All news

Mix-or-Split schedules LLM inference across devices, edge and cloud using a fine-grained latency model.

Joint caching of large models with parameter sharing and inference offloading in mobile edge computing.

DIAMM splits multimodal model inference across edge devices to cut latency.

FairCache shows prefix and context caching makes multi-tenant LLM serving unfair and proposes cache-aware scheduling to fix it.

PrefixQuant adds prefixed tokens to absorb activation outliers, enabling simpler static LLM quantization.

BitNet v2 adds low-bit, sparsified activations to 1-bit LLMs to cut inference cost further.

CeeTrain speeds up geo-distributed model training across mixed GPUs and networks spanning cloud, edge and devices.

A framework that detects and reduces QoS violations caused by noisy neighbors in multi-tenant public clouds.

ESR-HGNN removes redundant semantic computation to speed up mini-batch heterogeneous GNN inference.

Hydra-MCS splits maximum common subgraph search across CPU and GPU for speedups.

Uses contextual sparsity to get around the memory bandwidth limit in edge LLM inference.

Strategies to handle silent data errors in workflows of moldable parallel tasks.

Proposes serving RAG workloads through decentralized collaboration of edge devices instead of the cloud.

Deep reinforcement learning task scheduling for mobile edge computing that accounts for providers overselling capacity.

Uses curiosity-driven reinforcement learning to schedule requests collaboratively across mobile edge and cloud servers.

Schedules dependent tasks and manages container image layers together across device, edge and cloud tiers.

Speeds up out-of-core graph analytics on SSDs by preprocessing graphs based on which vertices stay active.

Pose-aware risk mitigation for end-effector actions in vision-language-action robot models.

CARSS combines online arrival forecasting with co-simulation to adapt job scheduling.

PROSPECT unifies streaming vision-language navigation with semantic-spatial fusion.

Language-guided state representations improve policy learning in world models.

An equilibrium scheduling strategy for multiple low-voltage distribution networks using Bayesian segmentation of their complementarity.

A bi-level flexibility trading scheme for economically scheduling several microgrids in one distribution system.

Survey of low-precision training for LLMs, including FP8 and lower formats.

Mix-attention fuses multimodal LLM features into sequential recommendation.

EcoxPU orchestrates disaggregated LLM inference across heterogeneous accelerators with agents to cut energy use.

A network flow scheduler for AI clusters that penalizes contention so latency-sensitive flows meet deadlines alongside bulk training traffic.

A Stackelberg game plus diffusion-based auction for offloading agentic AI tasks in IoT.

Distills reliable recommendation rationales from LLMs for conversational recommenders.

DWI speeds up on-device LLM decoding by picking the width of a nested model per step.