Research internship at Signal Labs, focused on causal structure learning outside the setting where most of it is benchmarked: real observational data, with no ground-truth graph to score against.

What I worked on

Causal discovery on real data. I applied constraint-based and score-based causal discovery algorithms to real-world observational datasets, and looked at the failure modes that do not show up on synthetic benchmarks. Three recur: there is no ground-truth graph, so the usual structural metrics are unavailable; the identifiability assumptions the algorithms rely on (causal sufficiency, faithfulness) are difficult to justify and often violated; and the recovered structure is unstable, changing noticeably across resamples and hyperparameter choices.

LLM-elicited causal priors. I used large language models as a source of causal background knowledge — variable semantics, plausible relations between variables, and edge orientations — and encoded that knowledge as priors and constraints for the discovery algorithms. The motivation is the ambiguity data alone cannot settle: observational data identifies a causal structure only up to an equivalence class, leaving edges unoriented, and background knowledge is the standard way to narrow it. LLMs are an unusual source for that knowledge, and how far it can be trusted is an open question.

Pipeline and tooling. I built an end-to-end pipeline covering dataset preprocessing, algorithm execution, and structure evaluation, so that experiments could be run reproducibly across combinations of datasets, algorithms, and prior configurations.