Discovering and Reasoning of Causality in the Hidden World with Large Language Models
Authors: Chenxi Liu, Yongqiang Chen, Tongliang Liu, Mingming Gong, James Cheng, Bo Han, Kun Zhang
Organizations: TMLR Group, Hong Kong Baptist University, Kowloon, Hong Kong SAR · Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Department of Philosophy, Carnegie Mellon University, Pittsburgh, PA 15213, USA · School of Computer Science, The University of Sydney, NSW 2008, Australia · School of Mathematics and Statistics, The University of Melbourne, Victoria 3010, Australia · The Chinese University of Hong Kong, Shatin, Hong Kong SAR
Revealing hidden causal variables alongside the underlying causal mechanisms is essential to the development of science. Despite the progress in the past decades, existing practice in causal discovery (CD) heavily relies on high-quality measured variables, which are usually given by human experts. In fact, the lack of well-defined high-level variables behind unstructured data has been a longstanding roadblock to a broader real-world application of CD. This procedure can naturally benefit from an automated process that can suggest potential hidden variables in the system. Interestingly, Large language models (LLMs) are trained on massive observations of the world and have demonstrated great capability in processing unstructured data. To leverage the power of LLMs, we develop a new framework termed Causal representatiOn AssistanT (COAT) that incorporates the rich world knowledge of LLMs to propose useful measured variables for CD with respect to high-value target variables on their paired unstructured data. Instead of directly inferring causality with LLMs, COAT constructs feedback from intermediate CD results to LLMs to refine the proposed variables. Given the target variable and the paired unstructured data, we first develop COAT-MB that leverages the predictivity of the proposed variables to iteratively uncover the Markov Blanket of the target variable. Built upon COAT-MB, COAT-PAG further extends to uncover a more complete causal graph, i.e., Partial Ancestral Graph, by iterating over the target variables and actively seeking new high-level variables. Moreover, the reliable CD capabilities of COAT also extend the debiased causal inference to unstructured data by discovering an adjustment set. We establish theoretical guarantees for the CD results and verify their efficiency and reliability across realistic benchmarks and real-world case studies.
Figures & tables
Content
Clarification
Method
COAT-MB (Algorithm 1 ; Section 3.3 )
inherited
COAT-PAG (Algorithm 2 ; Section 3.4 )
new
COAT-ADJ (Algorithm 3 ; Section 3.5 )
new
Theoretical Analysis
Identifiability on Markov Blanket (Section 4.1 )
inherited
Identifiability on Partial Ancestral Graphs (Section 4.2 )
new
Identifiability on Adjustment Sets (Section 4.3 )
new
Table 1: Comparison of main results with conference version ( 2024 ).
Figure 1: An example where PAG is less informative when limited in MB(Y) . (a) is the true causal structure indicating causal directions; (b) and (c) are the partial ancestral graphs. The circle mark ( ∘ ) means it is undetermined to be arrow head or tail. The node out of MB(Y) is colored.
Figure 2: Examples of Partial Ancestral Graph and two Markov-equivalent mixed graphs.
Figure 3: Illustration of COAT framework in eliciting the Markov Blanket of the rating score from the paired text reviews. COAT first (a) adopts an LLM to read, comprehend, and relate the rich knowledge about reviews on apples. The LLM needs to propose a series of candidate factors such as apple sizes and smells, along with some meta-information such as annotation guidelines. Based on the candidate factors, COAT then (b) prompts another LLM to annotate the unstructured review into structured data. (c) The CD algorithm then finds causal relations among the factors, and constructs feedback based on samples where the ratings can not be well explained by the existing factors. By looking into the new samples, the LLM is expected to use more related knowledge to uncover the desired causal factors.
Figure 4: Illustration of the prompt template for factor proposal in COAT .
Figure 5: Illustration of variables that could be discovered with COAT . Let Y be the target variable, and W be a factor that has been discovered, and also assume a latent variable U=w(X)∈MB(Y) . Conditioning on W facilitates the discovery of U .
Figure 6: An example where arrow heads cannot be revealed within MB(N)(Y) .
Figure 7: An example where COAT-PAG is conducted in several loops. Starget are colored.
Figure 8: An illustration of COAT-MB with uncertainty estimation. Deeper colors of discovered factors indicate lower uncertainty.
Assumption
Notes
I. LLM Capability Assumptions
Propose valid factors, i.e., pΨ>0 and CΨ>0 (Definition 3 ).
Validated in Figure 10 (c) and Table 15
Accurate annotation with independent noise (Assumption 1 ).
Validated in Figure 10 (a-b) and Table 5
II. Causal Discovery Assumptions ( Spirtes et al., 1995 )
Faithfulness and Markov condition
Depends on the specific algorithm (e.g., FCI, LiNGAM)
Validity of conditional independence tests
Table 2: Summary of Core Assumptions for COAT Methods.
Figure 9: The discovered causal graphs in AppleGastronome . Compared to the ground truth results, directly adopting LLMs to reason the causal relations can easily elicit many false positive edges. In contrast, the relations recovered by COAT have a high precision and recall. The directed edge between “taste” and “juiciness” can not be recovered by COAT because of the limitations of FCI.
LLMs
Method
MB
NMB
OT
Recall
Precision
F1
GPT-4o
META
3.00 ± 0.82
1.00 ± 0.00
4.00 ± 0.82
0.60 ± 0.16
0.37 ± 0.09
0.46 ± 0.12
DATA
3.00 ± 0.00
0.67 ± 0.47
0.00 ± 0.00
0.60 ± 0.00
0.83 ± 0.12
0.69 ± 0.04
DATA+CoT
4.67 ± 0.58
1.00 ± 0.00
0.33 ± 0.58
0.93 ± 0.12
0.78 ± 0.10
0.85 ± 0.10
COAT
4.00 ± 0.82
0.00 ± 0.00
0.00 ± 0.00
0.80 ± 0.16
1.00 ± 0.00
0.88 ± 0.10
GPT-4
META
2.67 ± 0.94
0.67 ± 0.47
2.33 ± 0.47
0.53 ± 0.19
0.46 ± 0.08
0.49 ± 0.13
DATA
3.00 ± 0.00
0.33 ± 0.47
0.00 ± 0.00
0.60 ± 0.00
0.92 ± 0.12
0.72 ± 0.04
Table 3: Full Results on the Apple Gastronome Benchmark.
LLM
Method
SHD
SID
Recall
Precision
F1
GPT-4o
pairwise
5.33 ± 2.62
0.00 ± 0.00
1.00 ± 0.00
0.45 ± 0.07
0.62 ± 0.07
COAT
1.00 ± 0.00
0.00 ± 0.00
1.00 ± 0.00
0.79 ± 0.03
0.89 ± 0.02
GPT-4
pairwise
8.67 ± 2.05
1.33 ± 1.89
0.93 ± 0.09
0.35 ± 0.05
0.51 ± 0.06
COAT
2.33 ± 0.47
1.00 ± 1.41
0.93 ± 0.09
0.70 ± 0.09
0.79 ± 0.04
GPT-3.5
pairwise
5.67 ± 4.03
3.00 ± 2.94
0.67 ± 0.24
0.49 ± 0.36
0.55 ± 0.32
COAT
2.00 ± 0.82
2.67 ± 2.05
0.75 ± 0.20
0.63 ± 0.21
0.68 ± 0.21
Table 4: Causal relation extraction results in AppleGastronome .
Figure 10: Quantitative evaluation of the causal capabilities of LLMs in COAT .
LLM
Test Object
T
P-value
GPT-4
Feature
0.2828
0.9997
Noise
0.0327
0.9962
GPT-3.5
Feature
0.4803
0.0325
Noise
0.0446
0.9962
Table 5: Independence tests of the annotation noises with annotated features and other noises in AppleGastronome .
V1
V2
X1
X2
FCI Orientation Rules
graph 1
Taste
Score
Nutrition
Market Potential
R 0, R 1
graph 2
Taste
Score
Juiciness
Size
R 0
graph 3
Smell
Score
Market Potential
Nutrition
R 0, R 4
graph 4
Juiciness
Score
Nutrition
Smell
R 0, R 2, R 4
graph 5
Market Potential
Score
Taste
Juiciness
R 3
Table 6: The variable specification for different PAGs.
Figure 11: The five causal graphs used in the setting on extending PAG. The filled nodes are present in the original PAG.
Edge Mark
Arrow Head
Accuracy
Precision
Recall
F1
DeepSeek-V3
DATA+CoT
48 ± 40
52 ± 36
54 ± 34
53 ± 35
DATA+FCI
66 ± 19
88 ± 11
76 ± 22
80 ± 17
COAT
88 ± 8
96 ± 9
88 ± 11
91 ± 9
GLM-4-Plus
DATA+CoT
60 ± 16
74 ± 17
70 ± 22
69 ± 16
DATA+FCI
54 ± 30
68 ± 27
58 ± 25
61 ± 24
Table 7: Results on extending Partial Ancestral Graph.
Edge Mark
Arrow Head
Accuracy
Precision
Recall
F1
DeepSeek-V3
DATA+CoT
80 ± 45
80 ± 45
80 ± 45
80 ± 45
DATA+FCI
80 ± 45
80 ± 45
80 ± 45
80 ± 45
COAT
100 ± 0
100 ± 0
100 ± 0
100 ± 0
GLM-4-Plus
DATA+CoT
50 ± 0
90 ± 22
100 ± 0
93 ± 15
DATA+FCI
80 ± 45
80 ± 45
80 ± 45
80 ± 45
Table 7: Results on extending Partial Ancestral Graph (continued).
Edge Mark
Arrow Head
Accuracy
Precision
Recall
F1
DeepSeek-V3
DATA+CoT
30 ± 45
30 ± 45
30 ± 45
30 ± 45
DATA+FCI
80 ± 27
100 ± 0
80 ± 27
87 ± 18
COAT
80 ± 27
100 ± 0
80 ± 27
87 ± 18
GLM-4-Plus
DATA+CoT
40 ± 42
60 ± 55
40 ± 42
47 ± 45
DATA+FCI
90 ± 22
100 ± 0
90 ± 22
93 ± 15
Table 7: Results on extending Partial Ancestral Graph (continued).
Edge Mark
Arrow Head
Accuracy
Precision
Recall
F1
DeepSeek-V3
DATA+CoT
20 ± 45
20 ± 45
20 ± 45
20 ± 45
DATA+FCI
40 ± 22
80 ± 45
40 ± 22
53 ± 30
COAT
80 ± 27
100 ± 0
80 ± 27
87 ± 18
GLM-4-Plus
DATA+CoT
80 ± 27
100 ± 0
80 ± 27
87 ± 18
DATA+FCI
40 ± 22
80 ± 45
40 ± 22
53 ± 30
Table 7: Results on extending Partial Ancestral Graph (continued).
Edge Mark
Arrow Head
Accuracy
Precision
Recall
F1
DeepSeek-V3
DATA+CoT
100 ± 0
100 ± 0
100 ± 0
100 ± 0
DATA+FCI
80 ± 27
100 ± 0
100 ± 0
100 ± 0
COAT
90 ± 22
100 ± 0
100 ± 0
100 ± 0
GLM-4-Plus
DATA+CoT
70 ± 45
80 ± 45
70 ± 45
73 ± 43
DATA+FCI
20 ± 27
40 ± 55
40 ± 55
40 ± 55
Table 7: Results on extending Partial Ancestral Graph (continued).
Edge Mark
Arrow Head
Accuracy
Precision
Recall
F1
DeepSeek-V3
DATA+CoT
10 ± 22
30 ± 45
40 ± 55
33 ± 47
DATA+FCI
50 ± 0
80 ± 45
80 ± 45
80 ± 45
COAT
90 ± 22
80 ± 45
80 ± 45
80 ± 45
GLM-4-Plus
DATA+CoT
60 ± 55
60 ± 55
60 ± 55
60 ± 55
DATA+FCI
40 ± 55
40 ± 55
40 ± 55
40 ± 55
Table 7: Results on extending Partial Ancestral Graph (continued).
Figure 12: The comparison between COAT-MB and COAT-PAG in the causal discovery setting.
Figure 13: The discovered causal graphs in Neuropathic . (b) shows the result based on directly prompting LLM to reason for the causal relations among all factors. Disconnected ones are dropped.
LLM
Method
Factor Proposal
PA
AN
OT
Acc
F1
GPT-4
Meta
3
5
6
0.91
0.59
DATA
2
2
0
0.95
0.50
DATA+CoT
3
4
13
0.81
0.35
COAT
3
6
3
0.96
0.80
GPT-3.5
Meta
3
5
6
0.91
0.59
Table 8: Factor proposal in Neuropathic .
Figure 14: The Causal Graph Identified by Algorithm 3 on two marketing data sets.
Hillstrom
Retail Hero
Factor Provider
Method
ATE (pp)
RMSE
ATE (pp)
RMSE
-
Ground Truth
6.09
-
3.32
-
Human Expert ( Dhawan et al., 2024 )
Bag-of-Words
7.57 ± 1.37
2.23
2.61 ± 2.08
2.42
N-MC IPW
4.81 ± 0.80
1.51
1.85 ± 2.01
2.49
N-IPW
5.23 ± 1.00
1.32
3.83 ± 1.29
1.39
DeepSeek-V3
DATA+IPW
4.31 ± 0.12
1.78
-1.68 ± 1.72
5.29
Table 9: ATE Estimation Results.
Figure 15: The comparison between COAT-MB and COAT-ADJ in the causal inference setting.
Method
Cluster Size
Group Size
MB
NMB
OT
Recall
Precision
F1
META
-
-
2.67 ± 0.94
0.67 ± 0.47
2.33 ± 0.47
0.53 ± 0.19
0.46 ± 0.08
0.49 ± 0.13
DATA
-
3
3.00 ± 0.00
0.33 ± 0.47
0.00 ± 0.00
0.60 ± 0.00
0.92 ± 0.12
0.72 ± 0.04
COAT
len(factor)+1
3
4.00 ± 0.82
0.33 ± 0.47
0.00 ± 0.00
0.80 ± 0.16
0.93 ± 0.09
0.85 ± 0.11
COAT
len(factor)+1
1
4.67 ± 0.58
0.00 ± 0.00
0.00 ± 0.00
0.93 ± 0.12
1.00 ± 0.00
0.96 ± 0.06
COAT
2
3
3.67 ± 1.53
0.00 ± 0.00
0.00 ± 0.00
0.73 ± 0.31
1.00 ± 0.00
0.82 ± 0.22
Table 10: Results about Ablation Study on Hyperparameters.
Review Generation
Factor Proposal
MB
NMB
OT
Recall
Precision
F1
DeepSeek-V3-0324
DeepSeek-V3-0324
3.75 ± 0.66
0.00 ± 0.00
0.44 ± 0.51
0.75 ± 0.13
0.91 ± 0.10
0.82 ± 0.12
GPT-OSS-120B
5.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
Qwen3-Next-80B-A3B
2.94 ± 0.10
0.33 ± 0.58
0.00 ± 0.00
0.59 ± 0.02
0.92 ± 0.14
0.71 ± 0.04
GPT-OSS-120B
DeepSeek-V3-0324
4.08 ± 1.18
0.33 ± 0.58
0.00 ± 0.00
0.82 ± 0.24
0.91 ± 0.15
0.86 ± 0.20
GPT-OSS-120B
5.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
Qwen3-Next-80B-A3B
4.33 ± 1.15
0.00 ± 0.00
0.00 ± 0.00
0.87 ± 0.23
1.00 ± 0.00
0.92 ± 0.14
Table 11: Cross-Model Validation of AppleGastronome Construction.
Figure 16: Heatmap for the cross-model validation of synthetic data.
Factor
Type
Factor Definition
Juiciness
Correlated
1 : Abundant or moist texture described in text
0 : Not mentioned or ambiguous description
-1 : Dry or lacking moisture
Color
Irrelevant
1 : Vibrant or appealing color mentioned
0 : Not mentioned or neutral description
-1 : Uneven, dull, or unappealing color
Table 12: Definitions of injected factors for robustness stress testing. Juiciness serves as a disturbing candidate (correlated with outcome but not in the true Markov Blanket), while Color represents a completely irrelevant factor.
Injection
Iteration
MB
NMB/OT
Injected Factors
F1
juiciness
1
0.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
0.00 ± 0.00
2
3.20 ± 0.98
0.20 ± 0.40
0.00 ± 0.00
0.75 ± 0.16
3
3.60 ± 0.80
0.00 ± 0.00
0.00 ± 0.00
0.83 ± 0.10
4
4.00 ± 0.89
0.00 ± 0.00
0.00 ± 0.00
0.88 ± 0.11
color
1
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
2
2.60 ± 0.49
0.60 ± 0.80
0.00 ± 0.00
0.64 ± 0.14
Table 13: Stress test result with non-MB factor injected in the initial set. For the juiciness factor, it was preserved in the first round, and then was removed after 1 iteration. For the color factor, since it is not relevant to the textual data, it was rejected in the first round by statistical testing.
Figure 17: Stress test result with a variable set of pure noises. The initial variables U1,U2,U3 are not relevant to the textual data, and therefore COAT-PAG returns no additional factors.
Figure 18: Stress test result with latent confounder for the treatment effect variables. For (a) and (b), there is no latent variable, and the result from COAT-ADJ indicates the edge T→Y is visible since there exists another node (aroma) that is not adjacent to Y and has an edge into T . For (c) and (d), the edge T→Y is invisible due to the latent confounder (juiciness).
Figure 19: An example input from Brain Tumor data. Each row contains random samples from the same category (top-down: glioma, meningioma, and no tumor).
Figure 20: Detailed descriptions of these two final factors in the Brain Tumor case study.
Figure 21: Final causal graph by COAT-MB in the Brain Tumor case study.
Figure 22: The prompt with feedback based on previous loop in the ENSO case study.
Figure 23: The final causal graph found by COAT-MB in the ENSO case study.
Figure 24: Example of one processed factor Innovation and Technology Focus during the first 200 trading days. The red star marks the highest value. In June 2007, there was a discussion about future competition in desktop operating systems and the trend towards web-based services Markoff (2007) .
Figure 25: The final causal graph by COAT-MB in the Stock News case study.
Figure 26: The descriptions of the proposed factors in the Stock News data.
Figure 27: The out-of-sample cumulative return of trading strategies based on different factors in the Stock News case study.
Buy and Hold
product focus
legal/regulatory issues
market strategy
innovation and technology focus
Expected Return
-3.50
0.29
0.38
0.48
0.72
Sharpe Ratio
-2.38
0.70
0.94
1.20
1.86
t-Stat
-3.00
0.89
1.19
1.51
2.35
Information Ratio
-
2.45
2.50
2.56
2.73
α
-
3.79
3.88
3.98
4.22
α t-Stat
-
3.09
3.15
3.23
3.44
Table 14: Performance of trading strategy according to each factor.
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
Notation
Description
Ψ
The LLM for factor proposal.
p ( pt )
Prompt used (at the t -th COAT round) for factor proposal.
Ψs
The LLM for factor parsing.
pp
Prompt used for factor parsing.
A
Algorithm for causal discovery. In this paper we mainly use FCI.
X
Unstructured data; treated as a high-dimensional random vector.
Appendix
Table 47
Figure 28: Illustration of prompts for generating AppleGastronome .
Figure 29: Illustration of examples in AppleGastronome .
Figure 30: Illustration of the prompt for factor proposal.
Figure 31: Illustration of the prompt for factor annotation.
Figure 32: Illustration of the prompt for feedback.
LLMs
COAT Round
Perception Score
Capacity Score
I(y;x∣hS)
GPT-4o
iter 1
0.82 ± 0.02
0.24 ± 0.01
0.28 ± 0.02
iter 2
0.44 ± 0.08
0.26 ± 0.02
0.21 ± 0.03
iter 3
0.38 ± 0.12
0.19 ± 0.09
0.24 ± 0.05
GPT-4
iter 1
0.78 ± 0.02
0.18 ± 0.02
0.45 ± 0.04
iter 2
0.61 ± 0.08
0.26 ± 0.05
0.28 ± 0.05
iter 3
0.39 ± 0.28
0.22 ± 0.00
0.29 ± 0.02
Appendix
Table 15: Full Results of Causal Metrics in Each Round for Each LLM on the Apple Gastronome Benchmark.
Figure 33: Ground truth and faithful (via FCI algorithm) causal graphs in AppleGastronome .
Figure 34: Causal graphs with GPT-4 in AppleGastronome .
Figure 35: Causal graphs with GPT-3.5 in AppleGastronome .
Figure 36: Causal graphs with Llama-2 in AppleGastronome .
Figure 37: Causal graphs with mistral Medium in AppleGastronome .
Figure 38: Causal graphs with Claude-3-Opus in AppleGastronome .
Figure 39: Causal graphs with DeepSeek-V2 in AppleGastronome .
Figure 40: Causal graphs with Llama-3-70b in AppleGastronome .
Figure 41: Causal graphs with mistral-Large in AppleGastronome .
Figure 42: Causal graphs with qwen-1.5-110B in AppleGastronome .
Figure 43: Causal graphs with GPT-4o in AppleGastronome .
Figure 44: Illustration of prompts for generating Neuropathic .
Figure 45: Illustration of examples in Neuropathic .
Figure 46: Ground truth and faithful (via FCI algorithm) causal graphs in Neuropathic .
Figure 47: Causal graphs with GPT-4 in Neuropathic .
Figure 48: Causal graphs with GPT-3.5 in Neuropathic .
Figure 49: Causal graphs with LLaMA-2-70b in Neuropathic .
Figure 50: Causal graphs with Mistral-Medium in Neuropathic .
LLM
Method
Factor Proposal
PA
AN
OT
Acc
F1
GPT-4
Meta
3
5
6
0.91
0.59
Data
2
2
0
0.95
0.50
COAT
3
6
3
0.96
0.80
Data (LiNGAM)
3
3
0
0.96
0.67
COAT (LiNGAM)
3
6
4
0.95
0.75
Appendix
Table 16: Causal discovery results in Neuropathic . PA, AN, and OT refer to the parents, ancestors, and others, respectively. Accuracy and F1 measure the recovery of the causal ancestors.
Figure 51: The discovered causal graphs in Neuropathic with LiNGAM.
Benchmark Name
Type
Source
Sample
Sample Size
Ground Truth
Apple Gastronome
Synthetic Data
This paper. It will be open-sourced under CC-BY 4.0 .
Textual
200 textual reviews.
Yes
Neuropathic
Semi Real-world Data
Tabular samples from Tu et al. (2019) on their GitHub repo under CC-BY 4.0 . Textual samples are generated by this paper.
Textual
100 synthetic textual data for prompt construction; 1000 tabular data for CI tests.
Yes
Brain Tumor
Real-world Data
An open Kaggle data set ( kaggle/brain-tumor-classification-mri ) with an open-sourced project. ( Bhuvaji et al., 2024 ) . MIT License.
Image
60 MRI samples for COAT and CI tests. The sample size is small due to the expensive API cost. One can easily construct a much larger data set by following our instructions.
No
Stock News
Real-world Data
An open Kaggle data set ( kaggle/stock-price-and-news-realted-to-it ) ( Innovations, 2023 ) . MIT License.
Textual with time index
200 samples for COAT experiments. 400 samples for empirical validation.
No
ENSO
Real-world Data
The NOAA 20th Century Reanalysis V3 data set ( Compo et al., 2011 ) from their website at https://psl.noaa.gov (with CC0 1.0 License ). We developed a set of simple tools for LLMs to utilize the official database.
NetCDF
Monthly observation for 1836/01 to 2015/12; 1.0 degree latitude x 1.0 degree longitude global grid (360x181); 90+ climate variables.
No
Appendix
Table 17: Summary of Benchmark Data
Figure 52: Schematic overview of COAT-MB output components for end users. It has four integrated modules (conceptual layout): factor definitions with uncertainty indicators (right), the inferred causal structure (upper-left), structured annotation results (middle-left), and annotation details (bottom-left).
Causal discovery seeks to uncover causal relations from data, typically represented as causal graphs, and is essential for predicting the effects of interventions. While expert knowledge is required to construct principled causal graphs, many statistical methods have been proposed to leverage observational data with varying formal guarantees. Causal Assumption-based Argumentation (ABA) is a framework that uses symbolic reasoning to ensure correspondence between input constraints and output graphs, while offering a principled way to combine data and expertise. We explore the use of large language models (LLMs) as imperfect experts for Causal ABA, eliciting semantic structural priors from variable names and descriptions and integrating them with conditional-independence evidence. Experiments on standard benchmarks and semantically grounded synthetic graphs demonstrate state-of-the-art performance, and we additionally introduce an evaluation protocol to mitigate memorisation bias when assessing LLMs for causal discovery.
Zihao Li, Fabrizio Russo
Department of Computing, Imperial College London, UK
Causal discovery is a cornerstone of scientific reasoning, yet whether large language models can perform it reliably remains an open question. Recent benchmarks show that even fine-tuned models plateau on simple causal graphs and degrade as complexity grows, but why they fail has not been established. We prove the failure is fundamental: supervised fine-tuning, direct preference optimization, and in-context learning all produce predictors that cannot distinguish between causal graphs generating similar observational data, and any attempt to do so requires the model's internal representations to grow unboundedly, violating the very conditions under which these methods work. We formalize this as a kernel obstruction theorem, establishing that the limitation is intrinsic to the learning paradigm, not any particular model or dataset. We propose Agentic Causal Bayesian Optimization (A-CBO), wherein a frozen language model serves as an interventional oracle answering targeted queries about intervention effects, while an external Bayesian loop concentrates beliefs over candidate graphs in logarithmically many rounds. Because the decision operates outside the space where the obstruction applies, A-CBO provably converges while the underlying model remains unchanged. On Corr2Cause, A-CBO matches fine-tuned baselines without any training. On Extended Corr2Cause, a new benchmark scaling to 24 variables with 18K test samples, A-CBO significantly outperforms both fine-tuning and preference optimization, with the advantage growing
Amartya Roy, Sonali Parbhoo
SIRE, IIT Delhi and Robert Bosch GmbH, India · Imperial College London
Revealing the underlying causal mechanisms in the real world is crucial for scientific and technological progress. Despite notable advances in recent decades, the lack of high-quality data and the reliance of traditional causal discovery algorithms (TCDA) on the assumption of no latent confounders, as well as their tendency to overlook the precise semantics of latent variables, have long been major obstacles to the broader application of causal discovery. To address this issue, we propose a novel causal modeling framework, TLVD, which integrates the metadata-based reasoning capabilities of large language models (LLMs) with the data-driven modeling capabilities of TCDA for inferring latent variables and their semantics. Specifically, we first employ a data-driven approach to construct a causal graph that incorporates latent variables. Then, we employ multi-LLM collaboration for latent variable inference, modeling this process as a game with incomplete information and seeking its Bayesian Nash Equilibrium (BNE) to infer the possible specific latent variables. Finally, to validate the inferred latent variables across multiple real-world web-based data sources, we leverage LLMs for evidence exploration to ensure traceability. We comprehensively evaluate TLVD on three de-identified real patient datasets provided by a hospital and two benchmark datasets. Extensive experimental results confirm the effectiveness and reliability of TLVD, with average improvements of 32.67% in Acc, 62.21% in CAcc, and 26.72% in ECit across the five datasets.
Huaming Du, Tao Hu, Yijie Huang +5
Southwestern University of Finance and Economics, Chengdu, Sichuan, China · Hunan University of Technology and Business, Xiangjiang Laboratory, Hunan, China · Emory University, Atlanta, Georgia, United States