Synergizing Large Language Models and Theory for Human-like Causal Reasoning

Yanxi Zhang; Xin Cong; Zhong Zhang; Xiao Liu; Dongyan Zhao; Yesai Wu

cs.CLMay 13, 2025

Synergizing Large Language Models and Theory for Human-like Causal Reasoning

Authors: Yanxi Zhang, Xin Cong, Zhong Zhang, Xiao Liu, Dongyan Zhao, Yesai Wu

Organizations: Center for Data Science, AAIS, Peking University · Wangxuan Institute of Computer Technology, Peking University · Tsinghua University · State Key Laboratory of General Artificial Intelligence

Abstract

Performing causal reasoning akin to humans is fundamental to achieving strong artificial intelligence. However, existing studies on causal reasoning and large language models (LLMs) often underestimate the role of human causal cognition. Human causal reasoning is far more than recognizing causes and effects. Humans often factor in many other considerations when establishing a causal link between two events, such as norm violations, intentionality, morality, counterfactual scenarios, etc. To simulate such reasoning, it is intuitive to identify cause-effect pairs using token causality formalisms while modeling the factors that influence human causal judgment using findings from cognitive science experiments. To support the evaluation and modeling of human-like causal reasoning in LLMs, we introduce HCR-Bench, a challenging benchmark with 1,093 instances and detailed human causal reasoning step annotations, and HCR-Reasoner, an approach that systematically integrates theories of token causality and cognitive science into LLMs for human-like causal reasoning. Results suggest that HCR-Reasoner consistently improves LLMs' causal alignment with humans, and that theory-guided reasoning is highly effective.

Figures & tables

Figure 1: An example in HCR-Bench . To simulate human causal reasoning, one needs to consider both causal and cognitive factors.

Dataset	Type	Valid. Labels	Caus.	Cogn.
COPA Roemmele et al. (2011)	Type	✓	✗	✗
e-CARE Du et al. (2022)	Type	✓	✗	✗
CLadder Jin et al. (2023)	Type	✗	✓	✗
Corr2Cause Jin et al. (2024)	Type	✗	✓	✗
QRData Liu et al. (2024)	Type	✓	✗	✗
Big-Bench Srivastava et al. (2023)	Token	✓	✗	✗

Table 1: Comparison of causal reasoning benchmarks. Type: type vs. token causality? Valid. Labels: labels human-validated? Caus.: causal structures annotated? Cogn.: cognitive factors annotated?

Category	Statistic	Total
Size	# Samples	1,093
Story	# Sentences (per sample)	8.49
	# Words (per sample)	162.70
Query	# Queried causal events (per sample)	1.02
Reasoning	# Events (per sample)	3.39
	# Causal events (per sample)	2.39

Table 2: HCR-Bench statistics.

Figure 2: The overview of HCR-Reasoner .

Methods	Acc. (C.)	Acc. (I.)	Acc.
Human Average	-	-	69.60%
Qwen2.5-32B-Instruct	65.25%	80.43%	68.98%
+ HCR-Reasoner	70.14%	80.43%	72.67%
Qwen2.5-72B-Instruct	65.89%	77.83%	68.82%
+ HCR-Reasoner	71.63%	79.13%	73.48%
DeepSeek-V3	66.45%	77.61%	69.20%

Table 3: Results of the pilot study. Human average is directly cited from Suzgun et al. (2023) . See § D.1 for complete results and significance testing results.

Methods	Accuracy	CE $\checkmark$	OE $\checkmark$
Open-source LLMs
Qwen2.5-32B-Instruct	62.67%	-	-
+ zero-shot CoT	61.94%	-	-
+ manual CoT	62.85%	-	-
+ HCR-Reasoner	64.78%	95.91%	93.96%
Qwen2.5-72B-Instruct	64.87%	-	-

Table 4: The results of different LLMs on HCR-Bench . CE

\checkmark

and OE

\checkmark

are the proportions of correctly identified causal and outcome events, respectively.

Figure 3: Fine-grained accuracies of different factors.

Figure 4: Results of the causal analysis.

Model	Vanilla $\rightarrow$ FO $\rightarrow$ FT $\rightarrow$ HCR-Reasoner
Open-source LLMs
Qwen2.5-32B	62.67% $\rightarrow$ 61.67% $\rightarrow$ 64.59% $\rightarrow$ 64.78%
Qwen2.5-72B	64.87% $\rightarrow$ 64.41% $\rightarrow$ 67.98% $\rightarrow$ 67.52%
DeepSeek-V3	63.49% $\rightarrow$ 67.25% $\rightarrow$ 65.87% $\rightarrow$ 67.61%
Closed-source LLMs
Gemini-2.0	60.20% $\rightarrow$ 58.19% $\rightarrow$ 60.84% $\rightarrow$ 64.96%

Table 5: Ablation results on HCR-Bench . Green and red highlights denote performance rise/drop compared with Vanilla.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: The causal representation of the case.

Factor	Value	Definition
$f_{\mathrm{occur}}^{E}$	True	$E$ actually occurs in the causal setting.
	False	Otherwise.
$f_{\mathrm{order}}^{E}$	/	An integer starting from 0, denoting the temporal order of $E$ relative to other events in the causal setting. If two events occur simultaneously, they should share the same order.
$f_{\mathrm{tc}}^{E}$	True	$E$ is an token cause of $O$ by Def. 1 . If 1) $f_{\mathrm{occur}}^{E}\land f_{\mathrm{occur}}^{O}$ holds and 2) $f_{\mathrm{nc}}^{E}$ holds while allowing contingencies based on AC2, then $f_{\mathrm{tc}}^{E}$ holds. [Proof: AC1 holds if $f_{\mathrm{occur}}^{E}\land f_{\mathrm{occur}}^{O}$ holds; AC2 holds if $f_{\mathrm{nc}}^{E}$ holds while allowing contingencies; AC3 already holds since we only annotate minimal/atomic and non-overlapping causal events.]
	False	Otherwise.
$f_{\mathrm{nc}}^{E}$	True	$E$ is a necessary cause of $O$ by Def. 2 . If but for $E$ , $O$ would not have occurred, then $f_{\mathrm{nc}}^{E}$ holds.

Appendix

Table 6: Annotation guideline.

Methods	Acc. (C.)	Acc. (I.)	Acc.
Human Average	-	-	69.60 $\pm$ 0.00
Qwen2.5-32B-Instruct	65.25 $\pm$ 0.32	80.43 $\pm$ 0.00	68.98 $\pm$ 0.24
+ zero-shot CoT	61.70 $\pm$ 0.78	75.22 $\pm$ 1.74	65.03 $\pm$ 0.49
+ manual CoT	65.89 $\pm$ 0.59	79.13 $\pm$ 1.06	69.14 $\pm$ 0.34
+ HCR-Reasoner	70.14 $\pm$ 1.83	80.43 $\pm$ 0.00	72.67 $\pm$ 1.38
Qwen2.5-72B-Instruct	65.89 $\pm$ 1.21	77.83 $\pm$ 1.30	68.82 $\pm$ 1.04

Appendix

Table 7: Complete results of the pilot study.

Model	Comparison	Accuracy Diff	$p$ -value ( $\pm$ std)	Significant	Significant (Bonferroni)
Qwen2.5-32B-Instruct	HCR-Reasoner vs. Vanilla	$0.0369$	$0.0020\pm 0.0004$	$\surd$	$\surd$
	HCR-Reasoner vs. Zero-shot CoT	$0.0765$	$0.0021\pm 0.0004$	$\surd$	$\surd$
	HCR-Reasoner vs. Manual CoT	$0.0353$	$0.0022\pm 0.0004$	$\surd$	$\surd$
Qwen2.5-72B-Instruct	HCR-Reasoner vs. Vanilla	$0.0465$	$0.0022\pm 0.0004$	$\surd$	$\surd$
	HCR-Reasoner vs. Zero-shot CoT	$0.0754$	$0.0022\pm 0.0003$	$\surd$	$\surd$
	HCR-Reasoner vs. Manual CoT	$0.0642$	$0.0020\pm 0.0004$	$\surd$	$\surd$

Appendix

Table 8: Statistical significance of HCR-Reasoner on Big-Bench Hard causal judgment.

Explore similar work

Better Think Thrice: Learning to Reason Causally with Double Counterfactual Consistency
Feb 18, 2026Victoria Lin, Xinnuo Xu, Rachel Lawrence +4Counterfactual Reasoning in Language Models Causal Reasoning in Language Models
Carnegie Mellon University · Microsoft Research Cambridge · Microsoft Research India
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
CausalFlip: A Benchmark for LLM Causal Judgment Beyond Semantic Matching
Feb 23, 2026Yuzhe Wang, Yaochen Zhu, Jundong LiLLM Evaluation Causal Reasoning in Language Models
University of Virginia Charlottesville, Virginia, USA

Feb 18, 2026J/K move · Enter open · S save

Better Think Thrice: Learning to Reason Causally with Double Counterfactual Consistency

Victoria Lin, Xinnuo Xu, Rachel Lawrence, Risa Ueno+3

Carnegie Mellon University · Microsoft Research Cambridge · Microsoft Research India

Counterfactual Reasoning in Language Models Causal Reasoning in Language Models Reasoning Consistency in Language Models Causal Interventions in Language Models

Open paper AskPDF arXiv

Despite their strong performance on reasoning benchmarks, large language models (LLMs) have proven brittle when presented with counterfactual questions, suggesting weaknesses in their causal reasoning ability. While recent work has demonstrated that labeled counterfactual tasks can be useful benchmarks of LLMs' causal reasoning, producing such data at the scale required to cover the vast potential space of counterfactuals is limited. In this work, we introduce double counterfactual consistency (DCC), a lightweight inference-time method for measuring and guiding the ability of LLMs to reason causally. Without requiring labeled counterfactual data, DCC verifies a model's ability to execute two important elements of causal reasoning: causal intervention and counterfactual prediction. Using DCC, we evaluate the causal reasoning abilities of various leading LLMs across a range of reasoning tasks and interventions. Moreover, we demonstrate the effectiveness of DCC as a training-free test-time rejection sampling criterion and show that it can directly improve performance on reasoning tasks across multiple model families.

Abstract

Explore similar work

Better Think Thrice: Learning to Reason Causally with Double Counterfactual Consistency

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

CausalFlip: A Benchmark for LLM Causal Judgment Beyond Semantic Matching