From Causal Plausibility to Causal Reliability: Evaluating LLMs as Calibrated Direct Causal-Edge Classifiers

Amit Kumar; Elnur Adl Zarabi; Suranjana Trivedy; Zhiqian Chen; Lei Zhang; Kaiqun Fu; Taoran Ji

cs.LGAug 24, 2026

From Causal Plausibility to Causal Reliability: Evaluating LLMs as Calibrated Direct Causal-Edge Classifiers

Authors: Amit Kumar, Elnur Adl Zarabi, Suranjana Trivedy, Zhiqian Chen, Lei Zhang, Kaiqun Fu, Taoran Ji

Organizations: Texas A&M University-Corpus Christi, USA · BITS Pilani Goa, India · Mississippi State University, USA · Northern Illinois University, USA ·

Abstract

Large language models (LLMs) are increasingly used to provide prior causal knowledge for structural causal discovery, yet whether their direct-edge judgments and confidence can be trusted remains unclear. We systematically evaluate 12 instruction-tuned open-weight models across six benchmark causal graphs, five prompting strategies, and four confidence sources: verbalized, logit-based, cross-prompt agreement, and cross-model agreement. Under our language-only pairwise protocol, our evaluation yields three key findings. (i) LLM-based causal judgments are strongly recall-dominant: models predict overly dense graphs with many false-positive edges, while prompting mainly shifts the precision-recall trade-off rather than resolving overprediction. Gains from model scale diminish on the largest graphs and do not eliminate miscalibration. (ii) LLMs often capture causal relatedness without reliably identifying directness or orientation. Relative to published reference graphs, models misclassify 40.0% of indirect and 36.0% of reversed non-edges as direct edges, versus 28.2% of other non-edges. Moreover, 80.8% and 84.6% of these false positives receive verbalized confidence of at least 80%, revealing substantial overconfidence in structurally incorrect predictions. (iii) Conventional confidence estimates are unreliable, whereas agreement offers a more promising signal. Logit-based confidence frequently collapses near 1.0 regardless of correctness, while cross-prompt and cross-model agreement achieve better mean calibration and discrimination, though their advantages are not statistically significant after Holm correction. A benchmark-familiarity audit further identifies potential familiarity in five model-dataset pairs, all involving AsiaM. Overall, our results suggest LLMs are better viewed as sources of externally validated soft causal priors than as direct evidence of causal structure.

Figures & tables

Figure 1: Illustration of unreliable pairwise direct-edge judgments. Solid arrows form the reference chain; dashed arrows show an indirect relation predicted as direct and a reversed relation. Confidence values are illustrative.

Figure 2: Evaluation pipeline for pairwise direct-edge classification, graph reconstruction, confidence calibration, and structural error analysis.

Figure 3: Precision–recall behavior by (a) prompt, averaged across models and datasets, and (b) model, averaged across prompts and datasets. The dashed diagonal denotes equal precision and recall; points above it are recall-dominant.

Dataset	Precision $\uparrow$		Recall $\uparrow$		F1 $\uparrow$		nSHD $\downarrow$
Dataset	S	L	S	L	S	L	S	L
AsiaM	$0.415\pm 0.175$	$\mathbf{0.608\pm 0.139}$	$0.707\pm 0.264$	$\mathbf{0.830\pm 0.210}$	$0.478\pm 0.149$	$\mathbf{0.680\pm 0.126}$	$0.281\pm 0.145$	$\mathbf{0.141\pm 0.062}$
River Status	$0.198\pm 0.048$	$\mathbf{0.216\pm 0.062}$	$0.592\pm 0.213$	$\mathbf{0.597\pm 0.292}$	$0.281\pm 0.050$	$\mathbf{0.299\pm 0.107}$	$0.357\pm 0.139$	$\mathbf{0.292\pm 0.090}$
COVID	$0.261\pm 0.181$	$\mathbf{0.412\pm 0.171}$	$0.443\pm 0.221$	$\mathbf{0.545\pm 0.252}$	$0.280\pm 0.144$	$\mathbf{0.403\pm 0.080}$	$0.191\pm 0.158$	$\mathbf{0.104\pm 0.039}$
Coal Gasifier	$0.049\pm 0.029$	$\mathbf{0.069\pm 0.057}$	$0.428\pm 0.247$	$\mathbf{0.505\pm 0.316}$	$0.078\pm 0.045$	$\mathbf{0.095\pm 0.053}$	$0.288\pm 0.204$	$\mathbf{0.225\pm 0.159}$
Hepar2	$0.062\pm 0.021$	$\mathbf{0.086\pm 0.024}$	$0.546\pm 0.173$	$\mathbf{0.593\pm 0.218}$	$0.106\pm 0.030$	$\mathbf{0.143\pm 0.029}$	$0.265\pm 0.144$	$\mathbf{0.191\pm 0.097}$

Table 1: Dataset-level edge-classification and graph-reconstruction performance. Values are mean

\pm

SD across model–prompt combinations within each scale group (S: 7 models; L: 5 models). Higher precision, recall, and F1 and lower nSHD indicate better performance.

Figure 4: Dataset-level mean F1 for small and large LLMs, averaged across models and prompts within each group. Lines connect group means; rightward shifts favor large LLMs.

Model	Nodes	Recall	Dev.	Edge F1	SHD $\downarrow$
Phi-4	8/7	0.857	14.3%	0.500	6
Qwen3-32B	7/7	0.857	0.0%	0.714	4
Qwen2.5-72B	8/7	1.000	14.3%	0.857	2
Gemma-4-31B	7/7	1.000	0.0%	0.800	3
Llama-3.3-70B	8/7	1.000	14.3%	0.941	1

Table 2: Structure recall for model–dataset pairs passing the benchmark-familiarity screen, all on AsiaM (7 nodes, 8 edges). Nodes reports generated/true counts, Dev. is node-count deviation, and SHD is the number of edge edits.

Figure 5: Edge precision–recall curves obtained by thresholding positive-edge confidence for small and large model groups, macro-averaged across six datasets.

Group	Method	ECE $\downarrow$	Brier $\downarrow$	AUROC $\uparrow$	Acc. $\uparrow$
Small	Cross-model	0.019 $\pm$ 0.013	0.149 $\pm$ 0.043	0.692 $\pm$ 0.043	0.792 $\pm$ 0.071
	Cross-prompt	0.038 $\pm$ 0.012	0.176 $\pm$ 0.043	0.661 $\pm$ 0.047	0.755 $\pm$ 0.060
	Logit-based	0.226 $\pm$ 0.054	0.354 $\pm$ 0.046	0.611 $\pm$ 0.037	0.549 $\pm$ 0.064
	Verbalized	0.229 $\pm$ 0.034	0.364 $\pm$ 0.032	0.358 $\pm$ 0.059	0.709 $\pm$ 0.057
Large	Cross-prompt	0.016 $\pm$ 0.012	0.127 $\pm$ 0.050	0.772 $\pm$ 0.047	0.820 $\pm$ 0.070
	Cross-model	0.024 $\pm$ 0.018	0.139 $\pm$ 0.056	0.704 $\pm$ 0.055	0.822 $\pm$ 0.070

Table 3: Calibration by model group and confidence source, reported as mean

\pm

SD across six datasets. ECE denotes squared-gap ECE. Bold and underlining indicate the best and second-best results within each group.

Figure 6: Agreement-confidence distributions pooled across six datasets: (a) cross-prompt agreement for Gemma-4-31B-IT and (b) cross-model agreement under Metadata prompting. Correct predictions concentrate more strongly near full agreement.

Non-edge Type	Valid Pairs	FP Count	FP Rate	High-Conf. FP Rate
Reversed direct	29,559	10,644	36.0%	30.4%
Indirect	161,520	64,568	40.0%	32.3%
Other	2,256,303	637,205	28.2%	22.1%

Table 4: False-positive behavior by non-edge type. FP Rate is the fraction of valid queries producing false positives; High-Conf. FP Rate is the fraction producing false positives with verbal confidence at least 80.

Figure 7: Verbal-confidence survival among false positives with parseable scores. Each curve shows the fraction at or above a given threshold; reversed-direct errors remain most concentrated at high confidence.

Model	Formulation	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$
Qwen3-4B	Binary	0.229	0.570	0.261	0.200
Qwen3-4B	Three-way	0.124	0.678	0.194	0.276
Phi-4	Binary	0.222	0.416	0.234	0.132
Phi-4	Three-way	0.224	0.627	0.303	0.145
Gemma-4-31B	Binary	0.177	0.816	0.270	0.242
Gemma-4-31B	Three-way	0.187	0.714	0.275	0.205

Table 5: Binary versus three-way direct-edge classification under Metadata prompting, macro-averaged across four datasets. Bold indicates the better formulation for each model and metric.

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Domain	Source	Nodes	Edges	Pairs	No-edge:Edge
Small networks ( $n<20$ )
AsiaM	Respiratory diagnosis	bnlearn; ( 21 )	7	8	42	4.25:1
River Status	Ecological quality	( 28 )	15	25	210	7.40:1
Medium networks ( $20\leq n\leq 50$ )
COVID	Vaccine risk-benefit	BayesFusion; ( 25 )	20	26	380	13.62:1
Coal Gasifier	Industrial risk	( 23 )	39	39	1,482	37.00:1

Appendix

Table A1: Reference-graph statistics. No-edge:edge denotes class imbalance over ordered variable pairs.

Dataset	Number of variables	Mean words per description	Code-like variable names (%)	Descriptions with acronyms (%)
AsiaM	7	5.3	0.0	0.0
River Status	15	3.7	6.7	0.0
COVID	20	6.8	100.0	45.0
Coal Gasifier	39	3.9	100.0	5.1
Hepar2	70	3.4	15.7	0.0
Munin1	186	7.6	100.0	27.4

Appendix

Table A2: Metadata-level indicators of variable-description complexity across datasets. The table reports simple descriptive measures of the node metadata used in the classification prompts. Lower values for the number of variables, code-like variable names, and acronym-containing descriptions generally indicate easier metadata conditions. Mean words per description is included as a descriptive measure of label length, but is not assumed to have a strictly monotonic relationship with performance because longer descriptions may either clarify or complicate variable meaning.

Figure A1: Prompt-level F1 performance across models and datasets. Each cell shows the best F1 score and prompt, with right-side bars summarizing prompt wins per model. The horizontal line separates large and small models.

Model	Dataset	Name-only				Metadata				CoT				Few-shot				Few-shot+CoT
Model	Dataset	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$
Gemma-4- E4B-IT	asiam	0.500	0.375	0.429	0.190	0.417	0.625	0.500	0.214	1.000	0.125	0.222	0.167	0.400	0.500	0.444	0.190	0.625	0.625	0.625	0.143
	river	0.200	0.240	0.218	0.205	0.239	0.440	0.310	0.233	0.212	0.560	0.308	0.300	0.167	0.920	0.282	0.557	0.169	0.880	0.284	0.529
	covid	0.071	0.038	0.050	0.097	0.435	0.385	0.408	0.074	0.421	0.308	0.356	0.076	0.202	0.654	0.309	0.197	0.162	0.654	0.260	0.255
	coal	0.042	0.026	0.032	0.041	0.125	0.256	0.168	0.067	0.096	0.308	0.146	0.093	0.038	0.846	0.073	0.562	0.043	0.718	0.082	0.423
	hepar2	0.096	0.325	0.148	0.094	0.093	0.415	0.152	0.115	0.096	0.179	0.125	0.062	0.046	0.626	0.086	0.335	0.049	0.561	0.091	0.281

Appendix

Table A3: Primary causal edge classification performance for small language models across six benchmark datasets and five prompt styles. Cell colors encode performance percentile within the full table (yellow = good, orange = bad), with intensity scaled per metric direction (P/R/F1 higher is better; nSHD lower is better). A dash denotes undefined precision because the model made no positive predictions.

Small LLMs
Model	Prompt	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	Avg. nSHD $\downarrow$
Gemma-4- E4B-IT	Name-only	0.154	0.176	0.150	0.111
	Metadata	0.224	0.386	0.266	0.126
	CoT	0.307	0.287	0.199	0.135
	Few-shot	0.144	0.733	0.202	0.421
	Few-shot + CoT	0.176	0.707	0.227	0.384

Appendix

Table A4: Macro-averaged primary causal edge classification performance for small (left) and large (right) LLMs across all datasets. Precision (P), recall (R), and F1 are macro-averaged over the six benchmark datasets for each model–prompt pair. Avg. nSHD denotes the macro-average of normalized structural Hamming distance, computed for each dataset as

\mathrm{SHD}/(n_{d}(n_{d}-1))

, where

n_{d}

is the number of variables in dataset

d

. Higher P, R, and F1 are better, while lower Avg. nSHD is better. Bold values indicate the best prompt for each model and metric.

	Small LLMs					Large LLMs
Dataset	Model	Avg. P $\uparrow$	Avg. R $\uparrow$	Avg. F1 $\uparrow$	Avg. nSHD $\downarrow$	Model	Avg. P $\uparrow$	Avg. R $\uparrow$	Avg. F1 $\uparrow$	Avg. nSHD $\downarrow$
asiam	Gemma-4-E4B-IT	0.588	0.450	0.444	0.181	Qwen3-32B-Instruct	0.603	0.825	0.694	0.124
	Llama-3.1-8B-Instruct	0.407	0.800	0.526	0.276	Gemma-4-31B-IT	0.705	0.900	0.780	0.090
	Ministral-8B-Instruct-2410	0.220	0.800	0.344	0.562	Qwen2.5-72B-Instruct	0.680	0.850	0.733	0.114
	Phi-4	0.470	0.825	0.593	0.210	Llama-3.3-70B-Instruct	0.612	0.800	0.638	0.167
	Phi-4-Mini-Instruct	0.356	0.275	0.305	0.190	Llama-3.1-70B-Instruct	0.438	0.775	0.552	0.210

Appendix

Table A5: Dataset-wise prompt-averaged primary causal edge classification performance for small and large LLMs on the six benchmark datasets. For each model–dataset pair, precision (P), recall (R), F1, and normalized structural Hamming distance (nSHD) are averaged across the five prompt styles. nSHD is computed as

\mathrm{SHD}/(n_{d}(n_{d}-1))

, where

n_{d}

is the number of variables in dataset

d

. Higher P, R, and F1 are better, while lower Avg. nSHD is better. Bold values indicate the best model within each dataset and metric, with small and large LLMs evaluated independently.

Model	Dataset	Name-only				Metadata				CoT				Few-shot				Few-shot+CoT
Model	Dataset	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$
Qwen3-32B- Instruct	asiam	0.500	0.625	0.556	0.190	0.667	0.750	0.706	0.095	0.727	1.000	0.842	0.071	0.583	0.875	0.700	0.119	0.538	0.875	0.667	0.143
	river	0.203	0.560	0.298	0.314	0.217	0.520	0.306	0.281	0.273	0.720	0.396	0.262	0.209	0.560	0.304	0.305	0.212	0.720	0.327	0.352
	covid	0.444	0.154	0.229	0.071	0.538	0.269	0.359	0.066	0.500	0.269	0.350	0.068	0.324	0.423	0.367	0.100	0.281	0.692	0.400	0.139
	coal	0.000	0.000	0.000	0.028	0.082	0.684	0.146	0.206	0.076	0.513	0.132	0.176	0.055	0.667	0.101	0.310	0.046	0.641	0.087	0.355
	hepar2	0.060	0.472	0.107	0.196	0.067	0.463	0.117	0.173	0.072	0.463	0.124	0.161	0.086	0.528	0.148	0.150	0.086	0.480	0.146	0.139

Appendix

Table A6: Primary causal edge classification performance for large language models across six benchmark datasets and five prompt styles. Cell colors encode performance percentile within the full table (yellow = good, orange = bad), with intensity scaled per metric direction (P/R/F1 higher is better; nSHD lower is better). A dash denotes undefined precision because the model made no positive predictions.

	Small LLMs				Large LLMs
Dataset	Model	AUROC $\uparrow$	ECE $\downarrow$	Brier $\downarrow$	Model	AUROC $\uparrow$	ECE $\downarrow$	Brier $\downarrow$
asiam	Gemma-4-E4B-IT	0.845	0.005	0.110	Qwen3-32B-Instruct	0.747	0.002	0.077
	Llama-3.1-8B-Instruct	0.689	0.004	0.158	Gemma-4-31B-IT	0.774	0.001	0.056
	Ministral-8B-Instruct-2410	0.540	0.228	0.428	Qwen2.5-72B-Instruct	0.822	0.006	0.065
	Phi-4	0.773	0.002	0.118	Llama-3.3-70B-Instruct	0.863	0.014	0.094
	Phi-4-Mini-Instruct	0.779	0.010	0.138	Llama-3.1-70B-Instruct	0.753	0.010	0.148

Appendix

Table A7: Cross-prompt agreement calibration results for small (left) and large (right) LLMs across all datasets. This method combines predictions across the five prompt styles for each fixed model and dataset. Bold values indicate the best model within each dataset for each metric (small and large LLMs evaluated independently).

		Small LLMs			Large LLMs
Dataset	Prompt	AUROC $\uparrow$	ECE $\downarrow$	Brier $\downarrow$	AUROC $\uparrow$	ECE $\downarrow$	Brier $\downarrow$
asiam	Name-only	0.772	0.008	0.135	0.632	0.022	0.114
	Metadata	0.711	0.014	0.105	0.656	0.012	0.054
	CoT	0.655	0.024	0.131	0.829	0.008	0.099
	Few-shot	0.671	0.012	0.207	0.871	0.018	0.095
	Few-shot + CoT	0.833	0.019	0.158	0.523	0.016	0.100

Appendix

Table A8: Cross-model agreement calibration results for small (left) and large (right) LLMs across all datasets. This method combines predictions across all models in each group for each fixed prompt style and dataset. Bold values indicate the best prompt within each dataset for each metric (small and large LLMs evaluated independently).

		Small LLMs				Large LLMs
Metric	Comparison	$\widetilde{\Delta}$	Raw $p$	Holm $p$	Comparison	$\widetilde{\Delta}$	Raw $p$	Holm $p$
ECE	CM vs. CP	-0.022	0.0625	0.1875	CM vs. CP	0.007	0.0312	0.1875
	CM vs. Logit	-0.196	0.0312	0.1875	CM vs. Logit	-0.189	0.0312	0.1875
	CM vs. Verb.	-0.207	0.0312	0.1875	CM vs. Verb.	-0.179	0.0312	0.1875
	CP vs. Logit	-0.173	0.0312	0.1875	CP vs. Logit	-0.196	0.0312	0.1875
	CP vs. Verb.	-0.191	0.0312	0.1875	CP vs. Verb.	-0.179	0.0312	0.1875

Appendix

Table A9: Pairwise two-sided Wilcoxon signed-rank tests over six paired dataset-level scores, small (left) and large (right) LLMs evaluated separately.

\widetilde{\Delta}

is the median difference (first method minus second). Holm correction is applied across the six pairwise comparisons within each model group and metric. CM = cross-model agreement, CP = cross-prompt agreement, Verb. = verbalized confidence. With

N=6

paired observations, the smallest attainable two-sided raw

p

-value is

2/2^{6}\approx 0.031

; no adjusted (Holm)

p

-value falls below 0.05.

Non-edge type	Model	False-positive rate	High-conf. false- positive rate	False positives with conf. $\geq 80$
Reversed direct	Ministral-8B-Instruct-2410	58.6%	45.0%	76.8%
Indirect	Gemma-4-31B-IT	56.3%	52.2%	92.7%
Other	Ministral-8B-Instruct-2410	56.4%	41.0%	72.7%

Appendix

Table A11: Worst model-level high-confidence false-positive behavior by non-edge type. For each non-edge category, we report the model with the largest high-confidence false-positive rate, along with its overall false-positive rate and the share of its false positives that have verbal confidence of at least 80.

Figure A2: Model classification performance ranking across datasets. Performance is averaged over five prompt styles.

Figure A3: Model classification performance as a function of dataset size. Datasets are ordered by increasing number of variables. Points show mean F1 averaged over prompt styles, and shaded bands show variation across models within each group.

Figure A8: Confidence distributions for Qwen2.5-72B-Instruct under the Chain-of-Thought prompt. Predictions are pooled across all six benchmark datasets. Blue curves denote correctly classified variable pairs, while orange curves denote misclassified pairs. Panel (a) shows verbalized confidence, and panel (b) shows logit-based confidence. Both confidence sources concentrate in the high-confidence region, with logit-based confidence collapsing almost entirely near 1.0 for both correct and incorrect predictions, indicating strong overconfidence and limited separation between reliable and unreliable predictions.

Figure A9: Agreement-based confidence distributions for Qwen2.5-72B-Instruct and the large-model ensemble. Panel (a) shows cross-prompt agreement for Qwen2.5-72B-Instruct, while panel (b) shows cross-model agreement under the Chain-of-Thought prompt. Correct predictions concentrate more strongly near full agreement, indicating that agreement-based confidence is more informative than raw confidence for identifying reliable predictions.

Figure A10: Confidence distributions for Gemma-4-31B-IT using the Metadata prompt.

Model	Dataset	P $\uparrow$	R $\uparrow$	F1 $\uparrow$	nSHD $\downarrow$	Non-edge FP Rate $\downarrow$	Orientation Error $\downarrow$	$A\!\rightarrow\!B$ Recall $\uparrow$	$B\!\rightarrow\!A$ Recall $\uparrow$
Qwen3-4B	River Status	0.247	0.760	0.373	0.305	72.5%	0.0%	0.760	–
	COVID	0.167	0.500	0.250	0.197	39.6%	18.8%	0.812	0.000
	Hepar2	0.067	0.813	0.124	0.291	60.6%	5.7%	0.909	0.000
	Munin1	0.016	0.637	0.031	0.311	62.9%	22.7%	0.879	0.000
Phi-4	River Status	0.287	0.920	0.438	0.281	71.3%	0.0%	0.920	–
	COVID	0.471	0.615	0.533	0.071	11.0%	5.9%	0.812	0.300

Appendix

Table A12: Dataset-level three-way classification results under Metadata prompting. Non-edge FP rate is the fraction of true no-edge pairs assigned either directed label. Orientation error is the fraction of detected true-edge pairs assigned the reversed direction. A dash indicates that the dataset contains no

B\!\rightarrow\!A

instances under the fixed unordered-pair ordering.

Small LLMs

Large LLMs

Dataset

Model

Nodes

Matches

Recall

Dev.

Risk

Model

Nodes

Matches

Recall

Dev.

Risk

asiam

Gemma-4-E4B-IT

10/7

0.286

0.429

Qwen3-32B-Instruct

7/7

0.857

0.000

Yes

Llama-3.1-8B-Instruct

8/7

0.143

Gemma-4-31B-IT

7/7

1.000

0.000

Yes

Ministral-8B-Instruct-2410

1/7

0.143

0.857

Qwen2.5-72B-Instruct

8/7

1.000

0.143

Yes

Phi-4

8/7

0.857

0.143

Yes

Llama-3.3-70B-Instruct

8/7

1.000

0.143

Yes

Phi-4-Mini-Instruct

5/7

0.429

0.286

Llama-3.1-70B-Instruct

10/7

1.000

0.429

Appendix

Table A13: Full node-level contamination results for all model–dataset pairs, small (left) and large (right) LLMs evaluated separately. Nodes reports generated/true counts, and Matches reports semantically aligned nodes. High-risk pairs (node-count deviation below 15% and node recall above 0.85) are highlighted.

Figure A11: Post-hoc temperature sensitivity of logit-based calibration on Munin1.

Figure A12: Post-hoc temperature sensitivity of logit-based calibration on AsiaM. AUROC is nearly unchanged because temperature scaling preserves the logit ranking, while Gemma-31B’s ECE decreases at higher temperatures, indicating overconfident Yes/No logits. Qwen2.5-72B remains stable, suggesting better calibrated logit confidence.

Figure A13: Post-hoc temperature sensitivity of logit-based calibration on River Status. The main effect of temperature appears in ECE rather than AUROC, showing that temperature changes confidence sharpness rather than discrimination. Gemma-31B benefits from softer logits, while Qwen2.5-72B is comparatively insensitive.

Figure A14: Post-hoc temperature sensitivity of logit-based calibration on COVID. AUROC remains stable across temperatures, confirming that prediction ranking is preserved. ECE varies modestly, with Gemma-31B showing some calibration benefit from temperature adjustment and Qwen2.5-72B remaining consistently well calibrated.

Figure A15: Post-hoc temperature sensitivity of logit-based calibration on Coal Gasifier. Gemma-31B shows a clear ECE reduction as temperature increases, suggesting its saved Yes/No logits are too sharp for reliable confidence estimates. Qwen2.5-72B has low and stable ECE, indicating stronger post-hoc calibration.

Figure A16: Post-hoc temperature sensitivity of logit-based calibration on Hepar2. Gemma-31B calibration improves substantially under higher-temperature softening, while AUROC remains nearly constant. This supports the project’s finding that logit confidence can be recalibrated without changing causal-edge predictions.

Causal Reasoning in Language Models

Causal reasoning in language models studies how models identify cause-and-effect structure, predict interventions, and answer counterfactual questions.

See all 102 papers

CauTion: Knowing When to Trust LLMs for Ensemble Causal Discovery
Jun 2, 2026Bo Peng, Kaiwen Wu, Sirui Chen +3Causal Discovery
Shanghai AI Laboratory · Shanghai Innovation Institute · Shanghai Jiao Tong University +2
Causality Elicitation from Large Language Models
Mar 4, 2026Takashi Kameyama, Masahiro Kato, Yasuko Hio +2Causal Reasoning in Language Models
Mizuho-DL Financial Technology Co., Ltd.
CausalBN-Bench: A Comprehensive Benchmark for Causal Learning Capability of LLMs
Apr 9, 2024Yu Zhou, Xingyu Wu, Jibin Wu +2LLM Evaluation Causal Discovery
Department of Data Science and Artificial Intelligence, The Hong Kong Polytechnic University, Hong Kong SAR 999077, China · College of Computer Science, Chongqing University, Chongqing 400044, China

Jun 2, 2026J/K move · Enter open · S save

CauTion: Knowing When to Trust LLMs for Ensemble Causal Discovery

Bo Peng, Kaiwen Wu, Sirui Chen, Zhiheng Wang+2

Shanghai AI Laboratory · Shanghai Innovation Institute · Shanghai Jiao Tong University+2

Causal Discovery

Open paper AskPDF arXiv

Causal discovery from observational data remains challenging due to the fundamental limitations of purely statistical methods, such as statistical distinguishability within equivalence classes and sensitivity to finite sample sizes. While large language models (LLMs) offer a promising source of domain knowledge to complement statistical inference, existing LLM-augmented methods are vulnerable to LLM errors and incur high token costs. Moreover, reliance on a single data-centric algorithm can make results sensitive to algorithm-specific biases. To address these limitations, we propose CauTion, a framework that reliably integrates LLM domain knowledge into an ensemble of statistical causal discovery algorithms through consensus filtering and LLM reliability estimation. CauTion proceeds in three stages. First, an algorithm ensemble utilizes a consensus voting to resolve up to 96% of edges on which algorithms agree, achieving near-perfect accuracy on the filtered consensus edges. Second, a trust-calibrated arbitration mechanism estimates the relative reliability of the LLM and the algorithms via an annotation-free trust calibration procedure, which is then utilized to govern a trust-weighted voting process that restricts LLM arbitration exclusively to edges with unreliable algorithmic evidence. Third, a cycle repair step is applied to guarantee the final causal graph is validly acyclic. Experiments on six datasets demonstrate that CauTion consistently outperforms both data-centric and LLM-augmented baselines, with larger gains on larger graphs and strong robustness to LLM errors. Code is available at https://github.com/OpenCausaLab/CauTion.

Abstract

Causal Reasoning in Language Models

Explore similar work

CauTion: Knowing When to Trust LLMs for Ensemble Causal Discovery

Causality Elicitation from Large Language Models

CausalBN-Bench: A Comprehensive Benchmark for Causal Learning Capability of LLMs