Causals: An Open Knowledge Graph for Research on Causal Reasoning

Miaozhe Lin, Qiongzheng Lin, Zhongwei Liang, Zijun LiuarXiv 2026 · arXiv

TL;DR

Researchers have introduced Causals, a large-scale, open knowledge graph containing over 369,000 nodes and 1.2 million edges derived from news reports. Designed to bridge the gap between "causal parrots" (models that merely repeat associations) and true causal reasoning, this resource provides the structural context—linking events to specific companies, commodities, and markets—necessary for Large Language Models (LLMs) to trace the real-world consequences of new events.

What Problem Does This Paper Address?

Large Language Models are surprisingly good at identifying causal relationships they’ve seen thousands of times in their training data. However, they often struggle to reason about novel events. If a major shipping lane is blocked today, an LLM needs to know which specific companies rely on that route, what commodities they move, and which downstream markets will be affected.

The challenge is that this information is scattered across thousands of news reports. Existing causal resources are often either too abstract (dealing with general concepts like "smoking causes cancer") or entirely synthetic. There is a lack of high-quality, entity-linked relational data that captures the "plumbing" of the global economy—the structural relationships through which an event's consequences actually unfold.

Methods

The authors developed a pipeline called NewsAgent to construct the graph:

  1. Extraction: LLMs identify entities, events, and relationships within news text, constrained by a specific extraction vocabulary.
  2. Entity Reconciliation: To ensure "Apple Inc." in one report is the same as "Apple" in another, the system uses bge-m3 embeddings for candidate retrieval combined with exact-name lookups and model-assisted resolution for ambiguous cases.
  3. Publication Filtering: To ensure privacy and safety, the team removed over 82,000 "Person" nodes, focusing the initial release on organizations, locations, and commodities rather than private individuals.
  4. Semantic Representation: The graph distinguishes between different types of edges. For example, PRODUCES or CONSUMES represents a structural relationship, while TRIGGERED or STATED_CAUSE represents an explicit causal claim extracted from the text.

Datasets

The primary contribution is the Causals v0.1.0 dataset. The paper situates Causals among other major benchmarks in the field:

Table 1. Complementary roles of selected resources. Sizes refer to the cited releases or paper, not current live services.

Unlike CauseNet, which focuses on web-scale concept relations, Causals focuses on specific, news-derived entities and events.

Key Insights and Ideas

  • Structural vs. Causal Edges: A key "aha moment" is the realization that explicit causal claims (e.g., "Event A caused Event B") make up only 1.23% of the graph. The vast majority of the graph (over 41%) consists of structural relationships like SUPPLIES or INPUT_TO. The authors argue that these structural links are the "routes of influence" that models must traverse to perform reasoning.
  • The Neighborhood Explosion: The graph is highly connected. While a single node might only have ~13 neighbors (1 hop), by 3 hops, the average reachable neighborhood expands to over 4,400 nodes.
  • Separating Knowledge from Mechanism: The authors emphasize that the graph provides the knowledge (who is connected to whom), but not the mechanism (how a disruption propagates). A buyer's production only stops if their inventory is low—information that must come from the reasoning model or additional context, not just the graph structure itself.
Figure 1. A neighborhood in Causals v0.1.0-2026-10-09. Labels and arrow directions reproduce the stored relationships.

Key Results

  • Connectivity: The graph is remarkably cohesive; the largest weakly connected component contains 91.33% of all released nodes.
  • Retrieval Properties: The study measured how quickly a search expands from a starting entity. As shown in the reachability analysis, the median number of nodes reached jumps from 1 at one hop to 1,255 at three hops.
Figure 3. Mean and median cumulative reachability from the same 100 seeds. The vertical axis is logarithmic.
  • Market-Direction Analysis: In a practical application, the researchers used the graph to predict the direction of financial assets based on news signals. Using a rule-based engine to propagate signals through the graph, they achieved 58.26% accuracy on 872 asset-days, outperforming an "always-up" baseline by 8.26 percentage points.

Limitations and Future Work

  • Missing Text: Due to licensing and privacy, the public release contains the graph structure but not the original article text. This limits the ability to perform passage-grounded explanations.
  • Static Snapshot: The current release is a single snapshot. Causal reasoning often requires understanding how relationships change over time, which a static graph cannot fully capture.
  • Future Directions: The authors invite community contributions to the ontology and plan to incorporate additional source types, such as textbooks and educational materials, to add "general mechanisms" (e.g., economic principles) to the event-specific data.

Summary updated

Abstract

Applying causal knowledge to a new event requires connecting the event to relevant entities and the relationships through which its consequences may unfold. We present Causals, an open knowledge graph for research on this form of reasoning. Its initial release, constructed from news reports, contains 369,184 nodes and 1,272,418 edges linking reported events, organizations, commodities, financial instruments, and other entities. Structural relationships, observations, and attributed causal claims retain distinct semantics and links to source records. We describe the construction and representation of the resource and characterize its connectivity and multi-hop retrieval properties. Causals provides shared relational knowledge for studying how language models retrieve relevant context and apply mechanisms to particular events. The dataset is available under CC BY 4.0 at https://causals.org/, and we invite contributions to its ontology, records, and evidence.
BibTeX
@misc{lin2026causals,
  author = {Miaozhe Lin and Qiongzheng Lin and Zhongwei Liang and Zijun Liu},
  title = {{Causals: An Open Knowledge Graph for Research on Causal Reasoning}},
  year = {2026},
  eprint = {2611.16267},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.16267},
}
Figure