Every Act Has Its Price: Compressed Moral Composition in Frontier LLMs

The paper introduces the **Moral Trolley Arena**, a two-stage benchmark designed to evaluate how large language models (LLMs) compose moral judgments from multiple signals, moving beyond traditional isolated act assessments. It features a calibration phase with a 229-scenario corpus grounded in Moral Foundations Theory, followed by a composite evaluation that combines calibrated acts into two-act moral items. Findings indicate that while composite judgments in ten frontier models are largely predicted by individual act strength, they exhibit compressed relationships and non-additive intensity anchoring, suggesting a need for moral audits to focus on the rules of moral evidence composition rather than isolated act rankings.