Generating Effective CoT Traces for Mitigating Causal Hallucination

Yiheng Zhao; Jun Yan

cs.CLApr 14, 2026

Generating Effective CoT Traces for Mitigating Causal Hallucination

Authors: Yiheng Zhao, Jun Yan

Organizations: Concordia University, Montreal, Canada

Abstract

Although large language models (LLMs) excel in complex reasoning tasks, they suffer from severe causal hallucination in event causality identification (ECI), particularly in smaller models ( $\leq$ 1.5B parameters). A promising approach to address this issue is to fine-tune them with Chain-of-Thought (CoT) traces. However, there is currently a lack of CoT trace dataset available for ECI. In this paper, we first investigate the essential criteria that effective CoT traces should possess to mitigate causal hallucination in smaller models. We then design a pipeline to generate CoT traces that meet these criteria. Moreover, since there is currently no metric for quantifying causal hallucination, we also introduce a new metric, the Causal Hallucination Rate (CHR), to quantify causal hallucination, guide the formulation of effective CoT trace criteria, and validate the effectiveness of our pipeline. Our experiments show that fine-tuning with the CoT traces generated by our pipeline not only substantially reduces causal hallucination in smaller LLMs but also improves mean accuracy. Moreover, the fine-tuned models exhibit strong cross-dataset and cross-difficulty generalization, as well as robustness under misleading intervention prompts.

Figures & tables

Figure 1: Illustration of causal hallucination in smaller models ( $\leq$ 1.5B parameters) on EventStoryLine Caselli and Vossen (2017) , where a large accuracy gap exists between causal and non-causal event pairs. Causal event pairs denote pairs with a causal relationship, while non-causal pairs do not. Accuracy is computed separately for each category as the proportion of correctly predicted instances.

Setting	PPL	Mean Token	CHR
Vanilla	-	-	84.55
Qwen	2.77	242	60.30
Llama	2.35	317	34.12
Perplexity	2.28	302	39.26
LongOnly	2.28	-	39.84

Table 1: Results validating whether the perplexity-based selection strategy is effective. “PPL" denotes perplexity, and “Mean Token" represents the average token length of each CoT trace. “Vanilla" denotes Qwen2.5-1.5B without CoT fine-tuning. “Qwen" and “Llama" use CoT traces generated by Qwen2.5-7B and Llama3.1-8B, respectively. “Perplexity" selects the lowest-perplexity trace across models, while “LongOnly" removes shorter Qwen traces from the perplexity-based selection set.

Setting	Mean Token	CHR
Vanilla	-	84.55
Qwen	242	59.79
Llama	317	34.68
LlamaThinking	482	30.60

Table 2: Comparison of CoT traces with different lengths. “Mean Token" represents the average token length of each CoT trace. “Vanilla" denotes Qwen2.5-1.5B without CoT fine-tuning. “Qwen” uses concise traces from Qwen2.5-7B, “Llama” uses longer traces from Llama3.1-8B, and “LlamaThinking” uses even longer CoT traces generated by Llama3.1-8B, guided by few-shot examples produced by Qwen3-235B-A22B (Thinking) Yang et al. (2025) .

Setting	PPL	CHR
Vanilla	-	84.55
Qwen	2.76	59.79
QwenR	2.75	55.98
Llama	2.35	34.68
LlamaR	2.36	42.81
LlamaR (Our prompt)	2.34	33.77

Table 3: Effect of applying rewriting on CoT traces of different lengths. “Qwen”, “Llama”, and “LlamaThinking” follow the same definitions as in Table 2 . “QwenR”, “LlamaR”, and “LlamaThinkingR” denote applying the prior study of rewriting prompt, while “LlamaR (our prompt)” applies our revised rewriting prompt.

Figure 2: Overview of the proposed CoT trace generation pipeline.

Methods	CHR	mAcc
GPT3.5-turbo	43.43	57.36
GPT4	53.30	51.40
Llama3.1-8B	60.59	58.97
Qwen3-30B-A3B	60.27	58.76
Dr.ECI	100.00	50.00
MuTQA	46.03	28.82

Table 4: Comparison results on EventStoryLine. “Vanilla” refers to the original model; “CoT” denotes zero-shot chain-of-thought prompting; “ICL” denotes in-context learning with the same exemplars as those used to elicit long CoT traces from Llama3.1-8B in our pipeline; “Binary” refers to fine-tuning with binary labels; and “Our” indicates fine-tuning using CoT trace data generated by our proposed pipeline.

Methods	CHR	mAcc
Qwen2.5-1.5B (Vanilla)	83.54	52.97
Qwen2.5-1.5B (w/o rewriting)	23.39	56.51
Qwen2.5-1.5B (w rewriting)	6.26	66.00
Llama3.2-1B (Vanilla)	76.43	55.58
Llama3.2-1B (w/o rewriting)	17.13	55.51
Llama3.2-1B (w rewriting)	9.14	63.44

Table 5: Ablation results on EventStoryLine. “Vanilla” refers to the original model; “w/o rewriting” denotes fine-tuning without rewriting the initial CoT traces that contain rich semantic explanations and reasoning steps. “w rewriting” denotes fine-tuning after rewriting these CoT traces using the target model.

Methods	PPL
Qwen2.5-1.5B (w/o rewriting)	3.23
Qwen2.5-1.5B (w rewriting)	3.18
Llama3.2-1B (w/o rewriting)	3.99
Llama3.2-1B (w rewriting)	3.65

Table 6: Perplexity comparison with and without CoT trace rewriting.

Causal-TimeBank
Methods	CHR	mAcc
Qwen2.5-1.5B (Vanilla)	84.55	51.53
Qwen2.5-1.5B (Our)	11.37	66.79
Llama3.2-1B (Vanilla)	67.79	45.85
Llama3.2-1B (Our)	4.26	61.51
MAVEN-ERE

Table 7: Cross-dataset results on Causal-TimeBank and MAVEN-ERE. “Vanilla” refers to the original model without any fine-tuning. “Our” denotes the model fine-tuned using the CoT trace data from the first fold of the five-fold cross-validation on EventStoryLine.

Methods	CHR	mAcc
Qwen2.5-1.5B (Vanilla)	54.52	59.76
Qwen2.5-1.5B (Our)	1.41	69.85
Llama3.2-1B (Vanilla)	93.94	51.62
Llama3.2-1B (Our)	4.51	59.45

Table 8: Document-level results on inter-sentence event pairs of EventStoryLine. “Vanilla” refers to the original model without any fine-tuning. “Our” denotes the model fine-tuned using the CoT trace data from the first fold of the five-fold cross-validation on EventStoryLine.

Methods	CHR	mAcc
Qwen2.5-1.5B (Vanilla)	83.54	52.97
Qwen2.5-1.5B (Our)	6.26	66.00
Qwen2.5-1.5B (Interv.)	2.26	63.08
Llama3.2-1B (Vanilla)	76.43	55.58
Llama3.2-1B (Our)	9.14	63.44
Llama3.2-1B (Interv.)	4.34	60.34

Table 9: Robustness results on EventStoryLine. “Vanilla” refers to the original model without fine-tuning. “Our” denotes the model fine-tuned with our pipeline and evaluated without any intervention prompts. “Interv.” represents the same fine-tuned model evaluated under incorrect intervention prompts during inference.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Training Set Statistics
Fold	Total	Pos	Neg
1	7021	2179	4842
2	7491	2122	5369
3	7378	2128	5250
4	7629	2213	5416
5	6697	2094	4603

Appendix

Table 10: Statistics of the five-fold split on EventStoryLine for both the training and test sets. “Pos” denotes causal samples, while “Neg” denotes non-causal samples.

Dataset	Total	Pos	Neg
Causal-TimeBank	7656	298	7358
MAVEN-ERE	19642	3359	16283

Appendix

Table 11: Statistics of the Causal-TimeBank and MAVEN-ERE datasets. “Pos” denotes causal event pairs, and “Neg” denotes non-causal event pairs.

Dataset	Total	Pos	Neg
EventStoryLine	13000	3000	10000

Appendix

Table 12: Document-level (inter-sentence) subset of the EventStoryLine. “Pos” denotes causal event pairs, and “Neg” denotes non-causal event pairs.

General Training Hyperparameters
Training Framework	TRL
Epochs	1
Per-device Batch Size	1
Gradient Accumulation Steps	8
Effective Batch Size	8
Learning Rate	$2\times 10^{-4}$

Appendix

Table 13: Training hyperparameters and LoRA configuration for fine-tuning.

Panel of Figure 3 — Figure 3: Qualitative comparison of Qwen2.5-1.5B outputs before and after fine-tuning on a non-causal event pair.

Panel of Figure 4 — Figure 4: Qualitative comparison of Llama3.2-1B outputs before and after fine-tuning on a non-causal event pair.

Qwen2.5-1.5B (Vanilla)
FPR	88.80
TNR	11.20
MCC	0.075
Qwen2.5-1.5B (Ours)
FPR	37.13 ( $\downarrow$ )
TNR	62.87 ( $\uparrow$ )

Appendix

Table 14: Additional evaluation results with FPR, TNR, and MCC on EventStoryLine. Lower FPR is better, while higher TNR and MCC are better. “Vanilla” refers to the original model and “Our” indicates fine-tuning using CoT trace data generated by our proposed pipeline.

LLM Hallucination Mitigation

Large language model hallucination mitigation develops ways to detect and reduce unsupported, fabricated, or inconsistent outputs.

See all 178 papers

CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models
Apr 13, 2026Linggang Kong, Lei Wu, Yunlong Zhang +4LLM Hallucination Detection Hallucination in Language Models
College of Electronic Engineering, National University of Defense Technology · Anhui Province Key Laboratory of Cyberspace Security Situation Awareness and Evaluation · Institute of Computer Application, China Academy of Engineering Physics
SERE: Structural Example Retrieval for Enhancing LLMs in Event Causality Identification
May 5, 2026Zhifeng Hao, Zhongjie Chen, Junhao Lu +5Causal Reasoning in Language Models Causal Reasoning
School of Computer Science, Guangdong University of Technology · College of Mathematics and Computer, Shantou University · Huawei Noah’s Ark Lab +1
CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness
Jul 21, 2026Ziming Wang, Yinghua Yao, Changwu Huang +2CoT Faithfulness Faithfulness of Language Model Explanations
Department of Computer Science and Engineering of Southern University of Science and Technology, Shenzhen, China · Center for Frontier AI Research, Agency for Science, Technology and Research (A*STAR), Singapore · School of AI and Liberal Arts, Beijing Normal-Hong Kong Baptist University, Zhuhai, China +1

Apr 13, 2026J/K move · Enter open · S save

CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models

Linggang Kong, Lei Wu, Yunlong Zhang, Xiaofeng Zhong+3

College of Electronic Engineering, National University of Defense Technology · Anhui Province Key Laboratory of Cyberspace Security Situation Awareness and Evaluation · Institute of Computer Application, China Academy of Engineering Physics

LLM Hallucination Detection Hallucination in Language Models Causal Interventions in Language Models LLM Interpretability

Open paper AskPDF arXiv

Despite the groundbreaking advancements made by large language models (LLMs), hallucination remains a critical bottleneck for their deployment in high-stakes domains. Existing classification-based methods mainly rely on static and passive signals from internal states, which often captures the noise and spurious correlations, while overlooking the underlying causal mechanisms. To address this limitation, we shift the paradigm from passive observation to active intervention by introducing CausalGaze, a novel hallucination detection framework based on structural causal models (SCMs). CausalGaze models LLMs' internal states as dynamic causal graphs and employs counterfactual interventions to disentangle causal reasoning paths from incidental noise, thereby enhancing model interpretability. Extensive experiments across four datasets and three widely used LLMs demonstrate the effectiveness of CausalGaze, especially achieving 3.3% improvement in AUROC on the TruthfulQA dataset compared to state-of-the-art baselines.

Abstract

LLM Hallucination Mitigation

Explore similar work

CausalGaze: Unveiling Hallucinations via Counterfactual Graph Intervention in Large Language Models

SERE: Structural Example Retrieval for Enhancing LLMs in Event Causality Identification

CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness