RuleSafe-VL: Evaluating Rule-Conditioned Decision Reasoning in Vision-Language Content Moderation

Researchers introduced RuleSafe-VL, a new benchmark for evaluating how well vision-language AI models apply explicit content moderation rules. The benchmark reveals significant gaps in rule-reasoning capabilities, with even top models achieving only 64.8% accuracy on rule-interaction recovery, indicating current safety systems may reach correct moderation decisions through superficial pattern-matching rather than genuine policy understanding.

RuleSafe-VL addresses a critical blind spot in AI safety evaluation: current benchmarks measure whether models produce correct moderation labels, not whether they reason through policy rules correctly. This distinction carries substantial weight for deployment risk. A model could classify prohibited content correctly while failing to understand the underlying rule structure, creating brittle systems that break under novel scenarios or adversarial inputs. The benchmark formalizes 93 atomic rules and 92 rule relationships across three policy families, generating 2,166 test cases that isolate specific reasoning failures.

The experimental results expose troubling weaknesses across frontier models. Rule-interaction recovery emerged as the dominant bottleneck—understanding how multiple rules interact in complex cases—where even best-performing models reached only 64.8% accuracy. Some safety-oriented models dropped below 7%, suggesting specialized safety training may not translate to rule reasoning. Decision-state prediction, requiring models to judge whether available evidence suffices for a moderation decision, peaked at 64.5%, indicating models struggle with evidential reasoning.

These findings have immediate implications for platform deployment. Content moderation at scale requires systems that generalize to unforeseen cases and explain their decisions to human reviewers. Models that achieve high benchmark scores through superficial learning will fail this requirement. The diagnostic task decomposition—identifying activated rules, recovering interactions, judging sufficiency, resolving outcomes—provides a replicable evaluation framework that other domains could adopt.

Future research should investigate whether scaling, architecture changes, or specialized training can close these reasoning gaps. The benchmark establishes baselines but reveals that robust rule-conditioned reasoning remains an unsolved problem in vision-language AI.

• →Current safety benchmarks mask whether AI models apply moderation rules correctly or just match labels through pattern recognition.
• →Rule-interaction recovery is the critical failure point, with top models reaching only 64.8% accuracy on this task.
• →Safety-oriented models sometimes underperform frontier models at rule reasoning despite specialized training.
• →The benchmark's four diagnostic tasks enable targeted evaluation of specific reasoning failures in content moderation.
• →Improving rule-conditioned reasoning remains essential for deploying trustworthy content moderation systems at scale.