EvoCorps: Proactive Discourse Moderation

• EvoCorps is an evolutionary multi-agent framework that proactively steers online discourse away from polarization.
• It employs specialized roles and a closed-loop intervention cycle to adapt and counter adversarial narratives in real time.
• Evaluations on the MOSAIC platform show enhanced sentiment stabilization and rationality compared to static moderation approaches.

EvoCorps is an evolutionary multi-agent framework developed for proactive depolarization of online discourse. Distinct from post-hoc or diagnostic governance methods, EvoCorps treats discourse moderation as a dynamic, closed-loop intervention task: it monitors, plans, and executes coordinated responses to evolving adversarial tactics in real time. This approach operationalizes intervention as a dynamic social game, wherein an organized team of agents steers conversational trajectories toward moderation and factual grounding before polarization peaks. The framework is instantiated on the MOSAIC social-AI simulation platform, with efficacy demonstrated across several key discourse metrics under adversarial and non-adversarial conditions (Lin et al., 9 Feb 2026).

1. Problem Formulation and Design Objectives

EvoCorps conceptualizes depolarization as a Stackelberg-style multi-agent game. The intervention team actively leads, shaping discourse evolution while users—including adversarial collectives—respond in a mean-field fashion. The central goal is to proactively steer collective sentiment ($e_t$) and viewpoint extremity ($v_t$) to moderate, evidence-based states before polarization emerges or is amplified.

Discourse occurs in a time-indexed post stream $\{\mathcal P_t\}_{t=1}^T$ over an underlying social graph. At each round $t$, the system observes $o_t = (\mathcal P_t, \mathcal C_t)$ (posts and comments). The ensemble user state is abstracted as $s_t = (v_t, e_t)$, with $v_t \in [0,1]$ (average viewpoint extremity) and $e_t \in [0,1]$ (aggregate sentiment).

Formally, the setting is a multi-agent Markov Decision Process (MMDP),

$$\mathcal M = \left(\mathcal N, \mathcal S, \{\mathcal A^i\}_{i \in \mathcal N}, P, R \right),$$

with agents $$\mathcal N = \{ \mathrm{Analyst}, \mathrm{Strategist}, \mathrm{Leader}, \mathrm{Amplifier} \}$$, control state $\mathcal S \equiv (v, e)$, joint action $a_t \in \prod_{i} \mathcal A^i$, and a reward function designed to decrease $v$ and increase $e$: $$R(s_t, a_t) = -\lambda_1 [v_{t+1} - v_t] + \lambda_2 [e_{t+1} - e_t].$$ An explicit adversarial model introduces malicious agents that inject extremized variants of news, employing coordinated multi-identity amplification to induce spikes in negative sentiment and extremity. Thus, EvoCorps must continually adapt as new adversarial frames emerge.

2. System Architecture and Component Roles

EvoCorps comprises four specialized roles, enacting a closed-loop intervention cycle:

1. Monitoring (Analyst): Extracts current mean-field metrics ($v_t, e_t$) from $o_t$ using like-weighted sentiment and extremity estimation. Triggers alerts when threshold levels of extremity or negative sentiment are exceeded.
2. Planning (Strategist): Receives alerts and consults an Action–Outcome Memory of past $(a_i, o_i, R_i)$ tuples, along with the current system state. Uses Tree-of-Thought structured reasoning to generate a coordinated intervention plan, including counter-argument narratives, rhetorical styles, amplifier role configuration, and timing strategies.
3. Grounded Generation (Leader): Retrieves candidate arguments from an Evidence Knowledge Base $K_t = \{ (k_i, p_i) \}$, where $k_i$ is a knowledge item and $p_i$ its persuasiveness score. Constructs multiple counter-message drafts, internally votes on persuasiveness, logic, and readability, and selects a final message.
4. Multi-Identity Diffusion (Amplifier): Implements diffusion via a set of distinct personas, each rephrasing and disseminating the leader’s core message to emulate authentic, community-level grassroots amplification and stylistic diversity.

A retrieval-augmented collective cognition core underpins the architecture:

• Evidence Knowledge Base ($K_t$): Items are retrieved from trusted sources if $\mathrm{relevance}(k', K_t) > \delta$ and are reinforced based on impact,

$$p_i \leftarrow p_i + \eta R(s_t, a_t) \mathbb{I}[k_i \in \mathcal K(a_t)].$$

• Action–Outcome Memory ($M_t$): Retains tuples only for interventions with $R_i > \epsilon$.

Data flows through the pipeline as $o_t \rightarrow s_t \rightarrow$ plan $\rightarrow$ message $\rightarrow$ diffusion $a_t \rightarrow$ environment $\rightarrow o_{t+1}$, with rewards feeding back to update $K_{t+1}$ and $M_{t+1}$.

3. Evolutionary Learning Mechanism

EvoCorps employs an evolutionary adaptation procedure rather than standard gradient-based optimization, directly evolving the knowledge base $K_t$ and memory $M_t$:

• Population: All knowledge items ($K_t$, as genotypes) and memory entries ($M_t$).
• A. Sampling: At each $t$, sample action $a_t \sim \pi(a \mid s_t, K_t, M_t)$.
• B. Fitness Evaluation: Fitness is determined by the reward:

$$F(s_t, a_t) = R(s_t, a_t) = -\lambda_1 \Delta v_t + \lambda_2 \Delta e_t.$$

Or, more broadly,

$$F = w_1 \Delta \mathrm{Pol} + w_2 \Delta \mathrm{Rat} - w_3 \Delta \mathrm{Ext},$$

with $$\Delta \mathrm{Pol} = \Delta(\text{Sentiment}) - \Delta(\text{Toxicity})$$, $\Delta \mathrm{Rat}$ incorporating Argument Quality Score (AQS), evidence usage, and fallacy rate, and $$\Delta \mathrm{Ext} = \Delta(\text{ViewpointExtremity})$$.

• C. Selection and Retention: Knowledge items used are reinforced: $p_i$ incremented by $\eta F$. If $F > \epsilon$, intervention is recorded in $M_{t+1}$.
• D. Variation: Mutation occurs by retrieving new $k'$ if $\mathrm{relevance} > \delta$. Crossover combines high-$p$ knowledge from multiple successful interventions.

Standard evolutionary operator analogs thus propagate high-fitness (effective) argument strategies and factual frames, while less effective knowledge items atrophy.

4. Simulation Environment and Empirical Results

EvoCorps is evaluated on MOSAIC, a social-AI simulation platform where generative agents interact in posting, liking, commenting, and sharing roles. Two primary datasets are ingested: NELA-GT-2021 and the COVID-19 Fake News collection. Adversarial news is synthesized by LLMs, with coordinated amplification by malicious agents. Factual corrections are delayed by four steps to simulate authentic reactive constraints.

Experiments comprise four regimes:

Case	Condition	Intervention
1	Benign, no adversary	None
2	Adversary only	None
3	Adversary + post-hoc fact-check	Static mitigation
4	Adversary + EvoCorps	Closed-loop intervention

Metrics include:

• Emotional polarization: Sentiment (mean mapped score, higher desired), toxicity (Perspective API, lower desired).
• Viewpoint extremity: Average extremity label.
• Argumentative rationality: Argument Quality Score (AQS), fallacy rate (% comments flagged), and evidence usage (% comments with verifiable attributions).

Observed results at $t = 30$:

• Higher sentiment: 39.2 (EvoCorps) vs. 29.0 (adversarial baseline)
• Lower extremity: 31.1 (EvoCorps) vs. 41.8
• Higher AQS and evidence usage, lower fallacy rate

Temporal plots indicate only EvoCorps achieves sentiment stabilization and significant suppression of extremity spikes.

5. Mechanistic Insights and Practical Considerations

Key findings include:

• Real-time adaptation: EvoCorps rapidly accommodates novel adversarial frames by evolving $K_t$ and $M_t$ without explicit retraining cycles.
• Role specialization: Ablation analysis attributes major improvements in sentiment and extremity to the Amplifier role (counter-amplification), while Strategist and Leader roles drive rationality gains. The Analyst stabilizes alert timing.
• Retrieval-augmented learning: The combination of fact retrieval and evolutionary memory mechanisms allows for persistent factual grounding and learning from prior interventions.

Potential extensions and operational pathways include:

• Multimodal incorporation (images/video into $K_t$),
• Human-in-the-loop moderation (calibration of LLM prompts via moderator input),
• Real-world deployment (platform-level APIs for injection, fact-grounding, downranking, all within transparent AI-labeled governance frameworks).

EvoCorps thus constitutes an adaptive, coordinated, and in-process governance layer for online discourse, surpassing the latency and rigidity of static, post-hoc policies by functionally shifting governance upstream—from late-stage mitigation to proactive, evolutionary intervention (Lin et al., 9 Feb 2026).