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

Researchers introduce Moral Trolley Arena, a new benchmark that measures how large language models compose multiple moral considerations into unified judgments. Testing ten frontier models reveals that composite moral reasoning follows compressed, non-additive patterns rather than simple addition of component moral signals.

The research addresses a critical gap in current LLM evaluation: existing moral benchmarks measure isolated preferences but fail to capture how models actually combine competing moral considerations in realistic scenarios. This matters because real-world ethical decisions rarely involve single moral factors—they require integrating multiple values simultaneously. Moral Trolley Arena uses a two-stage methodology that first calibrates individual moral acts across foundational ethics frameworks, then combines them to measure composite reasoning patterns.

The findings reveal concerning inconsistencies in how frontier models process moral composition. Rather than additively combining moral signals, models exhibit "compression" where composite judgments fall systematically below what component strengths would predict. Additionally, models show intensity-dependent anchoring effects and foundation-specific biases that persist after controlling for component values. These non-linearities suggest models lack robust compositional reasoning for ethics.

For the AI industry, these results highlight that moral audits based on isolated preference rankings provide false confidence about model alignment. A model scoring well on individual moral benchmarks may still fail at ethically complex decisions requiring simultaneous consideration of competing values. This compounds risks in high-stakes deployments where models must navigate genuine moral tradeoffs.

The convergence of distorted composition patterns across different providers suggests these aren't isolated implementation quirks but fundamental limitations in current training approaches. Future work should investigate whether composition failures reflect architectural constraints or training data limitations, and whether targeted interventions during model development can improve moral reasoning robustness.

• →Frontier LLMs combine moral considerations through compressed, non-additive functions rather than simple addition of component signals.
• →Models exhibit systematic intensity anchoring and foundation-specific residuals indicating compositional reasoning failures.
• →Current moral benchmarks measuring isolated preferences provide incomplete assessment of actual ethical reasoning capabilities.
• →Moral composition patterns converge across different model providers, suggesting fundamental rather than idiosyncratic limitations.
• →Enhanced moral auditing frameworks should measure composition rules alongside isolated preference rankings.