Tokens per task versus CausalDSScore (lower is better): the frontier (closed) and open-weight models are linearly separable in the (log) cost–quality plane.Abstract
Large language models (LLMs) increasingly act as integrated data-science agents, combining abstract reasoning with advanced tool use. Yet the relevant benchmark landscape largely divides into symbolic causal reasoning benchmarks without realistic data analysis or data analysis benchmarks without a principled causal data-generating structure. Furthermore, existing causal evaluation datasets are often restricted to curated examples from existing sources, with diversity coming from limited templatized variations rather than from systematic generation of novel synthetic causal structures. We introduce CausalDS, a benchmark for evaluating causal reasoning in agentic data-science workflows. Each benchmark instance is a scene consisting of a sampled structural causal model (SCM) with generated observational data and an accompanying synthetic natural-language story grounded in a realistic domain. We optionally ground the composition of the benchmark components in empirical distributions obtained from real-world datasets, thus retaining empirical structure while reducing the “causal parrot” risk through completely synthetic generation. From each scene, we then derive tasks spanning all three of Pearl’s rungs, with typical data-science prediction tasks appearing as Rung 1. Most tasks include a data science coding component, where the model typically needs to use several tools to arrive at the final answer due to the frequent presence of imperfect observations, which are generated by an observation model. Additionally, recognizing when a question admits no warranted answer and abstaining is treated as a first-class scored outcome. The benchmark thus jointly evaluates symbolic causal reasoning, data science, uncertainty quantification, abstention, and tool use/coding.
The paper introduces a benchmark generator for causal reasoning in agentic data-science workflows: every benchmark instance is synthetically generated rather than curated. In it, we:
- Generate fully synthetic scenes pairing a sampled causal graph and structural causal model with tabular data and natural-language story;
- Add an optional observation layer that varies the data-analysis difficulty without changing the causal side;
- Derive tasks spanning all three rungs of Pearl’s hierarchy which include non-answerable questions (i.e., testing absention);
- Make the benchmark fully parameterizable, so that it can be tailored to specific evaluation goals or grounded in empirical distributions obtained from real-world corpora.
We thus test the agents along five separate axes of agentic causal data science: symbolic causal reasoning, data science, uncertainty quantification, abstention, and tool use/coding.
We evaluate six contemporary agents — Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5, along with the open-weight Qwen 3.6 35B, Kimi K2.6, and Gemma 4 26B — on a 100-scene exam with realistically grounded composition. Claude Opus 4.8 leads overall, and the evaluated capabilities dissociate: symbolic causal reasoning is largely mastered across the field, while abstention, uncertainty quantification, and coding/tool-use/reasoning efficiency separate the models.
