Mosaic enables low-latency multi-agent planning by combining lightweight state tracking and constraint-guided coordination, achieving up to 32% faster execution and 10% point higher success rates.

This work systematically studies memory-efficient alternatives to backpropagation for LLMs, highlighting that FmAD and ZO save activation memory but incur higher compute, slower convergence, and lower accuracy.

Atom is an on-device system that speeds up video-language pipelines by decomposing large models into reusable modules, enabling parallel execution and reducing latency by ~30% with minimal performance loss.

Spry is a federated learning algorithm that enables finetuning LLMs using Forward-mode Auto Differentiation; to achieve low memory footprint, high accuracy, and fast convergence.

Flow addresses the challenge of statistical heterogeneity in federated learning through creating a dynamic personalized model for each input instance through a routing mechanism.

Flash uses client-side early-stopping training to facilitate detection of concept drifts and the server-side drift-aware adaptive optimization in a Federated Learning setup, to solve concept drift detection and adaptation challenges.

Multi-agent chatbots, like CommunityBots, improve user engagement and response quality over single-agent systems in multi-topic conversations by effectively managing topic and agent transitions.