Abstract
Scientific discovery catalyzes human intellectual advances, driven by the cycle of hypothesis generation, experimental design, evaluation, and assumption refinement. Central to this process is causal inference, uncovering the mechanisms behind observed phenomena. While randomized experiments provide strong inferences, they are often infeasible due to ethical or practical constraints. However, observational studies are prone to confounding or mediating biases. While crucial, identifying such backdoor paths is expensive and heavily depends on scientists' domain knowledge to generate hypotheses. We introduce a novel benchmark where the objective is to complete a partial causal graph. We design a benchmark with varying difficulty levels with over 4000 queries. We show the strong ability of LLMs to hypothesize the backdoor variables between a cause and its effect. Unlike simple knowledge memorization of fixed associations, our task requires the LLM to reason according to the context of the entire graph.
Introduction
In the field of causal inference, researchers typically begin with a Directed Acyclic Graph (DAG) to map variables and their causal directions. However, real-world scientific models are frequently incomplete. Identifying "known unknowns"—specifically confounders that influence both cause and effect, or mediators that explain the underlying mechanism—is a significant bottleneck that traditionally requires extensive domain expertise.
This research investigates whether Large Language Models (LLMs) can automate this hypothesizing phase. By analyzing a partial graph, can an LLM act as a "proxy expert" to suggest the most plausible missing links? The study moves beyond simple association to test if models can reason about the underlying mechanisms of a system.
Methodology
The researchers formalized a novel task termed LLM-assisted causal variable inference. To evaluate this, they developed a multi-stage benchmark comprising four levels of increasing difficulty:
- Out-of-Context Identification: A multiple-choice task where the model selects the missing variable from a list containing the correct answer and unrelated distractors.
- In-Context Identification: A more rigorous version where distractors are other variables from the same causal graph, testing the model's understanding of specific topological roles.
- Hypothesizing in Open World: The model generates five original suggestions for a missing variable without predefined options.
- Iterative Hypothesizing: The model must reconstruct a chain of multiple missing mediators between a cause and an effect sequentially.
Evaluation of open-ended suggestions utilized Semantic Similarity and an LLM-as-Judge approach, where GPT-4 rated suggestions on a scale of 1 to 10 based on contextual fit.
Datasets
The study utilized eight diverse causal graphs, including standard synthetic graphs from the BNLearn Repository (e.g., Cancer, Asia, Alarm), an expert-curated Alzheimer’s Disease graph, and a complex Neuropathic Pain graph containing over 200 nodes to ensure the models were tested on data likely absent from their training sets.
Key Results and Insights
Topological Awareness
LLMs demonstrated an ability to use graph structure rather than relying solely on variable names. Performance was significantly higher when identifying mediators and colliders compared to source or sink nodes, suggesting that models excel at filling in the "middle" of causal chains.
Performance Benchmarks
- Top Performers: GPT-4 and GPT-4o dominated the benchmark, achieving near-perfect accuracy in multiple-choice tasks and high scores in open-ended generation.
- Open-Source Gap: Models like Mixtral and Llama-3 performed well on simpler graphs but struggled as complexity increased (e.g., the Insurance and Alarm graphs).
- Semantic Alignment: In open-world tasks, GPT-4 achieved semantic similarity scores between 0.50 and 0.54, with LLM-as-Judge scores between 7.3 and 7.6, indicating high contextual relevance even without exact word matches.
Reasoning vs. Memorization
To address concerns about data leakage, the authors tested models on a recent Alzheimer’s graph published after the models' training cut-offs. While models exhibited some "parametric memory" of famous datasets, they remained capable of reasoning effectively on unfamiliar structures. Furthermore, providing textual descriptions of the graph's domain significantly boosted performance, particularly for smaller models.
Discussion and Future Work
While LLMs show promise, the current framework assumes that the edges between known and missing variables are predefined. In true discovery scenarios, models would need to propose both the variable and its placement within the graph.
Future research may explore Retrieval-Augmented Generation (RAG) to incorporate the latest scientific literature into the model's reasoning process. Additionally, "human-in-the-loop" systems could allow scientists to validate LLM-proposed variables using statistical data, creating a more robust pipeline for automated scientific discovery.
Summary updated