Sequential-AMPC uses recurrent neural networks to approximate nonlinear model predictive control offline, reducing online computation for embedded hardware control systems.
AI agents using Claude Code autonomously discovered novel adversarial attack algorithms for LLMs that outperform 30+ existing methods in jailbreaking and prompt injection attacks.
Agentic Variation Operators replace fixed mutation/crossover in evolutionary search with autonomous coding agents consulting lineage and domain knowledge.
TuneShift-KD enables knowledge distillation and transfer of fine-tuned specialized knowledge to newer LLM architectures without access to original training data.
Multi-dimensional evaluation framework for uncertainty attribution methods in explainable AI addressing inconsistent evaluation across heterogeneous proxy tasks.
UI-Voyager is a self-evolving mobile GUI agent using rejection fine-tuning and credit assignment to learn from failed trajectories in long-horizon tasks.
RAVEN applies generative pretraining to structured electronic health records using recurrence-aware next-visit event prediction on 1M+ patient dataset.
DreamerAD enables efficient RL for autonomous driving via latent world model achieving 80x speedup by compressing diffusion sampling from 100 to 1 step.
Multilevel Euler-Maruyama method accelerates diffusion model solving via multi-level approximators with polynomial speedup in HTMC regime.
KARMA applies LLMs to personalized search at Taobao by addressing knowledge-action gap through regularized multimodal alignment for next-item prediction.
Deletion-Insertion Diffusion language models replace masking paradigm with discrete diffusion processes for improved computational efficiency and generation flexibility.
DepthCharge framework measures knowledge depth in LLMs through adaptive probing across domains, addressing inability to sustain accurate responses in domain-specific details.
Study of prospective memory failures in LLMs when formatting constraints conflict with complex tasks.
MDKeyChunker: Structure-aware document chunking and single-call LLM enrichment for improved RAG pipelines.
Transformer-based approach for polarization detection in social media using threshold tuning and class weighting.
LLMORPH: Automated metamorphic testing tool for LLMs using metamorphic relations to verify correctness without labeled test data.
Benchmark of Qwen 2.5 1.5B quantized LLM inference across mobile, NPU, and GPU platforms measuring throughput and efficiency trade-offs.
Comprehensive review of energy-efficient software-hardware codesign for ML from TinyML to LLMs, addressing memory and data movement bottlenecks.
Dual-gated approach for autonomous compute modulation in asynchronous multi-agent reinforcement learning on edge devices.
Using sparse autoencoders to replace opaque vision foundation model representations with human-interpretable features for medical imaging.
LLM-informed planning framework for object search in partially-known environments using LLM probability estimates and prompt selection.
Perturbation-based method to trace and analyze linguistic representations in deep language models without imposing linearity constraints.
Security analysis of quantized edge-deployed LLMs showing knowledge extraction attacks remain effective despite quantization noise.
DeepXube: Open-source Python package combining deep reinforcement learning and heuristic search to automate pathfinding problem solving.
Praxium: AI-based system for diagnosing microservice anomalies in cloud applications using telemetry and dependency analysis.
MTP-D: Self-distillation method to improve multi-token prediction in LLMs, addressing acceptance rates and joint training challenges for faster inference.
AttentionPack optimizes vision-language model inference with memory-efficient decoding for long sequences.
ORACLE orchestrates NPC daily activities in digital environments using contrastive learning with Transformer-CVAE.
LLM-based ambient assistant for evidence-based medical guidelines that surfaces targeted questions during physician consultations.
Systematic study reveals pricing reversal phenomenon where cheaper reasoning LLMs often cost more in practice across diverse tasks.
End-to-end optimized machine vision system for low-light scenarios with minimal detected photons per inference.
COVTrack++ enables multi-object tracking for open-vocabulary categories including unseen objects using continuous video data.
DeepIn framework for self-interpretable neural networks that identifies minimal representations needed for DNN expressiveness.
KG-M3PO framework combines knowledge graphs, vision, and reinforcement learning for multi-task robotic manipulation with online 3D scene graphs.
ML-based multi-layer security framework for Industrial IoT addressing resource constraints and threats across network layers.
MedAidDialog multilingual multi-turn medical dialogue dataset for conversational AI in healthcare with improved realism over template-based systems.
TSRL framework uses reinforcement learning to dynamically optimize training curriculum for deepfake detection, modeling training as an MDP.
Visual study of UMAP projections examining geometric patterns in embedding difference vectors of antonym and synonym word pairs.
Applies quantum convolutional neural networks to solve partial differential equations on quantum simulators for scientific computing applications.
HEART-PFL framework for personalized federated learning using hierarchical directional alignment and adversarial knowledge transfer to handle data heterogeneity.
UniScale explores synergistic data and model scaling for search ranking, demonstrating that joint architectural and data design improvements outperform model scaling alone.
DVM enables real-time kernel generation for dynamic AI models, addressing compilation overhead and memory footprint issues in runtime compilation.
C-STEP introduces physics-informed safety measures for reinforcement learning in robotics, using intrinsic rewards for safe navigation in continuous domains.
CGRL framework addresses poor generalization of GNNs on out-of-distribution data using causal-guided representation learning to avoid spurious correlations.
Proposes method to quantify self-awareness in intelligent systems by identifying invariant cognitive processes that change slower than acquired skills.
Neuro-symbolic system using attention-based encoders and differentiable reasoning rules to detect human fatigue from eye-tracking and fNIRS signals.
Investigates joint effects of differential privacy and fairness constraints on federated classification systems across distributed servers.
Studies relationship between fair model representations and fair recommendations in recommender systems, examining demographic attribute classification.
Analysis of why self-distillation degrades LLM reasoning capability by suppressing epistemic verbalization and expression of uncertainty.
Composer 2 model specialized for agentic software engineering with long-term planning and coding abilities trained via continued pretraining and reinforcement learning.