---
title: "Mitigating Causal Bias in LLMs via Potential Outcomes"
description: "The paper studies causal bias in LLMs on document-level Event Causality Identification (ECI), the task of deciding whether… 17 references."
authors: ["Yiheng Zhao", "Yuanliang Li", "Shreya Savant", "Jun Yan"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: EACL 2026"
source: "https://aclanthology.org/2026.findings-eacl.220/"
doi: "10.18653/v1/2026.findings-eacl.220"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/zhao-2026-mitigating-causal-bias-llms-potential/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhao-2026-mitigating-causal-bias-llms-potential/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/) / Mitigating Causal Bias in LLMs via Potential Outcomes Framework and Actual Causality Theory

# Mitigating Causal Bias in LLMs via Potential Outcomes Framework and Actual Causality Theory

**Authors:** [Yiheng Zhao](https://md.research.elamur.ai/free/en/authors/yiheng-zhao/), Yuanliang Li, Shreya Savant and [Jun Yan](https://md.research.elamur.ai/free/en/authors/jun-yan/) · **Year:** 2026 · **Venue:** [Findings of the Association for Computational Linguistics: EACL 2026](https://md.research.elamur.ai/free/en/venues/findings/) ([Findings of the Association for Computational Linguistics 2026](https://md.research.elamur.ai/free/en/venues/findings/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.findings-eacl.220/) · [PDF](https://aclanthology.org/2026.findings-eacl.220.pdf) · [DOI 10.18653/v1/2026.findings-eacl.220](https://doi.org/10.18653/v1/2026.findings-eacl.220) · [OpenAlex](https://openalex.org/W7140076111)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper studies causal bias in LLMs on document-level Event Causality Identification (ECI), the task of deciding whether one event in a text causes another. Earlier work mostly examined sentence-level cases and reported that models over-predict causality. To test this more carefully, the authors re-partition the EventStoryLine corpus into a structure-controlled benchmark. It varies text length, event count, and lexical distance between events one at a time, filters out confounding factors, and balances causal and non-causal pairs. Their evaluation shows that LLMs fail in different ways: some consistently hallucinate causal links, some consistently miss them, and some switch between the two depending on text structure. To reduce this bias, they frame ECI as a causal inference problem, drawing on the potential-outcomes framework and on Halpern and Pearl's (HP) account of actual causality. Instead of asking for a direct yes/no answer, the framework has LLMs estimate how likely the later event is with and without the earlier one, and calls the pair causal if the difference is positive. This lowers the Causal Bias Index (CBI) and raises Overall Accuracy (OA), making causal judgments more trustworthy.

## Key points
- **Benchmark:** The dataset is built from EventStoryLine with three parts, each split into three subsets:
  - text length: 0–300, 300–600, and 600–1,000 words;
  - events per document: 0–20, 20–40, and 40–60;
  - event distance: 0–50, 50–200, and 200–1,000, following MAVEN-ERE.
- **Method:** The framework asks a factual question and a counterfactual question modelled on HP condition AC2, then computes ΔP = P1 − P2.
  - A pair with ΔP > 0 is labelled causal.
  - GPT3.5-turbo and Qwen2.5-14B are combined as an ensemble with repeated sampling.
- **Results against GPT3.5-turbo:** CBI drops by 12.2%, 15.6%, and up to 19.4% across the three parts, with OA gains of 4.2–5.9%.
- **Results against Qwen2.5-14B:** CBI drops by 16.7%, 13.5%, and 5.4%, with OA gains of 3.2–6.5%.
- **Analysis and limitation:**
  - Adding LLaMA3.1-8B raises average OA. On short texts (0–300 words), it cuts CBI from 7.2% to 0.8%.
  - OA improves steadily as the number of samples grows.
  - The ensemble with sampling increases inference cost and latency.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Concordia University](https://md.research.elamur.ai/free/en/universities/concordia-university/) — University, Montreal, Canada · 13 works · 673 citations

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,954 works · [🤝 General Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/general-social-sciences/) · 254 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,894 works

## Cite

Cite as: Yiheng Zhao, Yuanliang Li, Shreya Savant and Jun Yan. 2026. Mitigating Causal Bias in LLMs via Potential Outcomes Framework and Actual Causality Theory. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{zhao2026mitigating,
  title = {{Mitigating Causal Bias in LLMs via Potential Outcomes Framework and Actual Causality Theory}},
  author = {Yiheng Zhao, Yuanliang Li, Shreya Savant and Jun Yan},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.220},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.220}
}
```

- [← previous Mitigating Action-Relation Hallucinations in LVLMs via Relation-aware Visual Enhancement](https://md.research.elamur.ai/free/en/papers/qin-2026-mitigating-action-relation-hallucinations-lvlms/)
- [next → Mitigating Context Interference for Reliable and Efficient Search Agents](https://md.research.elamur.ai/free/en/papers/xue-2026-mitigating-context-interference-reliable-efficient/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/41/)

## References

17 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. [Gpt-4 technical report](https://md.research.elamur.ai/free/en/papers/openai-2023-gpt-4-technical-report/). arXiv preprint arXiv:2303.08774.
2. Sadi A Assaf and Sadiq Al-Hejji. 2006. Causes of delay in large construction projects. International journal of project management, 24(4):349–357. [find](https://scholar.google.com/scholar?q=Sadi+A+Assaf+and+Sadiq+Al%2DHejji%2E+2006%2E+Causes+of+delay+in+large+construction+projects%2E+International+journal+of+project+management%2C+24%284%29%3A349%E2%80%93357%2E)
3. Ruichu Cai, Shengyin Yu, Jiahao Zhang, Wei Chen, Boyan Xu, and Keli Zhang. 2025. Dr. eci: Infusing large language models with causal knowledge for decomposed reasoning in event causality identification. [find](https://scholar.google.com/scholar?q=Ruichu+Cai%2C+Shengyin+Yu%2C+Jiahao+Zhang%2C+Wei+Chen%2C+Boyan+Xu%2C+and+Keli+Zhang%2E+2025%2E+Dr%2E+eci%3A+Infusing+large+language+models+with+causal+knowledge+for+decomposed+reasoning+in+event+causality+identificatio)
4. Cheng Liu, Wei Xiang, and Bang Wang. 2024. Identifying while learning for document event causality identification. arXiv preprint arXiv:2405.20608. [find](https://scholar.google.com/scholar?q=Cheng+Liu%2C+Wei+Xiang%2C+and+Bang+Wang%2E+2024%2E+Identifying+while+learning+for+document+event+causality+identification%2E+arXiv+preprint+arXiv%3A2405%2E20608%2E)
5. Jintao Liu, Zequn Zhang, Zhi Guo, Li Jin, Xiaoyu Li, Kaiwen Wei, and Xian Sun. 2023. Kept: Knowledge enhanced prompt tuning for event causality identification. Knowledge-based systems, 259:110064. [find](https://scholar.google.com/scholar?q=Jintao+Liu%2C+Zequn+Zhang%2C+Zhi+Guo%2C+Li+Jin%2C+Xiaoyu+Li%2C+Kaiwen+Wei%2C+and+Xian+Sun%2E+2023%2E+Kept%3A+Knowledge+enhanced+prompt+tuning+for+event+causality+identification%2E+Knowledge%2Dbased+systems%2C+259%3A110064%2E)
6. Qing Lyu, Kumar Shridhar, Chaitanya Malaviya, Li Zhang, Yanai Elazar, Niket Tandon, Marianna Apidianaki, Mrinmaya Sachan, and Chris Callison-Burch. 2025. Calibrating large language models with sample consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 1926… [find](https://scholar.google.com/scholar?q=Qing+Lyu%2C+Kumar+Shridhar%2C+Chaitanya+Malaviya%2C+Li+Zhang%2C+Yanai+Elazar%2C+Niket+Tandon%2C+Marianna+Apidianaki%2C+Mrinmaya+Sachan%2C+and+Chris+Callison%2DBurch%2E+2025%2E+Calibrating+large+language+models+with+sample+)
7. Hieu Man, Minh Van Nguyen, and Thien Huu Nguyen. 2022. Event causality identification via generation of important context words. In Proceedings of the 11th Joint Conference on Lexical and Computational Semantics (* SEM) at NAACL 2022. [find](https://scholar.google.com/scholar?q=Hieu+Man%2C+Minh+Van+Nguyen%2C+and+Thien+Huu+Nguyen%2E+2022%2E+Event+causality+identification+via+generation+of+important+context+words%2E+In+Proceedings+of+the+11th+Joint+Conference+on+Lexical+and+Computationa)
8. Paramita Mirza and Sara Tonelli. 2014. An analysis of causality between events and its relation to temporal information. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2097–2106. [find](https://scholar.google.com/scholar?q=Paramita+Mirza+and+Sara+Tonelli%2E+2014%2E+An+analysis+of+causality+between+events+and+its+relation+to+temporal+information%2E+In+Proceedings+of+COLING+2014%2C+the+25th+International+Conference+on+Computation)
9. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. [Training language models to follow instructions with human feedback](https://md.research.elamur.ai/free/en/papers/ouyang-2022-training-language-models-follow-instructions/). Advances in neural information processing systems, 35:27730–27744.
10. A As shown in Table 7 and Table 8, GPT3.5-turbo achieves the optimal F1 score compared to non-LLM methods, regardless of whether the event pairs are Intra-Sentence or Inter-Sentence. Method Konstantinos I Roumeliotis, Nikolaos D Tselikas, and Dimitrios K Nasiopoulos. 2023. Llama 2: Early adopters’ u… [find](https://scholar.google.com/scholar?q=A+As+shown+in+Table+7+and+Table+8%2C+GPT3%2E5%2Dturbo+achieves+the+optimal+F1+score+compared+to+non%2DLLM+methods%2C+regardless+of+whether+the+event+pairs+are+Intra%2DSentence+or+Inter%2DSentence%2E+Method+Konstantin)
11. Donald B Rubin. 1974. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology, 66(5):688. [find](https://scholar.google.com/scholar?q=Donald+B+Rubin%2E+1974%2E+Estimating+causal+effects+of+treatments+in+randomized+and+nonrandomized+studies%2E+Journal+of+educational+Psychology%2C+66%285%29%3A688%2E)
12. Shirong Shen, Heng Zhou, Tongtong Wu, and Guilin Qi. 2022. Event causality identification via derivative prompt joint learning. In Proceedings of the 29th international conference on computational linguistics, pages 2288–2299. [find](https://scholar.google.com/scholar?q=Shirong+Shen%2C+Heng+Zhou%2C+Tongtong+Wu%2C+and+Guilin+Qi%2E+2022%2E+Event+causality+identification+via+derivative+prompt+joint+learning%2E+In+Proceedings+of+the+29th+international+conference+on+computational+lin)
13. Zhengwei Tao, Zhi Jin, Yifan Zhang, Xiancai Chen, Xiaoying Bai, Yue Fang, Haiyan Zhao, Jia Li, and Chongyang Tao. 2024. A comprehensive evaluation on event reasoning of large language models. arXiv preprint arXiv:2404.17513. — [arXiv:2404.17513](https://arxiv.org/abs/2404.17513) [find](https://scholar.google.com/scholar?q=Zhengwei+Tao%2C+Zhi+Jin%2C+Yifan+Zhang%2C+Xiancai+Chen%2C+Xiaoying+Bai%2C+Yue+Fang%2C+Haiyan+Zhao%2C+Jia+Li%2C+and+Chongyang+Tao%2E+2024%2E+A+comprehensive+evaluation+on+event+reasoning+of+large+language+models%2E+arXiv+pr)
14. Xiaozhi Wang, Yulin Chen, Ning Ding, Hao Peng, Zimu Wang, Yankai Lin, Xu Han, Lei Hou, Juanzi Li, Zhiyuan Liu, and 1 others. 2022a. [Maven-ere: A unified large-scale dataset for event coreference, temporal, causal, and subevent relation extraction](https://md.research.elamur.ai/free/en/papers/wang-2022-maven-ere-unified-large-scale/). arXiv preprint arXiv:2211.07342.
15. Intra-Sentence P (%) R (%) F1 (%) DPJL KEPT CPATT 65.3 50.0 79.4 70.8 68.8 81.3 67.9 57.9 80.4 GPT3.5-turbo 78.5 83.3 80.8 *Table 7: Comparison experiment results between GPT3.5-turbo and non-LLMs methods on intra-sentence event pairs of EventStoryLine* Euysup Shim, Brad Carter, and Seongchan Kim. 2… [find](https://scholar.google.com/scholar?q=Intra%2DSentence+P+%28%25%29+R+%28%25%29+F1+%28%25%29+DPJL+KEPT+CPATT+65%2E3+50%2E0+79%2E4+70%2E8+68%2E8+81%2E3+67%2E9+57%2E9+80%2E4+GPT3%2E5%2Dturbo+78%2E5+83%2E3+80%2E8+%2ATable+7%3A+Comparison+experiment+results+between+GPT3%2E5%2Dturbo+and+non%2DLLMs+met)
16. Ya Su, Hu Zhang, Guangjun Zhang, Yujie Wang, Yue Fan, Ru Li, and Yuanlong Wang. 2025. Enhancing event causality identification with llm knowledge and concept-level event relations. In Proceedings of the 31st International Conference on Computational Linguistics, pages 7403–7414. [find](https://scholar.google.com/scholar?q=Ya+Su%2C+Hu+Zhang%2C+Guangjun+Zhang%2C+Yujie+Wang%2C+Yue+Fan%2C+Ru+Li%2C+and+Yuanlong+Wang%2E+2025%2E+Enhancing+event+causality+identification+with+llm+knowledge+and+concept%2Dlevel+event+relations%2E+In+Proceedings+of+t)
17. GPT3.5-turbo vs non-LLMs methods Method CPATT GPT3.5-turbo Inter-Sentence P (%) R (%) F1 (%) 74.9 69.4 60.1 78.6 66.7 73.7 *Table 8: Comparison experiment results between GPT3.5-turbo and non-LLMs methods on inter-sentence event pairs of EventStoryLine* [find](https://scholar.google.com/scholar?q=GPT3%2E5%2Dturbo+vs+non%2DLLMs+methods+Method+CPATT+GPT3%2E5%2Dturbo+Inter%2DSentence+P+%28%25%29+R+%28%25%29+F1+%28%25%29+74%2E9+69%2E4+60%2E1+78%2E6+66%2E7+73%2E7+%2ATable+8%3A+Comparison+experiment+results+between+GPT3%2E5%2Dturbo+and+non%2DLLMs+met)

*Updated 2 November 2026*