Source-linked AI summary
Causal Discovery in the Era of Agents
Yujia Zheng, Vishal Verma, Mantej Gill, Haoyue Dai, Peter Spirtes, Kun Zhang
TL;DR
LLM-based causal discovery can blur whether graph claims come from data and assumptions or from textual associations and hallucinated mechanisms. The paper proposes causal-learn+, where agents assist with workflow tasks while formal algorithms produce the causal evidence, and a Big Five case study illustrates an accessible, transparent pipeline that preserves this grounding.
Problem
LLM-based causal discovery can make a graph’s evidential status ambiguous by mixing data, algorithmic assumptions, textual associations, prompt wording and hallucinated mechanisms.
Method
causal-learn+ coordinates agent assistance with explicit assumptions, user decisions, diagnostics and formal algorithms that alone compute causal discovery tests, searches, orientations and graph outputs.
Results
The Big Five case study illustrates an accessible and transparent causal discovery workflow whose resulting graph remains grounded in data, assumptions, procedures, diagnostics and domain knowledge.
Takeaways & Limitations
Agents should assist causal discovery, while algorithms discover and scientists judge, so agent fluency improves access and interpretation without creating causal evidence.
Takeaways & Limitations
The paper identifies hidden agent actions such as dropping variables, changing preprocessing or altering constraints as decisions that must remain visible and user-approved.
Abstract
from arXiv · showhide
Recent attempts to combine large language models (LLMs) with causal discovery ask models to infer pairwise directions, propose graph structures, or inject language-model outputs as priors and constraints. These approaches promise faster analysis, but they also obscure whether a causal evidence is supported by data and assumptions or by textual associations, prompt artifacts and hallucinated mechanisms. We argue for a different role for agents in causal discovery. Agents should inspect data, retrieve context, explain method assumptions and clarify graph outputs, but they should not supply edges, orientations, priors, constraints or causal conclusions. We propose the principle that agents assist the workflow, while causal claims remain grounded in data, explicit assumptions, formal algorithms, diagnostics and user or domain-expert decisions. We instantiate this principle in causal-learn+, an online platform that coordinates data analysis, preprocessing, method recommendation, expert-knowledge incorporation, formal discovery and interpretation around the algorithmic ecosystem of causal-learn. A case study on Big Five personality data illustrates agent-assisted pipeline of causal discovery without turning language-model unreliability into causal evidence. The platform is available at causallearn.com.
1 Introduction
Causal discovery turns observational data into inspectable causal hypotheses when interventions are costly or impossible, but its assumptions make analysis difficult and LLM-generated causal evidence unreliable. The paper therefore positions agents as workflow assistants—not sources of causal structure—and instantiates this principle in causal-learn+.
- Motivation: Causal discovery learns causal relations from observational data, which are often more obtainable than interventions that are expensive, slow, unethical, or impossible.It supports structured causal hypotheses that scientists can inspect, challenge, and refine.
- Challenges: Practitioners must manage scientific context, variable inclusion, missingness, algorithm choice, partially identified graphs, and limits on causal interpretation.These demands arise because causal discovery depends on assumptions and are especially difficult for users without specialized causal-methodology training.
- LLM limitations: LLM outputs used for directions, graphs, priors, constraints, or refinements can conflate dataset evidence with textual associations, prompt wording, common beliefs, or hallucinated mechanisms.Because language models predict patterns in text, their outputs may obscure the basis of a discovered relation when injected into graph construction.
- Proposed principle: Agents should assist causal discovery, not replace it, by coordinating tools, retrieving context, clarifying assumptions, supporting method selection, and helping interpret outputs.Users retain control over decisions affecting final conclusions, while agents remain embedded in a rigorous causal-analysis pipeline.
- System contribution: causal-learn+ implements this principle as an online agentic system around causal-learn, broadening access while preserving causal conclusions grounded in reliable analytical evidence.The system treats agents as assistants to causal learning rather than replacements for discovery algorithms.
2 Agent Cannot Provide Causal Evidence
Agents can assist causal-discovery workflows, but their outputs are not causal evidence. Causal claims must remain grounded in data, explicit assumptions, formal procedures, diagnostics, and visible user or expert decisions.
- Evidence boundary: Agent outputs such as literature summaries, temporal-order suggestions, mechanism recognition, and explanations can guide users but do not constitute data-based causal evidence.Their usefulness creates the risk that users mistake coherent assistance for evidence.
- Evidence boundary: If agents propose edges, orientations, priors, constraints, or thresholds, the resulting graph no longer reflects a well-specified discovery method alone.Its relations may instead reflect data, algorithmic assumptions, pretrained beliefs, prompt wording, or hallucination.
- Interpreting graph outputs: Fluent explanations can overstate partially identified outputs: CPDAGs represent Markov equivalence classes, while PAGs represent models that may include latent confounders and selection variables.Latent nodes are not automatically named scientific constructs, and Granger relations are not automatically intervention claims.
- Workflow decisions: Hidden workflow actions—including dropping variables, preprocessing, adding forbidden edges, changing temporal tiers, lowering rank thresholds, or orienting edges—alter algorithmic inputs or outputs.These actions should be visible user decisions rather than silent agent behavior.
- Operating principle: Graph-structure evidence should come from observed data, explicit assumptions, formal algorithms, diagnostics, and user or expert decisions, with agent outputs restricted to context, guidance, or explanation.Provenance is part of what a graph mark means; agent inputs enter the formal core only when explicitly adopted as assumptions or interpretations.
3 How Agents Should Help Causal Discovery
Agents should assist causal discovery by simplifying, explaining and documenting the workflow while leaving causal conclusions grounded in formal algorithms, explicit assumptions and user judgment. Their suggestions, choices, settings and outputs should remain distinct, visible, traceable and reversible so assistance does not become causal evidence.
- Data understanding: Agents can inspect data by summarizing variables, detecting missingness, flagging unusual distributions and identifying possible identifiers or time stamps before discovery begins.They can also call visualization or statistical tools to help users notice problems before method selection or interpretation.
- Preprocessing: Agents can explain preprocessing alternatives and run checks, but consequential transformations must remain visible, deliberate user decisions rather than hidden assistant actions.Scaling, discretization, missing-data handling, variable exclusion and item recoding can all affect the discovered graph.
- Method and knowledge support: Agents can translate scientific questions into methodological considerations, recommend suitable algorithm families and incorporate expert knowledge as context without directly creating edges, orientations or constraints.Users or domain experts should explicitly approve any role for retrieved or model-provided knowledge in the analysis.
- Tool coordination: Agents can coordinate analysis tools and execute selected algorithms, but must not alter the tests, scores, constraints, search, orientation rules or graph files defining formal discovery.Tool coordination is distinct from causal judgment.
- Interpretation and system principles: Agents can explain graph notation, assumptions and domain vocabulary, communicate results and clarify limits without converting an assumption-conditioned graph into a causal fact.The overall design keeps agents around, rather than inside, the inferential core, with suggestions, user choices, settings and outputs exposed as distinct objects.
4 causal-learn+: Agentic Causal Discovery with Trust
causal-learn+ is an online agentic environment that coordinates data analysis, preprocessing, method recommendation, expert-knowledge incorporation, formal discovery and interpretation around a protected algorithmic core. Agents improve access and understanding, while causal evidence remains produced by explicit algorithms and accepted user or expert decisions.
- Workflow design: causal-learn+ coordinates the causal-discovery workflow while algorithms from causal-learn perform the discovery step itself.Agents do not infer causal relations, preserving the evidential semantics of the underlying methods.
- Workflow design: Agents inspect uploaded data by summarizing variables, missingness, data types, sample size, distributions and simple relationships.This helps users assess dataset readiness and identify variables requiring attention.
- Workflow design: Agents suggest preprocessing choices and present candidate method families with assumptions and diagnostics for user selection, but these suggestions are not causal evidence.Users may accept scaling, missing-data handling, exclusion or recoding decisions; recommendations include constraint-based, score-based, functional, latent-variable and time-series methods.
- Interpretation and expertise: Agents provide contextual knowledge and explain graph marks, diagnostics and method outputs without automatically creating graph priors or changing what algorithmic outputs license.Domain constraints require explicit approval by the user or domain expert, while reports state assumptions behind graph-level statements or interpretations.
- Formal discovery: Formal causal discovery is the protected core, with conditional-independence tests, scores, rank constraints, graph search, orientation rules and graph outputs computed by explicit algorithms.No LLM or agent participates in these calculations, which remain tied to the selected data, assumptions, parameters and diagnostics.
- Access: The online platform at causal-learn.com lowers the threshold for causal discovery by avoiding package installation, local configuration and coding while preserving formal algorithms and explicit user decisions.Users can conduct the workflow online without writing code for every step.
5 Real-world Example
The Big Five case study shows how causal-learn+ supports latent-variable causal discovery while keeping causal claims grounded in data, assumptions, algorithms, diagnostics, and domain expertise. Agents assist with data understanding, preprocessing, method selection, and interpretation without supplying causal evidence or conclusions.
- Data and latent structure: The dataset contains 50 standardized observed indicators from nearly 20,000 complete responses, requiring decisions about question meanings, missingness, reverse wording, and scaling.Responses use a five-point agreement scale, and missing values are removed before analysis.
- Data and latent structure: Latent personality dimensions are inferred from response patterns, so standard observed-variable methods may obscure measurement structure and yield difficult-to-interpret mechanisms.Hidden variables may resemble known dimensions, combine question groups, or have semantics that remain unresolved.
- Causal discovery results: RLCD handles partially observed models with causally related hidden variables, while GIN helps orient relations between hidden variables in the resulting graph.The graph groups many indicators associated with the same Big Five dimension near common hidden variables, consistent with existing psychological knowledge.
- Agent-assisted workflow: Agents can summarize questions, flag preprocessing issues, explain method choices, retrieve measurement context, and translate graph marks and latent-variable caveats.They should present structures such as a cluster resembling conscientiousness or O2 →O4 as hypotheses for domain review, not confirmed traits or mechanisms.
- Agent-assisted workflow: Causal-learn+ makes the workflow accessible and transparent while keeping the graph grounded in data, assumptions, discovery procedures, diagnostics, and domain knowledge.Agents assist users but do not add orientations, remove responses, or tune rank thresholds to match expected Big Five structure.
6 Discussion
Agents should assist causal-discovery workflows without creating causal evidence: graph marks remain grounded in data, assumptions, formal procedures, diagnostics, and scientific decisions. Future work should operationalize this standard through benchmarks, audit trails, expert approval interfaces, and evaluations of interpretation.
- 6 Discussion: Causal evidence remains grounded in data, assumptions, formal procedures, diagnostics, and scientific decisions rather than agent fluency.Agents can help people reach and understand evidence, but fluency cannot create it.
- 6 Discussion: Agents assist, algorithms discover, and scientists judge, while the inferential core remains explicit, inspectable, and tied to formal methods and user-approved assumptions.Agents can support data understanding, preprocessing, method choice, expert context, tool use, and interpretation.
- 6 Discussion: Future work should operationalize this standard with benchmarks, audit trails, expert approval interfaces, and evaluations of whether agent guidance reduces overinterpretation.The proposed infrastructure separates data, assumptions, agent recommendations, user decisions, and algorithmic outputs.