aiqbench analysis
This research introduces a principled gating mechanism for multi-agent reinforcement learning where communication occurs only when the KL divergence between agent belief distributions exceeds a specific threshold, replacing unstable REINFORCE-based policies with interpretable information exchange criteria. Evaluations on Predator-Prey and MPE benchmarks demonstrate that this approach significantly reduces variance in coordination tasks while maintaining competitive success rates compared to standard continuous or binary gating methods. The method also yields improved latent representations even when the gate remains inactive, highlighting its dual benefit of efficient communication and enhanced internal state estimation for multi-agent systems.
When to Communicate: Belief Distributions and KL Divergence for Principled Gating in Multi-Agent RL
arXiv:2608.14559v1 Announce Type: cross Abstract: Effective communication in multi-agent reinforcement learning requires agents to decide not only \textit{what} to communicate, but when? Existing approaches either communicate at every timestep or learn a binary gate through REINFORCE policy gradients \cite{singh2019}, a high-variance signal that produces unstable and uninterpretable gating behavior. I propose a principled alternative: agents communicate only when the KL divergence between their learned belief distributions exceeds a fixed threshold. Each agent maintains a belief distribution over a latent world state computed as a softmax over its LSTM hidden state, and communicates only when belief disagreement is large enough to justify information exchange. I evaluate this approach on the Predator-Prey benchmark from IC3Net \cite{singh2019} across two environment sizes with 5 seeds each, and on MPE simple\_spread \cite{lowe2017}, comparing against IC3Net, CommNet, and an independent controller. On PP 10$\times$10, IC3Net outperforms KL-belief at all thresholds. On the harder PP 20$\times$20, a threshold ablation over $\varepsilon \in \{0.1, 0.3, 0.5, 1.0\}$ reveals an inverted U-shape: $\varepsilon=0.5$ achieves 73.84 average steps and 42\% success rate versus IC3Net's 75.31 steps and 31\%, a gap of 1.47 steps and 11 percentage points with tighter seed variance. On MPE, the belief head improves mean reward by 12 points and reduces variance by 26$\times$ even when gating is inactive, suggesting two orthogonal contributions: principled gating when beliefs can converge, and improved latent representations that benefit coordination regardless.
Abstract:Effective communication in multi-agent reinforcement learning requires agents to decide not only \textit{what} to communicate, but when? Existing approaches either communicate at every timestep or learn a binary gate through REINFORCE policy gradients \cite{singh2019}, a high-variance signal that produces unstable and uninterpretable gating behavior. I propose a principled alternative: agents communicate only when the KL divergence between their learned belief distributions exceeds a fixed threshold. Each agent maintains a belief distribution over a latent world state computed as a softmax over its LSTM hidden state, and communicates only when belief disagreement is large enough to justify information exchange. I evaluate this approach on the Predator-Prey benchmark from IC3Net \cite{singh2019} across two environment sizes with 5 seeds each, and on MPE simple\_spread \cite{lowe2017}, comparing against IC3Net, CommNet, and an independent controller. On PP 10$\times$10, IC3Net outperforms KL-belief at all thresholds. On the harder PP 20$\times$20, a threshold ablation over $\varepsilon \in \{0.1, 0.3, 0.5, 1.0\}$ reveals an inverted U-shape: $\varepsilon=0.5$ achieves 73.84 average steps and 42\% success rate versus IC3Net's 75.31 steps and 31\%, a gap of 1.47 steps and 11 percentage points with tighter seed variance. On MPE, the belief head improves mean reward by 12 points and reduces variance by 26$\times$ even when gating is inactive, suggesting two orthogonal contributions: principled gating when beliefs can converge, and improved latent representations that benefit coordination regardless.
Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2608.14559 [cs.AI]
(or arXiv:2608.14559v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2608.14559 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 320 words
Related coverage
DI Series: Serving Kimi-K3 MXFP4 Wide-EP Disaggregated on AMD Instinct MI300X and MI325X
AMD · Oct 8, 2026
Game-Theoretic Disempowerment: Loss of Control to Agencyless AI
LessWrong · Sep 27, 2026
Angles of attack for continual learning safety
LessWrong · Jun 16, 2026
A Solution to Cryptographic Boxes for Unfriendly AI
LessWrong · Jul 19, 2026