Topic
multi-agent systems
Exit-and-Join Dynamics Enable Decentralized Coalition Formation in Multi-Agent Systems
A new paper by Zhu and Quanyan presents a decentralized dynamical process for coalition formation driven by unilateral exit-and-join decisions. Agents evaluate moves using the Aumann-Dreze value, leading to equilibrium characterizations and stability analysis through Lyapunov representations.
New Framework Verifies Safety of Multi-Agent AI Communication for Autonomous Logistics
A new framework uses decision tree distillation to formally verify learned communication policies in multi-agent systems, targeting safety-critical autonomous logistics operations. The approach achieves 97.9% fidelity to neural policies and verifies 18 temporal logic properties with 88.9% satisfaction, including collision probabilities below 1% thresholds.
Evaluator Bias Spreads Like a Contagion in Multi-Agent LLM Systems, New Research Finds
A new paper from arXiv introduces 'Contagion Networks,' a formal framework to measure how evaluation biases propagate when large language models serve as evaluators in multi-agent systems. In a controlled experiment using DeepSeek-chat, researchers found consistent bias propagation between agents, and demonstrated that increasing the evaluator committee from one to three agents reduces effective contagion by 72.4%.
Hierarchical Control in Multi-Agent Games: LLM Planning with RL Execution Outperforms Flat Learning
Researchers propose a hierarchical architecture where a large language model (LLM) acts as a centralized strategic controller selecting among specialized RL skill policies for a team of agents. In a 2v2 King of the Hill environment, the LLM+RL system achieved a 46.4% win rate, statistically equivalent to hand-crafted behavior trees (51.5%), and significantly outperformed flat RL. A user study found 60% of participants perceived the LLM+RL agents as the most human-like.
Argent Signaling Protocol Mitigates Semantic Drift in Multi-Agent AI Systems
Researchers introduce the Argent Signaling Protocol (ASP), a machine-readable header that tags AI responses with certainty, grounding, stochasticity, and assumption indices. In tests on document-grounded QA, ASP improved pass rates from 11.1% to 33.3% on a small model and blocked 100% of ungrounded outputs in multi-agent mode.
LLM Agents Fail Classical Consensus Tests, But Filters Improve Reliability, Study Finds
A new study from researchers (Anand, Sribalaji C, Pappas, George J) examines whether classical resilient consensus theory applies to LLM agents in multi-agent systems. Framing LLM agreement as a Byzantine consensus game, the authors found that prompted LLM agents fail to reach agreement that is achievable in principle, even across temperatures and horizons. Wrapping agents with classical resilient consensus filters improved agreement, though the benefit depends on the underlying topology's robustness.
New Framework Prevents Artificial Hivemind in Autonomous Agent Economies Using Entropy Control
Researchers propose the Behavioral Protocol Framework (BPF), an entropy-controlled pluralistic alignment system to prevent the 'artificial hivemind' effect in autonomous agent economies. The framework integrates three modules: Mentalizing-based Social Intelligence, Pluralistic Alignment, and Verifiable Execution Kernel. Anticipated results show improved stability, efficiency, and trustworthiness of agent-native economic systems.
CoAgent: New Concurrency Control Protocol Cuts Multi-Agent LLM Conflicts by 40%
Multi-agent LLM systems face concurrency conflicts when multiple agents mutate shared state. A new protocol, MTPO, leverages LLM judgment to repair conflicts, achieving 1.4x speedup and near-serial correctness. The implementation, CoAgent, outperforms traditional locking and optimistic concurrency control.
New Study Measures Trust Between AI Agents, Revealing Formation, Breakage, and Recovery Dynamics
A preprint on arXiv introduces a behavioral measure to quantify trust between language-model agents using costly verification in a cooperative game. Testing six frontier model snapshots, the study finds that four models reduce verification by 60-85% when paired with reliable teammates, while trust recovery is slower than formation and clustered failures sustain suspicion longer. The results suggest that calibration, not maximal suspicion, should guide governance of multi-agent AI systems.
Early Diagnosis of Wasted Computation in Multi-Agent LLM Systems via Failure-Aware Observability
A research paper proposes a trace-based observability framework for multi-agent LLM systems that diagnoses wasted computation before final evaluation. On 165 GAIA traces, warned failed runs spent 58.1% of tokens after the first warning. A pilot using warnings reduced post-warning token fraction from 0.638 to 0.304, supporting a layered design with cheap online signals and deeper semantic checks.
EdgeCitadel: Hybrid NATS-MQTT Orchestration Platform for Edge Multi-Agent Systems
EdgeCitadel is an edge multi-agent orchestration platform built around a single NATS 2.10 server with an MQTT adapter. It combines MQTT connectivity, JetStream-backed persistence, direct peer delegation, and a passive aggregator. A testbed spanning ARM64, x64, and Android clients demonstrates the hybrid architecture.