Mosaic framework cuts multi-agent AI planning time by 30%
Failed actions were the bottleneck—Mosaic uses memory and optimization to fix them.
Multi-agent planning with LLMs has been notoriously slow, limiting real-world deployment. A new paper from researchers (Panchal et al., accepted to ICML 2026) identifies failed actions as the primary culprit, stemming from poor state tracking under partial observability and redundant or conflicting coordination. Their solution, Mosaic, rethinks both memory and coordination. It introduces agent-centric semantic memory that stores objects in relative coordinates, enabling lightweight geometric transformations without heavy inference. For coordination, it leverages Integer Linear Programming (ILP) to allocate actions at each planning step, enforcing physical feasibility and inter-agent constraints.
On AI2-THOR and search-and-rescue benchmarks, Mosaic delivers significant gains: 27–32% faster execution, 30–33% fewer LLM calls, 25–31% fewer steps, and 4–10 percentage points higher success rates. These results show that efficient memory and constraint-guided coordination are critical for scalable, low-latency multi-agent embodied planning. The framework could enable more responsive robot teams in warehouses, disaster response, and other real-time settings.
- Mosaic achieves 27-32% faster execution and 30-33% fewer LLM calls compared to prior multi-agent planners.
- Uses agent-centric semantic memory with relative coordinates for accurate state tracking and geometric reasoning.
- Employs Integer Linear Programming at each step to enforce physical feasibility and resolve conflicts between agents.
Why It Matters
Makes multi-agent robot teams practical for real-time tasks like search-and-rescue and warehouse automation.