Causal Cartographer: From Mapping to Reasoning Over Counterfactual Worlds

Abstract: Causal world models are systems that can answer counterfactual questions about an environment of interest, i.e. predict how it would have evolved if an arbitrary subset of events had been realized differently. It requires understanding the underlying causes behind chains of events and conducting causal inference for arbitrary unseen distributions. So far, this task eludes foundation models, notably LLMs, which do not have demonstrated causal reasoning capabilities beyond the memorization of existing causal relationships. Furthermore, evaluating counterfactuals in real-world applications is challenging since only the factual world is observed, limiting evaluation to synthetic datasets. We address these problems by explicitly extracting and modeling causal relationships and propose the Causal Cartographer framework. First, we introduce a graph retrieval-augmented generation agent tasked to retrieve causal relationships from data. This approach allows us to construct a large network of real-world causal relationships that can serve as a repository of causal knowledge and build real-world counterfactuals. In addition, we create a counterfactual reasoning agent constrained by causal relationships to perform reliable step-by-step causal inference. We show that our approach can extract causal knowledge and improve the robustness of LLMs for causal reasoning tasks while reducing inference costs and spurious correlations.

• The paper introduces a two-stage framework using CTG-Extract to build causal graphs and CTG-Reason for structured, stepwise counterfactual inference.
• It demonstrates up to a 70% reduction in inference cost and improved numerical reasoning by enforcing local causal constraints.
• Empirical results across multiple LLMs validate enhanced robustness and practicality for real-world decision-making and policy analysis.

Causal Cartographer: Mapping and Reasoning Over Counterfactual Worlds

This paper introduces Causal Cartographer, a two-stage framework for explicitly extracting, representing, and reasoning over causal relationships in natural language, with a focus on addressing limitations in current LLMs for real-world counterfactual inference. The methodology allows the construction of large-scale causal knowledge graphs from naturalistic, unstructured sources—most notably, news media—and provides a structured approach for reliable counterfactual evaluation by LLM agents.

Overview of Framework

The proposed system comprises two primary agent modules:

1. Causal Extraction Agent (CTG-Extract): Employs a retrieval-augmented generation (graphRAG) approach to extract entities and causal relationships from collections of news documents. The extracted information is stored as a network of typed variables and directed causal links, each annotated with semantic attributes, provenance, and document-level world grounding. This enables not only aggregation but disambiguation and robust mapping of concepts as they recur across differing contexts.
2. Counterfactual Reasoning Agent (CTG-Reason): Performs step-wise causal reasoning over the constructed graph, leveraging structure-aware inference schemes that strictly adhere to identified causal pathways. This agent reduces dependency on global context and minimizes confounding effects from irrelevant correlations—directly addressing a major limitation of general LLM-based causal reasoning.

CausalWorld Network Construction

A notable contribution is the creation of CausalWorld, a dataset of nearly one thousand causal variables and over 1,300 edges, extracted from real-world news events with a specific thematic focus (oil prices, green energy, macroeconomic events). The network exhibits pronounced sparsity and community structure, matching expectations from causal structure learning (Sparse Mechanisms Shift hypothesis), and is intentionally constructed from overlapping "worlds" (document-derived instantiations).

Key engineering choices include:

• World Grounding: Each causal node is instantiated across multiple world contexts, permitting empirical construction and matching of counterfactual pairs via K-matching—facilitating factual-to-counterfactual comparisons with some identifiability guarantees.
• GraphRAG for Variable Matching: The use of neural embeddings and graph traversal for efficient variable matching avoids many pitfalls of string-based or pattern-matching approaches in prior work, ensuring robust mapping of synonymic or dependent phenomena.

Figures provided in the paper further clarify the structural diversity and inter-connectivity of the CausalWorld network, highlighting overlapping ownership of nodes by worlds, existence of cycles, and bridge nodes critical for information propagation.

Step-by-Step Causal Reasoning and Evaluation

The reasoning pipeline moves beyond pure prompt-based (chain-of-thought) strategies by enforcing strict adherence to the local Markov blanket (generalized as "causal blanket") for any inferential step, thus bounding the agent’s information set. The inference agent executes:

• Abduction: Infers exogenous variables by traversing the graph in anticausal (children-to-parent) fashion.
• Intervention: Introduces $do$-operator-based interventions, manipulating target variables and appropriately modifying the local graph structure.
• Prediction: Proceeds recursively in a causally disciplined step-wise manner, only accessing necessary ancestral information as dictated by the graph structure.

Ground-truth for counterfactual inference is established by empirical matching (K-matching) between worlds, with a theoretical argument that if a causal blanket is fully matched, the resultant estimand is equivalent to a true counterfactual in the underlying SCM. This methodological innovation enables direct, evaluation-ready datasets from naturally occurring data, overcoming the classic roadblock of missing counterfactuals in observational settings.

Experimental Results and Numerical Claims

Experiments benchmark several LLMs (OpenAI o3-mini, GPT-4.1, and LLaMA-3.1-8B) on both synthetic and real-world subsets of the CausalWorld Counterfactual Reasoning (CausalWorld-CR) dataset, comparing the step-by-step agent to prior standard CausalCoT (Chain-of-Thought) methods. Notable findings include:

• On boolean and trend queries, step-by-step causal conditioning matches or slightly outperforms chain-of-thought, and achieves up to 70% reduction in inference cost (context window and output length), directly benefiting efficiency for small models.
• Numerical query performance reveals heightened robustness for the stepwise approach, particularly with less capable LLMs; strict adherence to causal constraints yields lower variance in outcome estimation and improves prediction magnitude error.
• The CTG-Reason approach enables reliable completion of tasks where CausalCoT fails (notably with LLaMA-3.1-8B), demonstrating substantially higher robustness to prompt and context window limitations.

These results serve to empirically validate the claim that explicit, stepwise causal reasoning—anchored in a real-world, structured causal graph—improves both the reliability and computational efficiency of LLM-based counterfactual inference.

Limitations and Practical Considerations

The paper acknowledges several important limitations:

• Order Dependence: Iterative extraction implies document order can influence final graph structure; robust, order-agnostic incremental updates would require further work.
• Graph Completeness: The method assumes access to a sufficiently complete causal graph for the target queries. Incompleteness or errors in extraction propagate to inference failures—a significant concern for deployment at web scale.
• Data Truthfulness: The reliance on news data and potential for adversarial/misinformation contamination introduces unavoidable risks to the integrity of downstream inferences.
• Numerical Reasoning Biases: LLMs’ well-known deficiencies in fine-grained numerical prediction persist, though the structure-enforced agent partially mitigates spurious reasoning.

For practical large-scale applications, these limitations necessitate careful source validation, disciplined monitoring of extraction pipelines, and potentially hybrid human-in-the-loop curation at critical junctions.

Broader Implications and Future Directions

The framework demonstrates how explicit causal knowledge extraction and graph-structured reasoning can bridge the gap between contemporary LLM capabilities and the requirements of robust, trustworthy decision-making systems operating in open-world natural language domains.

Future avenues of research suggested by this work include:

• Extending to multimodal causal extraction (e.g., integrating images, tables, or time series).
• Enabling dynamic, online updating of causal graphs in light of new evidence, with mechanisms to handle adversarial noise.
• Application to high-stakes settings (policy analysis, epidemiology, risk assessment), with additional layers of verification and interpretability.
• Incorporation into curricula for deliberate training or fine-tuning of LLMs explicitly on counterfactual and intervention tasks, using the constructed datasets as a scalable benchmark.

Conclusion

Causal Cartographer operationalizes the long-held desideratum of robust, transparent, and empirically evaluable causal reasoning over real-world natural language data. By explicitly extracting, structuring, and exploiting causal relationships, the framework advances LLM-based agents closer to reliable, context-aware general reasoning and opens new paths for scalable deployment in practical, high-impact domains.