CausalDS: Benchmarking Causal Reasoning in Data-Science Agents
CausalDS is a benchmark testing if AI agents can perform causal reasoning in data science workflows. It creates realistic scenarios with causal models, synthetic stories, and noisy data to challenge models on symbolic reasoning, coding tools, uncertainty, and when to abstain from answering. The research shows that robust agents solve the formal problem rather than just mimicking the story they are given.