Research map / Multi-agent systems and games / LLM agents

DRAMA: Next-Gen Dynamic Orchestration for Resilient Multi-Agent Ecosystems in Flux

CVPR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DRAMA, a three-layer embodied multi-agent orchestration framework with affinity-based Hungarian scheduling, hierarchical trust-chain task takeover, and collective spatial inference, evaluated in VirtualHome-Social under dynamic agent populations.

Problem

Embodied multi-agent systems typically use static agent-task assignments that cannot adapt when agents join, drop out, or recover, or when tasks change, leading to failures and inefficiency.

Approach

Models agents and tasks as resource entities; uses a dual-capacity Hungarian assignment over an affinity matrix (workload availability minus predicted distance) with event-triggered rescheduling, an ordered guardian trust chain per task for immediate takeover, and an MLP-based collective spatial prior learned from pooled agent observations.

Key results

• Across static and dynamic scenarios, DRAMA reduces Average Steps by about 4% on average and up to 7-8% in highly dynamic scenarios versus baselines (Table 1).
• DRAMA is the only method that succeeds in agent dropout and recovery scenarios; all baselines fail those (Table 2).
• Throughput improves by roughly 10-20% over existing methods across scenarios; conflict rate is lowest among dynamic schedulers.

Surprising findings

• MCTS achieves lower conflict rate than DRAMA in several settings, attributed to its static predefined scheduling policy.
• Performance of DRAMA is highly consistent across four different backbone LLMs, suggesting the framework's gains are not tied to one model.

Limitations

• Maximum of five agents, described as the upper bound of current system capacity.
• Baselines were re-implemented and adapted to VirtualHome-Social; efficiency evaluation gave all methods extra inputs (final target and progress), and robustness evaluation followed original designs without handover support.
• Evaluation limited to a single simulated environment with a custom, self-designed task configuration.

Experimental setup

Datasets: VirtualHome-Social

Benchmarks: VirtualHome-Social

Metrics: Average Steps Conflict Rate Throughput scenario success

Baselines: CoELA MCTS AgentVerse-Static AgentVerse-Dynamic ProAgent

Builds on: Hungarian algorithm CoELA ProAgent AgentVerse MASTER

Closest papers by meaning

• Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.86)
• Recognize Your Orchestrator: An Entropy Dynamics Perspective for LLM Multi-Agent Systems (ICML 2026 · similarity 0.86)
• From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.86)
• Augmented Runtime Collaboration for Self-Organizing Multi-Agent Systems: A Hybrid Bi-Criteria Routing Approach (AAAI 2026 · similarity 0.86)
• AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.86)
• Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.85)