ACED-Bench tests whether LLM agents can design experiments and reach causal conclusions

Key points

• Benchmark pairs LLM-written scenarios with hidden probabilistic graphical models that supply exact ground-truth answers.
• Main leaderboard ACED-CORE has 180 tasks; ACED-QUAL adds a 150-question yes/no qualification check.
• Seven contemporary LLMs and two earlier models were tested across three active agent designs.
• Reported failures include testing the wrong intervention, losing relevant evidence, and continuing after sufficient evidence appeared.

Why it matters for marketers: Content teams relying on AI for research summaries should note that agents may gather evidence poorly or stop too late, so outputs need human checks.