New JAIR Papers on Algorithmic Collusion, Social Choice, and Mereology

The latest issue of the Journal of Artificial Intelligence Research (JAIR) includes three new papers covering distinct areas of AI research.

Does an Intermediate Price Facilitate Algorithmic Collusion? Author Janusz Meylahn examines whether adding a middle price option in a duopoly influences reinforcement‑learning agents’ tendency to collude. Using decentralized Q‑learning with one‑period memory over up to fifteen possible prices, the study shows that while more collusive equilibria appear when a third price is available, the overall likelihood of collusion does not rise. The work builds on earlier evidence that RL agents can learn collusive strategies and highlights the algorithm’s convergence properties and basin calculations as key methodological strengths.¹.org]

Representative Social Choice: From Learning Theory to AI Alignment Tianyi Qiu proposes a representative social choice framework for aggregating preferences in large populations, such as jury trials or language‑model alignment. By framing deep questions as statistical learning problems, the paper applies generalization bounds from machine learning—e.g., Rademacher complexity—and extends Arrow’s Impossibility Theorem using combinatorial tools. This approach offers potential solutions for aligning language models with human values while acknowledging inherent challenges of democratic aggregation.¹.org]

Quantized Mereology Thomas Bittner investigates merging quantum information theory with classical mereology. The paper argues that crisp part–whole relations are captured by classical bits, whereas vagueness requires qubits, allowing entities to exist partially within multiple wholes simultaneously. Building on the RCC5 formalism, it demonstrates how qubit representations enable reasoning about vague spatial relationships like “partially overlaps” that were previously intractable under purely classical logic systems.¹.org]