VulGNN: lightweight graph neural network for software vulnerability detection, achieving comparable performance to LLMs with lower computational requirements.
LiteCoST: two-pillar framework using chain-of-structured-thought with small LMs for accurate, low-latency document question answering over long, noisy documents.
MemRerank: preference memory framework that distills user purchase history for personalized product reranking in LLM-based shopping agents.
FlexMem: training-free approach for long video understanding in multimodal LLMs using visual memory mechanism inspired by human video watching behavior.
Fine-tuning CLIP to improve vision-language model understanding of negation in text and images.
PRISM: 270K-sample multi-view retail video dataset for training embodied vision-language models in real-world deployment.
Research on downsides of AI in industrial edge-cloud systems including maintenance, optimization and safety challenges.
Self-improving code generation for LLMs using semantic entropy and behavioral consensus without external resources.
Anomaly detection methods for spacecraft telemetry optimized for edge deployment using neural architecture search.
Systematic study on security and reliability risks of using LLMs as evaluators (LLM-as-a-Judge) paradigm, covering adversarial vulnerabilities.
Method for reducing hallucinations in vision-language models through intermediate representation editing without full retraining.
AGFT: alignment-guided fine-tuning to improve adversarial robustness of vision-language models while preserving zero-shot generalization.
Study on adversarial prompt injection attacks against multimodal large language models using imperceptible visual perturbations.
Lightweight uncertainty quantification method for neural networks using gradient norms and isotropy assumption, suitable for large language models.
M-MiniGPT4: multilingual vision-language model supporting 11 languages using translated data and alignment training.
MemFactory: unified framework for training and inference of LLM agent memory systems with RL optimization for extraction, updating, and retrieval.
Study on meaning representations for natural language generation in dialogue systems, analyzing task impact on conversational AI performance.
Research on language models trained on child language data from BabyView dataset to understand data requirements and linguistic knowledge emergence.
Quantization and adaptive distillation method for deploying large vision models with multiple LoRAs on edge devices.
Mean Masked Autoencoder (MMAE) with flow-mixing for encrypted network traffic classification using self-supervised learning.
Small-scale language model study investigating multilingual acquisition patterns and optimal input structures for bilingual learning.
World model planning approach for origami generation with constraint satisfaction for multi-step geometric reasoning.
Agent system for automatically generating method illustration figures for scientific papers using compositional planning.
Multi-turn image editing agent with planning and reflection mechanisms to handle context-aware constraints and error accumulation.
Cognitive neuroscience study comparing neural language model representations with human brain processing of linguistic constructions.
Empirical comparison of multi-agent coordination structures for automated machine learning research using LLM-based agents.
Masked augmentation method for continuous image tokenizers addressing posterior collapse in visual generation models.
Closed-loop benchmark environment for training autonomous 6G network management agents with tool use and learning from environmental feedback.
LLM-based approach for extracting multiple narrative paths from data with user guidance and agenda support.
Framework analyzing challenges in multimodal active learning when modalities are missing or have varying difficulty levels.
Visual analytics system for knowledge editing workflows in LLMs with layer-level analysis and detailed intervention guidance.
Event-driven simulation framework using LLM-based agents to study social dynamics in controlled environment with real-time data streams.
Edge-based computer vision system for detecting violent behavior in public spaces with focus on latency and privacy constraints.
Applies quantum chemistry methods to study phase transitions and emergent capabilities in deep learning via spectral analysis of training dynamics.
Vision-Language-Action model architecture decoupling intent and action through latent world modeling to improve high-level decision making and training stability.
Comprehensive benchmark suite for SVG code generation, sketching, editing and understanding with four tasks including novel Sketch2SVG and editing datasets.
Method for editing Implicit Neural Representations by discovering deformation eigenmodes from Gram operators induced by INR features without retraining.
Sentence-level readability scoring system for German ESG reports to improve consumer comprehension of sustainability information.
Compares GraphRAG to simpler VectorRAG approaches for retrieval-augmented generation, analyzing effectiveness of graph-based relation representation versus vector indexing.
Evaluates LLM performance in automated RDF knowledge graph generation from cloud system logs for improved interpretability and root-cause analysis.
Large-scale field experiment measuring generative AI assistant impact on e-commerce customer service worker performance at Alibaba, showing task diagnosis and solution generation.
Validation study testing whether interview-informed LLM agents can simulate user responses in product discovery and concept testing scenarios.