EvoCorps: An Evolutionary Multi-Agent Framework for Depolarizing Online Discourse

TL;DR

EvoCorps reframes online discourse governance as a dynamic, closed-loop multi-agent game that proactively mitigates polarization under adversarial amplification. It combines role specialized coordination, a retrieval-augmented cognition core, and evolutionary learning to adapt strategies over time within a mean-field MMDP. Implemented on the MOSAIC simulation platform with multi-source news streams, EvoCorps demonstrates improvements in emotional polarization, viewpoint extremity, and argumentative rationality compared with adversarial and post-hoc baselines. The work highlights the potential of in-process interventions for platform governance while acknowledging limitations of simulated environments and the need for human-in-the-loop validation and multimodal extensions.

Abstract

Polarization in online discourse erodes social trust and accelerates misinformation, yet technical responses remain largely diagnostic and post-hoc. Current governance approaches suffer from inherent latency and static policies, struggling to counter coordinated adversarial amplification that evolves in real-time. We present EvoCorps, an evolutionary multi-agent framework for proactive depolarization. EvoCorps frames discourse governance as a dynamic social game and coordinates roles for monitoring, planning, grounded generation, and multi-identity diffusion. A retrieval-augmented collective cognition core provides factual grounding and action--outcome memory, while closed-loop evolutionary learning adapts strategies as the environment and attackers change. We implement EvoCorps on the MOSAIC social-AI simulation platform for controlled evaluation in a multi-source news stream with adversarial injection and amplification. Across emotional polarization, viewpoint extremity, and argumentative rationality, EvoCorps improves discourse outcomes over an adversarial baseline, pointing to a practical path from detection and post-hoc mitigation to in-process, closed-loop intervention. The code is available at https://github.com/ln2146/EvoCorps.

Figures (7)

• Figure 1: Online interaction evolves from normal user communication to malicious agents spreading harmful information, leading to amplified negative emotions and polarization. Existing harm detection and post-hoc intervention (e.g., content removal or labeling) occur only after diffusion and are insufficient to mitigate polarization.
• Figure 2: Workflow of the closed-loop evolutionary system. In the perception and decision phase, an analyst agent monitors incoming posts and evaluates opinion stance, extremity, popularity, and sentiment to trigger warnings. A strategist agent then formulates core arguments, leader styles, and amplification scales. In the execution and amplification phase, the leader agent, supported by the USC mechanism, retrieves evidence, generates candidate contents, and selects optimal outputs via voting, while amplifier agents coordinate role-based identities to expand and propagate the content. In the feedback and iteration phase, an analyst agent evaluates intervention effects against baselines and produces structured reports, which are used by the strategist agent to guide strategy adjustment and support subsequent decision making.
• Figure 3: Sentiment trajectories over time under Case 1/2/3/4. The dashed line marks clarification at $t{=}5$. Case 2 continues to decline, Case 3 partially mitigates, and Case 4 declines more slowly and stabilizes relative to Case 2/3.
• Figure 4: Reward trajectories across intervention rounds. Reward is defined in Eq. \ref{['eq:reward']} from changes in sentiment and viewpoint extremity. The cumulative sum grows steadily, while the cumulative average rises in early rounds and then stabilizes, reflecting diminishing marginal gains as the discourse state becomes more stable.
• Figure 5: Proportion of Agent Comments.