Method for extracting and clustering traffic scenarios from real-world highway data using conditional VAE for autonomous vehicle testing.
DeepStage: Deep reinforcement learning framework for autonomous defense against multi-stage APT attacks using provenance graphs and stage estimation.
Quantum transfer learning architecture combining pretrained classical models with variational quantum classifiers for image classification on noisy hardware.
TorchNWP: Compiler library tool for coupling AI models with traditional numerical models, enabling Fortran-Python interoperability for weather prediction.
Reward prediction model for robot manipulation using vision foundation models to infer dense task rewards from camera images without privileged state.
Evaluation metric for generative models assessing whether synthetic data preserves multivariate dependence structures for downstream inference tasks.
DesertFormer: Transformer-based semantic segmentation pipeline for off-road desert terrain classification in autonomous navigation systems.
Multi-fidelity surrogate modeling framework for airfoil optimization combining low-fidelity simulations with Gaussian processes and genetic algorithms.
Ensemble self-training approach for unsupervised neural machine translation using multiple models with auxiliary languages and token-level ensemble decoding.
Auto-Prov: End-to-end framework using LLMs to construct provenance graphs from system logs for anomaly detection and threat interpretation.
Framework for locating knowledge in mixture-of-experts LLMs by analyzing cross-lingual inconsistencies, advancing interpretability of expert routing.
Self-supervised learning method for medical image segmentation using contrastive learning and counterfactual generation to handle imperfect AI labels.
Multi-agent routing architecture for AI reasoning systems with dynamic execution graphs, addressing cascade failure propagation in agent delegation networks.
Safe reinforcement learning framework for robots using temporal logic constraints to enforce safety and operational requirements during training.
TAP-GPT uses pretrained LLMs for few-shot Alzheimer's disease prediction from multimodal biomedical tabular data.
OPERA framework for data pruning to improve efficiency and effectiveness of dense retriever finetuning.
Adaptive contracts framework for cost-effective AI delegation, balancing evaluation noise against evaluation costs.
LLM-driven pipeline for anonymizing text by replacing PII with realistic surrogates while preserving data utility.
Approach for developing Tharu language LLM using synthetic data generation and human validation to address low-resource language gap.
Analysis of multimodal LLM segmentation capabilities through layerwise probing and attention mechanisms.
Red-teaming alignment framework (CRAFT) that improves LLM robustness against jailbreaks by optimizing hidden representations.
Multi-agent RL framework for dynamic memory controller optimization with explainable energy and latency objectives.
Offline RL framework (PIER) for fuel-efficient maritime routing using physics-informed models and historical vessel data.
Multimodal LLM framework for ride-hailing dispute resolution combining visual and logical reasoning with transparency.
AI coding agent that bootstraps itself by re-implementing its own specification, demonstrating meta-circular properties similar to compiler bootstrapping.
Hardware-aware lossless compression technique (ZipServ) for efficient LLM inference with reduced memory and bandwidth requirements.
Research on conditional attention mechanism (L2A) that reduces computational costs for long-context LLM inference by selectively attending to relevant tokens.
arXiv paper proposing Riemannian Mirror Descent, generalizing first-order optimization methods to Riemannian manifolds with convergence guarantees.
Unified LLM for search, recommendation, and reasoning over large heterogeneous catalogs generating unambiguous item references under latency constraints.
Studies semi-factual explanations in XAI showing elaborated counterfactuals are preferred by users for understanding ML predictions and exploring alternatives.
Semantic ID-based generative retrieval system deployed at Spotify balancing long-term preferences with intent-aware podcast discovery using contextual signals.
Per-domain Q-value functions using graph neural networks for efficient policy learning in planning, cheaper than state-value alternatives.
Emergent Trust Learning: lightweight trust-based control algorithm for AI agents enabling cooperation in competitive multi-agent environments with shared resources.
Theoretical analysis proving graph transformers have structural benefits over GCNs for node-level prediction via Gaussian process limits.
HeiSD: Hybrid speculative decoding for Vision-Language-Action robot control models combining drafter-based and retrieval-based acceleration with kinematic awareness.
rSDNet: Robust neural learning framework defending against both label noise and adversarial attacks using alternative loss functions to standard cross-entropy.
State space models for biomolecular dynamics modeling, accelerating MD simulations while preserving temporal relationships for drug discovery applications.
Adaptive guidance mechanism for RAG-enhanced masked diffusion models to handle retrieval-prior conflicts when context is noisy or inconsistent.
Sensi: LLM agent architecture for ARC-AGI-3 using curriculum-based test-time learning with perception-action separation and active hypothesis testing.
Stochastic set-valued optimization framework for robust machine learning using multi-objective optimization with hyperbox set representations.
Uses shuffle products and finite-state automata to model overlapped speech for alignment and speaker-attributed transcription via marginalizing serializations.
CoVerRL framework escapes consensus trap in label-free LLM reasoning by using generator-verifier co-evolution to maintain output diversity and avoid reinforced systematic errors.
Framework integrating HPC, ML, and quantum computing for drug discovery, replacing trial-and-error with quantitative precision in molecular dynamics.
ChopGrad reduces memory costs in video diffusion model training by using truncated backpropagation with pixel-wise losses instead of accumulating activations.
Language model infers stage-play layouts (scenes, positions, movements) from narrative text without explicit spatial cues, testing spatial reasoning capabilities.
Text-to-Stage: LLM-based spatial reasoning from narratives to generate stage-play layouts. Probes compositional reasoning in language models.
Physics-informed ML surrogates for power grid simulation validation. Application of ML to scientific computing.
Research on how LLMs generate verbal confidence scores. Investigates timing and computation of uncertainty estimates in black-box models.
scicode-lint: LLM-generated patterns for detecting methodology bugs in scientific Python code. Addresses sustainability of ML-specific linters.
LoST: level-of-semantics tokenization method for 3D shape generation improving autoregressive 3D generative models.