Meta is releasing 10 tasks from WildArtifactBench, an internal evaluation framework designed to measure the…

Meta is releasing 10 tasks from WildArtifactBench, an internal evaluation framework designed to measure the practical utility of multimodal agents.

WildArtifactBench evaluates agents on complex, real-world tasks across different deliverable formats.

Instead of relying only on strict ground-truth answers, Meta says the framework uses:

• Win rates

• Elo scores

• Judgments from both human and agentic preference evaluators

First posted to our Telegram.