ABAS Simulator Tests Deliberative Polling Against Strategic Attacks

A team from multiple institutions (Alssadi, Alawaji, Kasula, Syed, Alfurhood, Zanker, Silaghi) has released a paper on arXiv (2606.11692) introducing ABAS—an LLM-based Agentic Bipolar Argumentation Simulator for evaluating alternative-based information systems in deliberative polling. The simulator formalizes a poll as a six-tuple of endorsing/opposing justifications, attack/enhance relations, and weighted shareholders. It spawns N autonomous agents, each with a latent opinion drawn from a desired distribution, who sequentially vote, author or choose justifications, and optionally submit argumentation links. The system then recommends K justifications ranked by observable endorsement mass, aiming to maximize coverage—the fraction of the total reason‑tag set represented in those recommendations. This addresses the NP‑hard Subsuming Justification Problem.

In controlled experiments, the team varied creativity rate (p_own), recommendation size (K), argumentation density (p_links), and population size (N) to measure effects on coverage and corpus diversity. Under an authenticated electorate (no Sybil attacks, only relation graph gameable), they stress‑tested scoring with coordinated strategic voting. A tag‑flood attack—where agents flood the system with many shallow justifications—collapses coverage dramatically. However, using author‑count relation weighting via a reversed‑PageRank rule resists the flood markedly better than uniform weights, suggesting a robust design principle for real‑world deliberative polling platforms.