Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2607.09603.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

Mosaic is a multi-agent embodied planning framework that reduces LLM-driven execution latency by combining agent-centric semantic memory (relative-coordinate object storage) with Integer Linear Programming for action allocation. On AI2-THOR and search-and-rescue benchmarks it reports 27-32% faster execution, 30-33% fewer LLM calls, and 4-10pp higher success rates.

Quantifies a bottleneck most multi-agent papers gloss over — failed actions dominate latency — and offers a concrete dual remedy. Useful reference for anyone building LLM agent stacks that coordinate.

Cited by

No citations on record.