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

Augmented Runtime Collaboration for Self-Organizing Multi-Agent Systems: A Hybrid Bi-Criteria Routing Approach

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

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 BiRouter, a bi-criteria next-hop routing method for self-organizing multi-agent systems that uses learned importance and gap scores plus agent reputation, trained on a synthetic routing dataset (MARS).

Problem

Coordinating LLM agents in decentralized, open networks where each agent only has local information, since existing MAS planners rely on static topologies or centralized global planning.

Approach

Each agent picks its successor with a softmax over reputation-gated mix of a learned importance score (task relevance, A*-like heuristic) and a learned gap score (contextual continuity, A*-like path cost), plus a Finisher agent for adaptive termination and successor-aware messages.

Key results

• In centralized MAS, BiRouter averages 91.73 across six benchmarks, versus 87.08 for GPTSwarm and 86.43 for LLM-Debate (best baselines listed in the table).
• In the decentralized SO-MAS setting, BiRouter reaches 91.99 on GSM8K and 89.63 on HumanEval, with lower token cost than DyLAN and MaAS on GSM8K.
• Removing the Finisher agent raises token overhead by about 69% on GSM8K and 50% on HumanEval.

Surprising findings

• Table 1 reports BiRouter at 100.00 on MultiArith and the average of 91.73 in a comparison where several baseline rows contain missing entries (--), so the averages are not fully comparable.

Limitations

• Decentralized evaluation simulates each agent with only three randomly chosen neighbors, not a real open network.
• Robustness tested only with agents whose prompts were rewritten to produce incorrect outputs.
• Authors note online adaptive optimization is future work, so the router is not updated during deployment.

Experimental setup

Datasets: MARS (synthetic multi-agent routing dataset) GSM8K MBPP

Benchmarks: MMLU GSM8K MultiArith SVAMP HumanEval MBPP

Metrics: accuracy pass@1 token cost

Baselines: Single-agent CoT ComplexCoT CoT-SC MacNet LLM-Blender LLM-Debate GPTSwarm G-Designer AgentVerse DyLAN MaAS

Builds on: A* search AFlow G-Designer MaAS DyLAN

Closest papers by meaning

• AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.90)
• HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems (ICML 2026 · similarity 0.88)
• DiSRouter: Distributed Self-Routing for LLM Selections (ICLR 2026 · similarity 0.87)
• COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
• SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.87)
• TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination (ICML 2026 · similarity 0.87)