Training-free retrieval-augmented generation framework for flood damage nowcasting using reinforced reasoning with geographically local context retrieval.
Method for applying LLMs to graph problems using human-interpretable graph encodings that address the mismatch between text-based models and graph reasoning.
Adaptive model selection framework for demand forecasting that accounts for horizon-induced performance degradation in inventory planning.
SphUnc framework combining hyperspherical representation learning with structural causal modeling for uncertainty decomposition in multi-agent systems.
Study on why multimodal models underperform unimodal counterparts in context-aided forecasting, investigating context quality issues in datasets.
Framework using optimal transport and distributionally robust optimization to analyze ML model vulnerabilities through constrained data perturbations for explainability.
AutoML approach combining deep unfolding with learnable proximal gradient descent for wireless optimization.
PLR: Plackett-Luce ranking method for efficient reordering of in-context learning examples in LLMs.
Cascading architecture for aircraft health diagnosis balancing accuracy and computational constraints.
Assessing robustness of climate foundation models under no-analog distribution shifts from climate change.
Masked autoencoder with normalizing flow for time series anomaly detection using foundation model approach.
Generalization bounds for overparameterized shallow networks related to distance from initialization.
QuanBench+: Unified benchmark for LLM-based quantum code generation across Qiskit, PennyLane, Cirq frameworks.
Hardware-efficient neuro-symbolic networks using Exp-Minus-Log operator for safety-critical edge deployment.
Apollo: Multimodal temporal foundation model for patient representations trained on 25 billion clinical records.
Bounded Ratio RL framework bridging trust region methods with PPO's clipped objective for on-policy learning.
Framework for adaptive hardware selection and tuning in cloud deep learning with cost-robustness tradeoffs.
LLM agents simulate individual behavior using self-report data for generative simulations of persons.
Meta-learning approach using hypergradients to automatically optimize data synthesis parameters for domain randomization in brain image segmentation.
Optimization system for LLM batch inference exploiting prefix sharing and throughput-oriented token batching to improve inference engine performance.
Hybrid Transformer architecture integrating quantum circuits into a single encoder layer's value projection while keeping other layers classical.
Imitation learning algorithm for robot control that maintains dynamic stability by addressing timing discrepancies in expert-agent switching for dynamic tasks.
Probing study of 25 modern language models to understand how they encode lexical identity and inflectional features across languages.
Multi-armed bandit framework incorporating machine learning-generated surrogate rewards using auxiliary data to improve sequential decision-making.
Improved value-iteration algorithms for solving multichain Markov decision processes under average-reward criterion.
Adaptive multi-task learning methodology for multi-sector portfolio optimization with transfer learning across asset classes.
LLM-based passage ranking with improved reasoning ability using reasoning models and step-by-step inference during test time.
Neural bandit framework for optimally selecting LLMs for subtasks in agentic AI pipelines to minimize cost while ensuring success.
Design principles and patterns for building robust, reliable GenAI-native systems integrating cognitive capabilities with traditional software engineering.
Scaling laws for post-training quantized LLMs stratified by task type to understand differential impacts on memorization, application, and reasoning.
Study of transformer capabilities in learning transitive relation inference across different graph structures relevant to LLM reasoning.
D-optimal design methodology for estimating main effects in black-box models with improved robustness to out-of-distribution evaluations.
Research on how LLMs develop universal sinusoidal representations of numbers and their interchangeability across different model families.
Unsupervised network anomaly detection using variational graph autoencoders for intrusion detection without requiring labeled datasets.
Self-consistency method for learning diffusion bridges via optimal control. Addresses simulation of conditioned diffusion dynamics for rare events.
Device-native autonomous agent system for privacy-preserving negotiations in insurance and B2B commerce. Runs locally without sending sensitive data to centralized servers.
KOCO-BENCH: benchmark for evaluating how LLMs acquire and apply domain knowledge in software development. Tests domain specialization effectiveness beyond knowledge possession.
veScale-FSDP: distributed training system enabling flexible sharding formats beyond fixed element-wise/row-wise patterns. Supports block-structured and quantization-aware training.
FlexServe: LLM inference system for mobile devices with hardware-based isolation using ARM TrustZone. Protects model weights and user data from compromised OS.
MotionGPT3 replaces diffusion with rectified flow for text-driven motion generation. Improves convergence and inference efficiency in continuous motion synthesis.
Video diffusion model learning joint distribution of videos and camera trajectories. Treats novel view synthesis and camera parameter recovery as unified task.
Pipeline and best practices for log analysis in AI systems. Standardized approach with concrete code examples for analyzing tool use and model behavior.
Linear probing study of how LLMs internally represent rhetorical questions. Analyzes rhetorical signals in model representations across social-media datasets.
BARD framework bridges autoregressive and diffusion vision-language models via progressive block merging and distillation. Improves inference efficiency of multimodal models.
Quantum-kernel SINDy framework for sparse identification of nonlinear dynamics. Addresses coefficient cannibalization problem in quantum feature maps for classical learning.
Dataset for grammatical error detection and correction in Romanian legal documents. Domain-specific NLP resource for training correction models.
Meta plans to use employee keystroke and mouse movement data to train AI models.
Tool for managing AI agent skills with versioning and profile support, inspired by pnpm's approach to deduplication and sharing across projects.
Analysis of 'tokenmaxxing' trends from Google Cloud Next and AIE Miami, focusing on LLM efficiency and context optimization.
Linux Foundation announces OCUDU Ecosystem Foundation for open source AI-RAN innovation and reference platform development.