iTAG: Inverse Design for Natural Text Generation with Accurate Causal Graph Annotations

Wenshuo Wang; Boyu Cao; Nan Zhuang; Wei Li

cs.CLApr 8, 2026

iTAG: Inverse Design for Natural Text Generation with Accurate Causal Graph Annotations

Authors: Wenshuo Wang, Boyu Cao, Nan Zhuang, Wei Li

Organizations: School of Future Technology, South China University of Technology, Guangzhou, China · China Mobile GBA Innovation Institute, Guangzhou, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China

Abstract

A fundamental obstacle to causal discovery from text is the lack of causally annotated text data for use as ground truth, due to high annotation costs. This motivates an important task of generating text with causal graph annotations. Early template-based generation methods sacrifice text naturalness in exchange for high causal graph annotation accuracy. Recent Large Language Model (LLM)-dependent methods directly generate natural text from target graphs through LLMs, but do not guarantee causal graph annotation accuracy. Therefore, we propose iTAG, which performs real-world concept assignment to nodes before converting causal graphs into text in existing LLM-dependent methods. iTAG frames this process as an inverse problem with the causal graph as the target, iteratively examining and refining concept selection through Chain-of-Thought (CoT) reasoning so that the induced relations between concepts are as consistent as possible with the target causal relationships described by the causal graph. iTAG demonstrates both extremely high annotation accuracy and naturalness across extensive tests, and the results of testing text-based causal discovery algorithms with the generated data show high statistical correlation with real-world data. This suggests that iTAG-generated data can serve as a practical surrogate for scalable benchmarking of text-based causal discovery algorithms.

Figures & tables

Figure 1: An example of the three-phase workflow of iTAG : i nverse design for T ext gener A tion with causal G raph.

Panel of Figure 2 — Algorithm 1 inverse-design-based concept substitution

Figure 2: Annotation accuracy of generated causal graphs across methods on claude-opus-4-1.

Method	fastText	TextCNN	TSCNN	RoBERTa	Human
Template-based	0.98	0.97	0.96	0.99	0.81
LLM-dependent	0.58	0.61	0.59	0.64	0.57
LLM-dependent+CA	0.55	0.57	0.56	0.60	0.54
LLM-dependent+ID	0.56	0.59	0.58	0.62	0.55
iTAG	0.52	0.54	0.53	0.57	0.51

Table 1: Naturalness and indistinguishability of generated text across methods on claude-opus-4-1.

Figure 3: Transferability of causal discovery evaluation on claude-opus-4-1. Rows (top to bottom) are RuleBayes, SCITE, and LLM-CG; columns compare iTAG-generated corpora (left) vs. matched real-world corpora (right) across $n{=}3$ – $10$ . We report $F1_{\mathrm{G}}$ ( $\uparrow$ ), SHD ( $\downarrow$ ), and SID ( $\downarrow$ , log-scale; plotted as $\mathrm{SID}+\varepsilon$ , $\varepsilon{=}10^{-3}$ ).

Pearson Corr.	Spearman Corr.	Linear Regr.
Metric Corr. p-value $F1_{\mathrm{G}}$ 0.928 $6.64\times 10^{-4}$ SHD 0.927 $7.69\times 10^{-4}$ SID 0.921 $1.78\times 10^{-3}$ Average 0.925 /	Metric Corr. p-value $F1_{\mathrm{G}}$ 0.926 $8.81\times 10^{-4}$ SHD 0.921 $1.33\times 10^{-3}$ SID 0.928 $5.27\times 10^{-4}$ Average 0.925 /	Metric R 2 95% CI $F1_{\mathrm{G}}$ 0.861 [0.703, 0.938] SHD 0.859 [0.699, 0.938] SID 0.848 [0.678, 0.932] Average 0.856 [0.694, 0.936]

Table 2: Centered (within-

n

) agreement between algorithm scores on iTAG-generated vs. real-world corpora (24 algorithm–size pairs), controlling for the shared monotonic dependence on

n

. We report Pearson

r

(two-sided

p

-values via a stratified permutation test that shuffles algorithm identities within each

n

-bucket;

B{=}10{,}000

permutations), Spearman

\rho

(two-sided), and linear-fit

R^{2}

with 95% CI.

Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Admissible domain	Tested range (wide)	Main sampling
Variable quantity $n$	$\{3,\ldots,10\}$	$\{3,\ldots,10\}$	$\mathrm{Unif}\{3,\ldots,10\}$
Expected density $p$	$(0,1)$	$[0.05,\,0.80]$	$\mathrm{Unif}(0.05,0.80)$
max_parents	$\{0,\ldots,n-1\}$	$\{1,\ldots,n-1\}$	$\mathrm{Unif}\{1,\ldots,n-1\}$
max_children	$\{0,\ldots,n-1\}$	$\{1,\ldots,n-1\}$	$\mathrm{Unif}\{1,\ldots,n-1\}$
Confounder ratio $\gamma_{c}$	$[0,1]$	$[0.00,\,0.80]$	$\mathrm{Unif}(0.00,0.80)$
Collider ratio $\gamma_{v}$	$[0,1]$	$[0.00,\,0.80]$	$\mathrm{Unif}(0.00,0.80)$

Appendix

Table 3: Phase 1 structural parameter domains, wide tested ranges, and the sampling distributions used in reported experiments. Aggregated metrics that pool over

n

use an equal number of samples per

n

(equivalently, a uniform mixture over

n

Swept parameter	Grid size	$\mathbf{F1_{\mathrm{Ga}}}$ range	SHD range	SID range	$\mathbf{F1_{\mathrm{D}}}$ range	Conclusion
$p$	$11$	$[0.948,\,0.971]$	$[0.86,\,1.24]$	$[0.42,\,0.68]$	$[0.508,\,0.536]$	small vs. changing $n$
max_parents	$\|\{1,\ldots,n-1\}\|$	$[0.952,\,0.968]$	$[0.91,\,1.18]$	$[0.46,\,0.62]$	$[0.511,\,0.528]$	negligible
max_children	$\|\{1,\ldots,n-1\}\|$	$[0.954,\,0.969]$	$[0.89,\,1.15]$	$[0.44,\,0.61]$	$[0.509,\,0.531]$	negligible
$\gamma_{c}$	$9$	$[0.950,\,0.967]$	$[0.92,\,1.21]$	$[0.45,\,0.64]$	$[0.507,\,0.534]$	small
$\gamma_{v}$	$9$	$[0.951,\,0.968]$	$[0.90,\,1.19]$	$[0.44,\,0.63]$	$[0.510,\,0.532]$	small
$\lambda$	$\|\{0,\ldots,n-2\}\|$	$[0.955,\,0.970]$	$[0.88,\,1.14]$	$[0.43,\,0.59]$	$[0.512,\,0.527]$	negligible

Appendix

Table 4: Wide-range one-factor sensitivity summary for Phase 1 parameters. Ranges are computed over the sweep grid as described in the text (mean metric at each grid point, then min/max over the grid). Overall, parameter-induced variations are small compared with the effects induced by changing

n

in the main figures.

Parameter	$p$ -value on $F1_{\mathrm{Ga}}$	partial $\eta^{2}$ on $F1_{\mathrm{Ga}}$	$p$ -value on $F1_{\mathrm{D}}$	partial $\eta^{2}$ on $F1_{\mathrm{D}}$
$p$	0.284	0.008	0.412	0.005
max_parents	0.517	0.004	0.623	0.003
max_children	0.489	0.005	0.571	0.004
$\gamma_{c}$	0.341	0.007	0.456	0.004
$\gamma_{v}$	0.378	0.006	0.502	0.004
$\lambda$	0.562	0.003	0.649	0.002

Appendix

Table 5: Stratified permutation ANOVA after controlling for

n

(10,000 permutations; Benjamini–Hochberg correction at

\alpha=0.05

). All corrected

p

-values exceed 0.05 and all partial

\eta^{2}<0.01

Provider	API model identifier	Usage in paper
Anthropic	claude-opus-4-1-20250805-thinking	Default backbone in main tables
OpenAI	gpt-5-pro-2025-10-06	LLM-CG backbone; robustness runs; verifier backbone option
DeepSeek	DeepSeek-R1	Robustness runs
Qwen	Qwen3-235B-A22B-Thinking-2507	Robustness runs

Appendix

Table 6: Backbone LLMs used in this paper.

Module	temp.	top- $p$	max tokens	Output
Phase 2 proposer/refiner ( Initial / Refine )	0.3	0.95	5000	JSON
Phase 2 verifier ( CounterfactualVerification )	0.2	1.00	5000	JSON
Phase 3 textual transformation	0.7	0.95	5000	JSON
LLM causal discovery prompt (for baselines/analysis)	0.0	1.00	5000	JSON

Appendix

Table 7: Decoding parameters used in API calls.

$n$	Template-based			LLM-dependent			LLM+CA			LLM+ID			iTAG
	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$
3	1.00	0.0	0.0	0.72	3.5	3.0	0.77	3.0	2.5	0.79	2.8	2.3	0.98	0.4	0.2
4	1.00	0.0	0.0	0.68	4.7	4.1	0.73	4.0	3.6	0.75	3.8	3.4	0.97	0.6	0.3
5	1.00	0.0	0.0	0.64	6.0	5.5	0.69	5.4	4.9	0.71	5.1	4.7	0.96	0.8	0.4
6	1.00	0.0	0.0	0.60	7.3	6.8	0.66	6.6	6.1	0.68	6.3	5.9	0.95	1.0	0.5
7	1.00	0.0	0.0	0.57	8.7	8.2	0.63	8.0	7.5	0.65	7.7	7.2	0.94	1.2	0.6

Appendix

Table 8: GPT-5-pro-2025-10-06: Experiment 1 (annotation accuracy) across different

n

Method	BERT	GPT-2	RoBERTa	DistilBERT	Avg.
Template-based	0.99	0.98	0.97	0.99	0.83
LLM-dependent	0.61	0.63	0.62	0.66	0.58
LLM+CA	0.58	0.60	0.59	0.63	0.56
LLM+ID	0.59	0.61	0.60	0.64	0.57
iTAG	0.54	0.56	0.55	0.59	0.53

Appendix

Table 9: GPT-5-pro-2025-10-06: Experiment 2 (naturalness; detectability AUC) averaged over

n\in\{3,\ldots,10\}

Algorithm	$n$	iTAG-generated corpora			Real-world corpora
		$\mathrm{F1}_{\mathrm{G}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{G}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$
RuleBayes	3	0.795	1.449	0.490	0.716	1.350	0.510
	4	0.661	2.902	1.706	0.659	2.770	1.510
	5	0.592	4.043	2.407	0.614	4.580	2.880
	6	0.555	5.516	4.722	0.553	6.110	4.790
	7	0.508	7.916	6.111	0.503	7.940	6.880

Appendix

Table 10: GPT-5-pro-2025-10-06: Experiment 3 transferability curves for different

n

Metric	Pearson Corr.	p-value	Spearman Corr.	p-value	$R^{2}$	95% CI
$\mathrm{F1}_{\mathrm{G}}$	0.923	$<10^{-4}$	0.891	$5.17\times 10^{-9}$	0.851	[0.725, 0.944]
$\mathrm{SHD}$	0.924	$<10^{-4}$	0.877	$1.85\times 10^{-8}$	0.855	[0.628, 0.957]
$\mathrm{SID}$	0.929	$<10^{-4}$	0.931	$1.42\times 10^{-11}$	0.864	[0.778, 0.941]
Average	0.925	/	0.900	/	0.856	/

Appendix

Table 11: GPT-5-pro-2025-10-06: Experiment 3 within-

n

centered agreement statistics.

$n$	Template-based			LLM-dependent			LLM+CA			LLM+ID			iTAG
	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$
3	1.00	0.0	0.0	0.70	3.8	3.2	0.75	3.2	2.8	0.77	3.0	2.6	0.97	0.5	0.3
4	1.00	0.0	0.0	0.66	5.1	4.5	0.71	4.4	4.0	0.73	4.2	3.8	0.96	0.7	0.4
5	1.00	0.0	0.0	0.62	6.5	6.0	0.67	5.9	5.4	0.69	5.6	5.2	0.95	0.9	0.5
6	1.00	0.0	0.0	0.58	8.0	7.4	0.63	7.3	6.8	0.65	7.0	6.5	0.94	1.1	0.6
7	1.00	0.0	0.0	0.55	9.6	9.0	0.60	8.9	8.3	0.62	8.5	7.9	0.93	1.3	0.7

Appendix

Table 12: DeepSeek-R1: Experiment 1 (annotation accuracy) across different

n

Method	BERT	GPT-2	RoBERTa	DistilBERT	Avg.
Template-based	0.99	0.98	0.98	0.99	0.84
LLM-dependent	0.63	0.65	0.64	0.68	0.60
LLM+CA	0.60	0.62	0.61	0.65	0.58
LLM+ID	0.61	0.63	0.62	0.66	0.59
iTAG	0.56	0.58	0.57	0.61	0.55

Appendix

Table 13: DeepSeek-R1: Experiment 2 (naturalness; detectability AUC) averaged over

n\in\{3,\ldots,10\}

Algorithm	$n$	iTAG-generated corpora			Real-world corpora
		$\mathrm{F1}_{\mathrm{G}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{G}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$
RuleBayes	3	0.791	1.548	0.442	0.716	1.350	0.510
	4	0.536	3.048	1.392	0.659	2.770	1.510
	5	0.526	4.016	2.397	0.614	4.580	2.880
	6	0.518	5.990	5.741	0.553	6.110	4.790
	7	0.501	6.070	8.028	0.503	7.940	6.880

Appendix

Table 14: DeepSeek-R1: Experiment 3 transferability curves for different

n

Metric	Pearson Corr.	p-value	Spearman Corr.	p-value	$R^{2}$	95% CI
$\mathrm{F1}_{\mathrm{G}}$	0.862	$<10^{-4}$	0.842	$2.74\times 10^{-7}$	0.743	[0.552, 0.896]
$\mathrm{SHD}$	0.901	$<10^{-4}$	0.879	$1.70\times 10^{-8}$	0.811	[0.575, 0.914]
$\mathrm{SID}$	0.893	$<10^{-4}$	0.865	$5.72\times 10^{-8}$	0.797	[0.672, 0.924]
Average	0.885	/	0.862	/	0.784	/

Appendix

Table 15: DeepSeek-R1: Experiment 3 within-

n

centered agreement statistics.

$n$	Template-based			LLM-dependent			LLM+CA			LLM+ID			iTAG
	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{Ga}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$
3	1.00	0.0	0.0	0.69	4.0	3.4	0.74	3.4	3.0	0.76	3.1	2.8	0.96	0.6	0.4
4	1.00	0.0	0.0	0.65	5.3	4.7	0.70	4.6	4.2	0.72	4.3	3.9	0.95	0.8	0.5
5	1.00	0.0	0.0	0.61	6.8	6.2	0.66	6.1	5.6	0.68	5.8	5.3	0.94	1.0	0.6
6	1.00	0.0	0.0	0.57	8.3	7.7	0.62	7.5	7.0	0.64	7.2	6.7	0.93	1.2	0.7
7	1.00	0.0	0.0	0.54	9.9	9.3	0.59	9.2	8.6	0.61	8.8	8.2	0.92	1.5	0.8

Appendix

Table 16: Qwen3-235B-A22B-Thinking-2507: Experiment 1 (annotation accuracy) across different

n

Method	BERT	GPT-2	RoBERTa	DistilBERT	Avg.
Template-based	0.99	0.98	0.98	0.99	0.84
LLM-dependent	0.62	0.64	0.63	0.67	0.59
LLM+CA	0.59	0.61	0.60	0.64	0.57
LLM+ID	0.60	0.62	0.61	0.65	0.58
iTAG	0.55	0.57	0.56	0.60	0.54

Appendix

Table 17: Qwen3-235B-A22B-Thinking-2507: Experiment 2 (naturalness; detectability AUC) averaged over

n\in\{3,\ldots,10\}

Algorithm	$n$	iTAG-generated corpora			Real-world corpora
		$\mathrm{F1}_{\mathrm{G}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$	$\mathrm{F1}_{\mathrm{G}}\uparrow$	$\mathrm{SHD}\downarrow$	$\mathrm{SID}\downarrow$
RuleBayes	3	0.549	1.347	0.520	0.716	1.350	0.510
	4	0.539	2.158	1.141	0.659	2.770	1.510
	5	0.524	3.453	2.549	0.614	4.580	2.880
	6	0.389	5.028	5.704	0.553	6.110	4.790
	7	0.379	5.108	5.784	0.503	7.940	6.880

Appendix

Table 18: Qwen3-235B-A22B-Thinking-2507: Experiment 3 transferability curves for different

n

Metric	Pearson Corr.	p-value	Spearman Corr.	p-value	$R^{2}$	95% CI
$\mathrm{F1}_{\mathrm{G}}$	0.879	$<10^{-4}$	0.845	$2.14\times 10^{-7}$	0.773	[0.586, 0.928]
$\mathrm{SHD}$	0.874	$<10^{-4}$	0.845	$2.14\times 10^{-7}$	0.765	[0.498, 0.936]
$\mathrm{SID}$	0.885	$<10^{-4}$	0.856	$9.51\times 10^{-8}$	0.783	[0.608, 0.936]
Average	0.879	/	0.849	/	0.774	/

Appendix

Table 19: Qwen3-235B-A22B-Thinking-2507: Experiment 3 within-

n

centered agreement statistics.

Backbone	Metric	Full ( $r/\rho/R^{2}$ )	w/o RuleBayes	w/o SCITE	w/o LLM-CG
GPT-5-pro-2025-10-06	$\mathrm{F1}_{\mathrm{G}}$	0.923/0.891/0.851	0.801/0.908/0.641	0.893/0.675/0.798	0.970/0.971/0.940
GPT-5-pro-2025-10-06	$\mathrm{SHD}$	0.924/0.877/0.855	0.953/0.945/0.909	0.836/0.796/0.699	0.957/0.970/0.916
GPT-5-pro-2025-10-06	$\mathrm{SID}$	0.929/0.931/0.864	0.938/0.927/0.880	0.904/0.911/0.817	0.964/0.973/0.929
GPT-5-pro-2025-10-06	Average	0.925/0.900/0.856	0.897/0.927/0.810	0.878/0.794/0.771	0.964/0.971/0.928
DeepSeek-R1	$\mathrm{F1}_{\mathrm{G}}$	0.862/0.842/0.743	0.775/0.869/0.600	0.940/0.934/0.883	0.863/0.963/0.745
DeepSeek-R1	$\mathrm{SHD}$	0.901/0.879/0.811	0.903/0.891/0.816	0.874/0.879/0.764	0.959/0.939/0.919

Appendix

Table 20: Leave-one-algorithm-out stability check for Experiment 3 (within-

n

centered).

Backbone	Metric	Pearson Corr. [95% CI]	Spearman Corr. [95% CI]	$R^{2}$ [95% CI]
GPT-5-pro-2025-10-06	$\mathrm{F1}_{\mathrm{G}}$	0.923 [0.849, 0.972]	0.891 [0.764, 0.966]	0.851 [0.725, 0.944]
GPT-5-pro-2025-10-06	$\mathrm{SHD}$	0.924 [0.793, 0.994]	0.877 [0.765, 0.964]	0.855 [0.628, 0.957]
GPT-5-pro-2025-10-06	$\mathrm{SID}$	0.929 [0.882, 0.970]	0.931 [0.869, 0.975]	0.864 [0.778, 0.941]
GPT-5-pro-2025-10-06	Average	0.925 [0.868, 0.965]	0.900 [0.818, 0.956]	0.856 [0.755, 0.921]
DeepSeek-R1	$\mathrm{F1}_{\mathrm{G}}$	0.862 [0.743, 0.942]	0.842 [0.765, 0.921]	0.743 [0.552, 0.896]
DeepSeek-R1	$\mathrm{SHD}$	0.901 [0.749, 0.957]	0.879 [0.798, 0.943]	0.811 [0.575, 0.914]

Appendix

Table 21: Stratified bootstrap stability for Experiment 3 (within-

n

centered),

B=10000

Variant	$F1_{\mathrm{Ga}}\uparrow$	SHD $\downarrow$	SID $\downarrow$	# LLM calls	Tokens / sample
iTAG (default Phase 3)	0.967	1.06	0.56	$1$	487
iTAG + Gen-ID (Phase 3)	0.974	0.89	0.44	$1+2K_{\text{gen}}$	3,842

Appendix

Table 22: Phase 3 ablation: default structure-preserving textual transformation vs. adding a bounded generation-time inverse-design loop (Gen-ID). All metrics are computed as corpus means under the Experiment 1 evaluation protocol (Section 4.1.3 ), aggregated over

n\in\{3,\dots,10\}

. Costs report the number of LLM calls per sample (one initial generation call plus, for Gen-ID,

K_{\text{gen}}

extraction calls and

K_{\text{gen}}

revision calls) and the average total token usage per sample.

	iTAG (default Phase 3)			iTAG + Gen-ID (Phase 3)
n	$F1_{\mathrm{Ga}}\uparrow$	SHD $\downarrow$	SID $\downarrow$	$F1_{\mathrm{Ga}}\uparrow$	SHD $\downarrow$	SID $\downarrow$
3	0.986	0.28	0.11	0.990	0.22	0.07
4	0.981	0.53	0.25	0.987	0.41	0.17
5	0.976	0.74	0.38	0.982	0.59	0.28
6	0.969	0.97	0.52	0.976	0.79	0.41
7	0.963	1.18	0.63	0.971	0.98	0.51

Appendix

Table 23: Phase 3 ablation results by variable quantity

n

(500 examples per

n

). Values are corpus means computed as in Experiment 1 (Section 4.1.3 ).

	Spurious edges per graph ( $FP$ )			Missed edges per graph ( $FN$ )
$n$	Median	Mean	95th pct.	Median	Mean	95th pct.
3	0	0.14	1	0	0.11	1
4	0	0.28	1	0	0.21	1
5	0	0.39	2	0	0.29	1
6	0	0.52	2	0	0.38	2
7	0	0.64	2	0	0.47	2

Appendix

Table 24: Error decomposition for iTAG in Experiment 1 by variable quantity

n

(500 examples per

n

). Spurious edges correspond to non-edge false positives (annotators infer a direct edge where

a_{ij}=0

at generation time), while missed edges correspond to omissions (annotators do not mark a generation-time edge as directly supported by the paragraph).

Setting	$\Delta_{300\to 500}(F1_{\mathrm{Ga}})$	$\Delta_{300\to 500}(\mathrm{SHD})$	$\Delta_{300\to 500}(\mathrm{SID})$
iTAG (Exp 1)	0.004	0.09	0.11
LLM-dependent (Exp 1)	0.008	0.24	0.43
LLM-dependent+CA (Exp 1)	0.006	0.17	0.29
LLM-dependent+ID (Exp 1)	0.007	0.21	0.38

Appendix

Table 25: Sample-size stability analysis for Experiment 1 metrics. Values are computed from the protocol in Section C.4.1 . The main paper fixes 500 examples per method per backbone because metrics stabilize well before 500 (Section 4.1.1 ).

Detector / Method	$k$ =50	$k$ =100	$k$ =200	$k$ =500	$k$ =1000
fastText / iTAG	$0.54\pm 0.04$	$0.53\pm 0.03$	$0.52\pm 0.02$	$0.52\pm 0.01$	$0.52\pm 0.01$
TextCNN / iTAG	$0.57\pm 0.04$	$0.55\pm 0.03$	$0.54\pm 0.02$	$0.54\pm 0.01$	$0.54\pm 0.01$
TSCNN / iTAG	$0.56\pm 0.04$	$0.54\pm 0.03$	$0.53\pm 0.02$	$0.53\pm 0.01$	$0.53\pm 0.01$
RoBERTa / iTAG	$0.61\pm 0.05$	$0.59\pm 0.03$	$0.58\pm 0.02$	$0.57\pm 0.01$	$0.57\pm 0.01$
fastText / Template-based	$0.97\pm 0.02$	$0.98\pm 0.01$	$0.98\pm 0.01$	$0.98\pm 0.01$	$0.98\pm 0.01$
TextCNN / Template-based	$0.96\pm 0.02$	$0.97\pm 0.01$	$0.97\pm 0.01$	$0.97\pm 0.01$	$0.97\pm 0.01$

Appendix

Table 26: Detector learning curves (

F1_{\mathrm{D}}

for detecting generated texts; lower is better) as a function of training size

2k

. Values are computed with mean

\pm

std over 3 random seeds, averaged over the three domains. The main paper reports the

k=500

column in Table 1 and notes that results are stable once

k

reaches 500 (Section 4.1.3 ).

$n$	fastText	TextCNN	TSCNN	RoBERTa	Human
3	0.51	0.53	0.52	0.56	0.50
4	0.53	0.55	0.54	0.58	0.52
5	0.51	0.54	0.52	0.57	0.50
6	0.54	0.56	0.55	0.59	0.53
7	0.52	0.53	0.53	0.56	0.51
8	0.50	0.52	0.51	0.55	0.49

Appendix

Table 27: Detection

F1_{\mathrm{D}}

(lower is better) for detecting iTAG-generated texts, broken down by variable quantity

n

. The last row reproduces the overall averages reported in Table 1 (aggregated over domains and

n=3

–

10

) for reference.

$n$	MIMIC-IV-Note	FinCausal 2025	JUSTICE	Total
3	38	42	35	115
4	54	58	51	163
5	72	68	74	214
6	83	79	86	248
7	78	82	76	236
8	67	64	69	200

Appendix

Table 28: Real-world bucket sizes

N_{n}

after filtering to

n\in\{3,\dots,10\}

. These bucket sizes are used when computing real-world corpus means in Experiment 3 (Section 4.1.2 ).

Explore similar work

Implicit Causal Graph Construction in Text via Chain Discovery
Apr 20, 2026Liesbeth Allein, Marie-Francine MoensCausal Discovery Causal Reasoning in Language Models
Department of Computer Science, KU Leuven, Belgium · Department of Electronics and Information Systems, Ghent University, Belgium
Why LLMs Fail at Causal Discovery and How Interventional Agents Escape
May 26, 2026Amartya Roy, Sonali ParbhooCausal Discovery Causal Reasoning in Language Models
SIRE, IIT Delhi and Robert Bosch GmbH, India · Imperial College London
CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists
May 25, 2026Junlin Yang, Dylan Zhang, Xiangchen Song +7Structural Causal Models LLM Agent Evaluation
Tsinghua University · University of Illinois Urbana-Champaign · Carnegie Mellon University +2

Apr 20, 2026J/K move · Enter open · S save

Implicit Causal Graph Construction in Text via Chain Discovery

Liesbeth Allein, Marie-Francine Moens

Department of Computer Science, KU Leuven, Belgium · Department of Electronics and Information Systems, Ghent University, Belgium

Causal Discovery Causal Reasoning in Language Models Causal Inference

Open paper AskPDF arXiv

Causal graphs in text are typically populated by observable, predefined events. In contrast, we study implicit causal graph construction from text by treating each described cause-effect pair as the begin- and endpoint of an underlying latent causal graph and using large language models (LLMs) to infer intermediate causal events. We compare end-to-end graph construction with methods that frame the task as causal chain discovery. In the latter, graphs are built either by aggregating inferred chains or by progressively expanding partial chains through an iterative search process. We further explore Wisdom of the Crowd extensions that access causal knowledge from multiple LLMs in post-hoc aggregation and collaborative inference settings. We analyze trade-offs among these approaches and evaluate the validity of inferred causal relations using a manually curated database of 1,560 scientifically validated causal pairs. This database-based evaluation is proposed as reliable, resource-efficient, and transferable to settings where ground-truth graphs are unavailable.

$n$	MIMIC-IV-Note	FinCausal 2025	JUSTICE	Total
3	38	42	35	115
4	54	58	51	163
5	72	68	74	214
6	83	79	86	248
7	78	82	76	236
8	67	64	69	200

$n$	MIMIC-IV-Note	FinCausal 2025	JUSTICE	Total
3	38	42	35	115
4	54	58	51	163
5	72	68	74	214
6	83	79	86	248
7	78	82	76	236
8	67	64	69	200

Abstract

Explore similar work

Implicit Causal Graph Construction in Text via Chain Discovery

Why LLMs Fail at Causal Discovery and How Interventional Agents Escape

CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists

$n$	MIMIC-IV-Note	FinCausal 2025	JUSTICE	Total
3	38	42	35	115
4	54	58	51	163
5	72	68	74	214
6	83	79	86	248
7	78	82	76	236
8	67	64	69	200