VLMs Outperform Bluesky Moderation System on ModerationBench Benchmark

Vision-Language Models using instruction- or example-driven guidance can triple the F1 score of existing production systems, reaching 0.60 compared to Bluesky's 0.22 on the new 4,000-post ModerationBench. Both reasoning from policy precepts and generalizing from precedents show high effectiveness.