On the Limitations of Zero-Shot Classification of
Causal Relations by LLMs (Work in Progress)
Vani Kanjirangat⇤ , Alessandro Antonucci and Marco Zaffalon
Istituto Dalle Molle di Studi sull’Intelligenza Artificiale (IDSIA), USI-SUPSI, Lugano (Switzerland)
Abstract
We aim to explore and analyze the capabilities and limitations of the large language models in under-
standing and distinguishing causal sentences under a zero-shot setting. We experiment on a multi-class
dataset of direct causal, conditional causal, and correlational sentences. In the experiments, the GPT
and Falcon models are validated against a fine-tuned BERT model under different settings to explore
zero-shot capabilities in causality detection. Zero-shot approaches exhibit good performance in other
classification tasks, such as sentiment analysis or question answering. Yet, for this task, the fine-tuned
approach seems superior, and the situation does not change if language cues are added or a few-shot
setting is considered. This is a preliminary analysis of a work in progress. Still, the results suggest that
identifying causal relations is a particularly challenging task that is hard to address in a zero-shot setup.
Keywords
Large language models, zero-shot classification, few-shot classification, causal inference.
1. Introduction
The adoption of large language models (LLMs) [1, 2] is rapidly growing, primarily because of the
zero-shot capabilities exhibited by these tools in a wide range of natural language processing
tasks, such as sentiment analysis or recommendations and knowledge-intensive tasks, such as
question answering and domain-specific entity recognition [3, 4, 5, 6, 7]. Despite such popularity,
it is essential to understand the limitations and address questions such as: where can these
models fall back? What are the possibilities of such fallbacks? How can we improve their
performance beyond prompting engineering [8, 9, 10]?
This paper is a preliminary report on our work (in progress) on evaluating the potential
of state-of-the-art LLMs in the field of causal inference. More specifically, we investigate the
performance of LLMs in a classification task with sentences possibly involving causal relations.
Our analysis focuses on zero- and few-shot capabilities of LLMs compared against a fine-tuning
setting with encoder-based BERT models, which are nowadays the most common choice for
classification tasks [11, 12, 13]. Our tests show some limitations of LLM approaches in the causal
domain, this being the case either for zero-shot and few-shot setups. Notably, the situation
remains the same even if language cues are provided. Such negative results are in line with
some recent works presenting LLMs as causal parrots [14], not yet capable of genuine causal
reasoning [15], beyond just distinguishing between causes and effects [16, 17].
In: R. Campos, A. Jorge, A. Jatowt, S. Bhatia, M. Litvak (eds.): Proceedings of the Text2Story’24 Workshop, Glasgow
(Scotland), 24-March-2024
⇤
Corresponding author: Vani Kanjirangat (vani.kanjirangat@idsia.ch).
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
151
2. Related Work
Recently, a plethora of research has been going on in the direction of exploiting the zero-
shot and few-shot capabilities of LLMs. Because of the vast amount of pre-trained data they
have been exposed to, large (>10B parameters) language models are considered to have an
inherent ability to generalise across unseen tasks [18, 19, 20]. For instance, the number of
parameters of the recent GPT-3 and GPT-4 models is about, respectively, 175B and 1.76T. Zero-
shot and few-shot techniques have been tried with different prompting strategies (e.g., the
chain of thought) for both classification and generation tasks. In many knowledge-intensive
tasks (e.g., question answering), translations, classification tasks (e.g., sentiment analysis) and
recommendations, those approaches seem compelling, provided that an adequate, prompt
engineering effort is achieved [21, 22, 23, 24]. These techniques may be inaccurate for many
other tasks, significantly when the complexity increases, such as multi-task classifications and
hard sequence labelling tasks, especially in domain-specific problems [25, 8]. Researchers have
come up with soft prompting approaches and parameter efficient tuning (PEFT) [26] approaches
such as P-tuning [27], prompt-tuning [28, 29] and variations of prompt infusions to overcome
these problems, while trying to achieve fine-tuning-based performances. The causal reasoning
ability of LLMs has been initially investigated in [30]. The authors observe a good performance
with a pairwise causal discovery task, counterfactual reasoning task and actual causality by
conducting experiments on datasets of cause-effect pairs. A critical review of the causality
inference and reasoning with LLMs on benchmark datasets is reported in [31]. The authors
specify the requirements of causal datasets and problems of evaluations with LLMs, such as
memorisation (the dataset could be part of LLM pre-training). They also indicate that LLMs can
answer many datasets by simply computing similarities between options and questions in a
vector space. Further, they also indicate that the good performance of LLMs can sometimes be
due to spurious language cues in the datasets. In the rest of the paper, we explore the capability
of LLMs with simple prompt-based approaches in identifying causality on a multi-class dataset.
3. Dataset Settings
For our analysis, we focus on the dataset from [32], developed to automate the identification of
causal language use in the scientific literature. The data source was a collection of PubMed1
abstracts with five main health topics – nutrition, diabetes, obesity, breast cancer, and cholesterol.
Two domain experts were asked to annotate the sentences manually. A good agreement
(Cohen’s kappa = 0.98) was reported. The original dataset refers to a multi-class setup with
four options: correlational, direct causal, conditional causal, and one without any relations [33].
The entities possibly involved in the causal and correlational relations are not provided. Thus,
the identification depends on the specific language patterns used in the input sentences. In
the correlational case, the sentence describes some association between variables. With direct
causal sentences, the cause and effect are directly mentioned, while in the conditional case, the
relation definition carries out an element of doubt. Finally, there are sentences with neither
causation nor correlation.
1
https://pubmed.ncbi.nlm.nih.gov.
152
We use the original dataset in the native multi-class setting and a binary classification task.
For the binary class, we drop the correlational sentences and combine the direct and conditional
causal sentences, thus having only two classes, one with no relations and the other with causal
relations. This is intended to allow for a focus on causal relation discrimination. The multi-class
dataset includes 1356 no-relation, 494 direct causal, 213 conditional and 998 correlational cases,
which makes up 3061 cases. In the binary class setting, we have 1356 no-relation cases and 707
cases of causal relations (which combines direct and conditional cases), with 2063 cases overall.
4. Methodology
To test the capability of LLM models in classifying causal and non-causal sentences under
zero-shot settings, we initially design suitable prompts to tackle the task with LLMs. We use
both binary and multi-class settings with the prompt including the text input in Fig. 1 and some
variations. For the binary settings, we just need to change the classes in the prompt.
system_msg = You are a helpful assistant for causal reasoning and cause-and-effect relationship discovery.
Your aim is to identify the entities and to categorize the input sentences into either direct causal relation
or conditional causal relation or correlational relation or no relationship exist
intro_msg = You will be provided with a text. Text: {text}
instructions_msg = Please read the provided text carefully to comprehend the context and content.
Examine the roles, interactions, and details surrounding the entities within the text.
Based only on the information in the text, categorize the causal relation as
0. no relation
1. direct causal
2. conditional causal
3. correlational
Your response should analyze the situation in a step-by-step manner, ensuring the correctness of the ultimate conclusion,
which should accurately reflect the likely causal connection based on the information presented in the text.
If no clear causal relationship is apparent,
select the appropriate option accordingly, i.e., ’no relation’.
option_choice_msg = Your response should analyze the situation in a step-by-step manner,
ensuring the correctness of the ultimate conclusion,
which should accurately reflect the likely causal connection between the two entities based on the information presented in the text.
If no clear causal relationship is apparent, select the appropriate option accordingly.
Then provide your final answer within the tags [answer], (e.g. 1).
Figure 1: A zero-shot prompt for a causal recognition task.
Following the indications from [32] and findings from [31], we create another prompt in-
cluding language cues intended to help the LLM in providing more accurate classifications. In
fine-tuning approach, we assume the model automatically captures these patterns given the
training data. In a zero-shot setting, with the absence of such training information, we want
to see the impact on model performance when some explicit domain knowledge is available.
We added the following cues – association, associated with, predictor for correlational, increase,
decrease, lead to, effective in, contribute to, reduce for causal and along with may, might, appear to,
probably for conditional causal. These cues were then added to the zero-shot prompt (ZS-Cues).
Further, we tried them in a few-shot setup (FS-Cues) with some examples from each class (e.g.,
two samples for each class). Finally, we also consider a 500-shot experiment with labelled
samples, used also to train the BERT model under the same settings (500 samples for training).
153
5. Results and Discussion
For the fine-tuning approach, we use the best-base-cased model [34] in both binary and multi-
class settings with k-fold cross-validation. We use SimpleTransformers2 with four epochs and a
learning rate of 2E-5. We experiment with GPT (3.5 Turbo) and the open-source Falcon model
(falcon-7b-instruct 3 and falcon-40b-instruct) in zero-shot settings. Falcon-7b-instruct is a 7B
parameters causal decoder-only model fine-tuned on a mixture of chats and instructions, while
falcon-40b-instruct is a bigger model with 40B parameters. From Tab. 1, it can be observed
that Falcon-7b and 40b give inferior performance compared to GPT. Comparing the two Falcon
models, the 40b outperformed the 7b model in multi-class and binary settings. This expected
result motivates us to focus our experiments on GPT models only.
Table 1
Zero-shot (F1) performance of GPT and Falcon LLMs.
Model Approach Binary class Multi-class
GPT 3.5 turbo Zero shot (ZS) 0.59 0.37
Falcon-7b-instruct Zero shot (ZS) 0.19 0.27
Falcon-40b-instruct Zero shot (ZS) 0.26 0.38
For further experiments, we use GPT to analyse the performance under different prompt
settings and compare them with fine-tuned BERT-based models. Tab. 2 shows that, in both
settings, the performance of GPT under zero-shot settings is poor and the fine-tuned BERT model
performs better. In the multi-class case, many conditional causal relations are misclassified
as direct causal. Yet, the accuracy does not improve significantly in the binary setting. The
addition of cues (ZS-Cues) improves the performance, showing the importance of specific
patterns that help in classification, especially with multi-class settings. In both cases, ZS-Cues
performed better than FS-Cues. This could be because the sentences in this dataset are quite
varied (extracted from the scientific literature) and we cannot pre-assume that the selected
sentences for few-shot experiments could be the best representative for a class.
Table 2
Comparison of zero-shot, with and without cues, and few-shot LLMs against fine-tuned BERT.
Model Approach Binary class Multi-class
GPT 3.5 turbo Zero shot (ZS) 0.59 0.37
GPT 3.5 turbo Zero shot with Cues (ZS-Cues) 0.66 0.51
GPT 3.5 turbo Few shot with Cues (FS-Cues) 0.62 0.50
BERT-base-cased Full Fine Tuning (FFT) 0.92 0.87
Tab. 3 reports more details on the zero-shot results. The relatively high recall values for
causal sentences denote a good ability of the model in detecting direct causal relation. Yet, the
same does not happen with conditional causal relations, typically misclassified as direct ones.
This also explains the higher performance in the binary class case.
2
http://simpletransformers.ai.
3
https://huggingface.co/tiiuae/falcon-7b-instruct.
154
Table 3
Results on multi-class and binary settings for zero-shot classifications.
Multi-class Binary
No Rel. Causal Cond. Causal Corr. No Rel. Causal
F1-score 0.45 0.39 0.12 0.54 0.59 0.58
Precision 0.68 0.27 0.10 0.60 0.85 0.45
Recall 0.34 0.70 0.14 0.48 0.45 0.85
For a deeper comparison against the FFT model, we prompt the GPT model with more
examples. An option would be fine-tuning the GPT model, but we keep this as a future study,
as here the focus is on prompting approaches. Further, there are restrictions on the number
of prompt tokens processed by GPT 3.5 model. As a reasonable prompting solution, we use
500 samples (corresponding to a 1:4 train-test split). The same samples are used to train BERT
under multi-class settings. This proportion makes the BERT performance comparable with
the one with k-fold cross validation (F1=0.81), while a drastic drop is obtained with a 1:9 split
(F1=0.36). For GPT, this setup requires a slight change in the prompt (Fig. 1), to include a list
of input text and give the corresponding predictions as a list. We then chunked the remaining
2449 test samples, each containing ten samples, to be passed on to the prompt. These steps are
intended to optimise prompt efficiency in terms of costs and time. The results are in Tab. 4.
Table 4
BERT-Based FFT vs. GPT models with 500-shot training samples.
Model No Rel. Causal Cond. Causal Corr. Avg.
BERT-base-cased 0.86 0.77 0.74 0.86 0.81
GPT 3.5 turbo 0.61 0.42 0.12 0.55 0.43
It can be observed that, with 500 samples, the performance of GPT model was better than its
zero-shot counterpart, but is not comparable with BERT model fine-tuned with the same 500
training samples. This seems to confirm, in the causal domain, the general findings discussed in
[35]. At the same time, it is also notable that simply adding pattern information, like in ZS-Cues
and FS-Cues, makes the LLM performance better than the 500-shot model.
Moreover, in some cases, GPT gives predictions not explicitly mentioned in the prompts. For
instance, the prediction was multi-label (neither of the labels included true prediction) or the
prediction was categorised into a new class (not described in the prompt). For the evaluations,
we had to remove such samples. With zero-shot multi-class, finally, we had 3055 instances and
2016 instances with binary class.
Some explicit results from the zero-shot experiments are in Figs. 2, 3 and 4. Tab. 5 reports
examples of misclassification from the zero-shot prompting of GPT. Some of these instances
are hard to classify even for human experts. GPT almost always classifies a negative causality
as a no-relation case (I2, 3, and 4). Some other misclassifications (I8) predicted as causal while
conditional in practice are too obvious. Similar patterns are observed in correlational cases (I10).
It can be also observed that the addition of cues helped in the correct classification of some
instances. The correct predictions are in bold. E.g., I7 and I9 show that the cue may could have
helped in the correct classification as conditional causal, and for I12 the cue associated with.
155
text: However, obesity seems to be associated with more wound complications.
The text states that "obesity seems to be associated with more wound complications."
This suggests a potential causal relationship between obesity and wound complications.
To determine the type of causal relationship, we need to analyze the information provided.
1. Direct Causal Relation: This type of causal relation implies that one event directly causes another event.
In this case, if obesity directly causes wound complications, we would expect to see a clear cause-effect relationship between the two.
2. Conditional Causal Relation: This type of causal relation implies that one event causes another event
under certain conditions or circumstances. In this case, if obesity causes wound complications
only under specific conditions, we would need more information to determine the specific conditions.
3. Correlational Relation: This type of relation implies that two events are statistically related or occur together,
but one does not necessarily cause the other. In this case, if obesity and wound complications
are correlated but not causally related, we would expect to see a statistical association between the two.
Based on the information provided, it is not clear whether the relationship between obesity and wound complications
is a direct causal relation, a conditional causal relation, or a correlational relation.
The text only states that obesity is associated with more wound complications,
but it does not provide information on the specific nature of the relationship or any conditions under which the relationship occurs.
Therefore, the appropriate categorization for the causal relationship in this text is: 3 (correlational)
Figure 2: An example of correct correlational classification.
text: This study is one of the first proofs of concept showing that a nutrition-sensitive agriculture value chain approach
can contribute to improved child nutrition in a remote pastoralist population.
To analyze the causal relationship in the provided text, we need to identify the entities involved and understand their interactions.
Entities:
1. Study
2. Nutrition-sensitive agriculture value chain approach
3. Improved child nutrition
4. Remote pastoralist population
Based on the information in the text, let’s analyze the causal relationship:
1. The study is one of the first proofs of concept.
2. The study shows that a nutrition-sensitive agriculture value chain approach can contribute to improved child nutrition.
3. The improved child nutrition is in a remote pastoralist population.
From the above analysis, we can conclude that there is a direct causal relationship between the nutrition-sensitive agriculture
value chain approach and improved child nutrition in a remote pastoralist population.
Therefore, the causal relationship in the text can be categorized as a direct causal relation. 1
Figure 3: An example of correct classification of a direct causal relation.
text: There are no indications that endogenous and exogenous gonadal hormones affect the radiation dose-response relationship.
The provided text states that there are no indications that endogenous and exogenous gonadal hormones affect
the radiation dose-response relationship.
Based on this information, we can conclude that there is no clear causal relationship between endogenous and exogenous gonadal hormones
and the radiation dose-response relationship.
Therefore, the answer is 0 (no relation).
Figure 4: An example of a direct causal relation misclassified as no relation.
6. Limitations
Finally, as our paper presents the results of a work in progress, let us discuss the limitations of
the present work and the possible enhancements we might consider for the ongoing work.
We have used only one dataset; hence, whether our findings remain valid in the general case
might be questionable. The dataset facilitates understanding how well LLMs identify the causal
descriptions embedded in scientific literature under a more challenging multi-class setting,
156
Table 5
Sample sentences predictions from different prompts and ground truth (GT) values.
I Samples ZS ZS-C FS-C GT
1 CUSE was able to distinguish between benign and malignant breast masses 0 0 0 1
with high sensitivity and specificity.
2 The presence of an in situ stent did not interfere with surgery. 0 0 0 1
3 There are no indications that endogenous and exogenous gonadal hormones 0 0 0 1
affect the radiation dose-response relationship.
4 BRBC does not improve survival of women with MBC in this study, though 0 0 3 1
longer follow up is warranted.
5 The decreasing trends in nutritional status and appetite level after SCI require 1 0 0 0
special attention.
6 iCBT for depression is an efficacious, accessible treatment option for people 1 1 0 0
with diabetes.
7 CETP gene variation may affect coronary risk apart from the level of HDL-C 0 2 2 2
8 A considerable decline in RHR has occurred in Tromsø over the past two decades 0 3 3 3
in men and women of all ages.
9 Corticosteroids may delay the time of onset of severe skin reactions and also 1 1 2 2
reduce the incidence of severe radiation dermatitis
10 In HIV-infected patients, daptomycin appears to be a useful agent for treating 1 1 1 2
resistant GPIs.
11 Adverse drug events resulting in intensive care unit admission in oncology 1 1 1 3
patients are common and often associated with significant morbidity, mortality,
and cost.
12 In patients with obstructive CAD by CCTA, the baseline use of statins was 1 3 3 3
associated with improved clinical outcomes.
including correlative and causal relations. Distinguishing between direct and conditional
causation is especially difficult. To the best of our knowledge, there are no datasets with
analogous characteristics, at least for multi-class settings. Yet, manually annotating scientific
abstracts and creating new benchmarks for deeper validation is a realistic and necessary effort.
Moreover, in the current paper, we have focused on the GPT model and compared it with
the open-sourced Falcon and BERT-based models. This can be enhanced by comparing with
different LLMS. Further, the focus was on using prompt-based techniques, which have a broad
scope to be explored. Based on the finding from [31], we investigate techniques such as
incorporating language cues while prompting LLMs. One major problem is that LLMs can be
sensitive to manually engineered prompt designs; hence, automating prompts and using soft
prompt techniques would be the way forward.
7. Conclusions and Outlooks
This work is a preliminary exploration to understand the capabilities and limitations of the GPT
model in causality identifications, specifically in multi-class settings. The experiments show
that GPT has limited zero-shot and few-shot capabilities in capturing such causal relations,
subject to the data in consideration. Focusing on the limitations, in the future, we would like to
enhance our experiments on a range of causal data to have conclusive generalisations on the
studied facts. Prompt engineering as such has a lot of potential to be explored, while hard-core
engineering of prompts may not be always beneficial. Hence, we also plan to explore PEFT
techniques such as soft prompting for causal detection and further extractions of causal graphs.
157
References
[1] A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, Improving Language Understanding
by Generative Pre-Training, Technical Report, OpenAI, 2018.
[2] T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, Y. Iwasawa, Large language models are zero-shot
reasoners, Advances in Neural Information Processing Systems 35 (2022) 22199–22213.
[3] L. Shu, H. Xu, B. Liu, J. Chen, Zero-shot aspect-based sentiment analysis, arXiv preprint
arXiv:2202.01924 (2022).
[4] A. Kumar, N. Jain, S. Tripathi, C. Singh, From fully supervised to zero shot settings for
twitter hashtag recommendation, arXiv preprint arXiv:1906.04914 (2019).
[5] D. Teney, A. v. d. Hengel, Zero-shot visual question answering, arXiv preprint
arXiv:1611.05546 (2016).
[6] Z.-Y. Dou, N. Peng, Zero-shot commonsense question answering with cloze translation
and consistency optimization, in: Proceedings of the AAAI Conference on Artificial
Intelligence, volume 36, 2022, pp. 10572–10580.
[7] Y. Yang, A. Katiyar, Simple and effective few-shot named entity recognition with structured
nearest neighbor learning, arXiv preprint arXiv:2010.02405 (2020).
[8] L. Floridi, M. Chiriatti, GPT-3: Its nature, scope, limits, and consequences, Minds and
Machines 30 (2020) 681–694.
[9] K. Elkins, J. Chun, Can GPT-3 pass a writer’s Turing test?, Journal of Cultural Analytics 5
(2020).
[10] W. Wang, V. W. Zheng, H. Yu, C. Miao, A survey of zero-shot learning: Settings, methods,
and applications, ACM Transactions on Intelligent Systems and Technology (TIST) 10
(2019) 1–37.
[11] X. Wang, G. Chen, G. Qian, P. Gao, X.-Y. Wei, Y. Wang, Y. Tian, W. Gao, Large-scale
multi-modal pre-trained models: A comprehensive survey, Machine Intelligence Research
20 (2023) 447–482.
[12] V. Khetan, R. Ramnani, M. Anand, S. Sengupta, A. E. Fano, Causal BERT: Language
models for causality detection between events expressed in text, in: Intelligent Computing:
Proceedings of the 2021 Computing Conference, Volume 1, Springer, 2022, pp. 965–980.
[13] S. Aftan, H. Shah, A survey on BERT and its applications, in: 2023 20th Learning and
Technology Conference (L&T), IEEE, 2023, pp. 161–166.
[14] M. Ze evi , M. Willig, D. S. Dhami, K. Kersting, Causal parrots: Large language models
may talk causality but are not causal, Transactions on Machine Learning Research (2023).
[15] C. Zhang, D. Janzing, M. van der Schaar, F. Locatello, P. Spirtes, Causality in the time of
LLMs: Round table discussion results of CLeaR 2023, Proceedings of Machine Learning
Research vol TBD 1 (2023) 7.
[16] L. Zhiheng, Z. Jin, R. Mihalcea, M. Sachan, B. Schölkopf, Can large language models
distinguish cause from effect?, in: UAI 2022 Workshop on Causal Representation Learning,
2022.
[17] A. Antonucci, G. Piqué, M. Zaffalon, Zero-shot causal graph extrapolation from text via
LLMs, arXiv preprint arXiv:2312.14670 (2023).
[18] E. Almazrouei, H. Alobeidli, A. Alshamsi, A. Cappelli, R. Cojocaru, M. Debbah, É. Goffinet,
D. Hesslow, J. Launay, Q. Malartic, et al., The falcon series of open language models, arXiv
158
preprint arXiv:2311.16867 (2023).
[19] S. Zhuang, B. Liu, B. Koopman, G. Zuccon, Open-source large language models are strong
zero-shot query likelihood models for document ranking, arXiv preprint arXiv:2310.13243
(2023).
[20] Z. Wang, Y. Pang, Y. Lin, Large language models are zero-shot text classifiers, arXiv
preprint arXiv:2312.01044 (2023).
[21] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan,
P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, Advances in
Neural Information Processing Systems 33 (2020) 1877–1901.
[22] S. Kim, S. J. Joo, D. Kim, J. Jang, S. Ye, J. Shin, M. Seo, The CoT collection: Improving
zero-shot and few-shot learning of language models via chain-of-thought fine-tuning,
arXiv preprint arXiv:2305.14045 (2023).
[23] Y. Yu, Y. Zhuang, R. Zhang, Y. Meng, J. Shen, C. Zhang, Regen: Zero-shot text classi-
fication via training data generation with progressive dense retrieval, arXiv preprint
arXiv:2305.10703 (2023).
[24] Z. Yang, Z. Gan, J. Wang, X. Hu, Y. Lu, Z. Liu, L. Wang, An empirical study of GPT-3 for
few-shot knowledge-based vqa, in: Proceedings of the AAAI Conference on Artificial
Intelligence, volume 36, 2022, pp. 3081–3089.
[25] M. Moradi, K. Blagec, F. Haberl, M. Samwald, GPT-3 models are poor few-shot learners in
the biomedical domain, arXiv preprint arXiv:2109.02555 (2021).
[26] Z. Fu, H. Yang, A. M.-C. So, W. Lam, L. Bing, N. Collier, On the effectiveness of parameter-
efficient fine-tuning, in: Proceedings of the AAAI Conference on Artificial Intelligence,
volume 37, 2023, pp. 12799–12807.
[27] X. Liu, K. Ji, Y. Fu, W. Tam, Z. Du, Z. Yang, J. Tang, P-tuning: Prompt tuning can be
comparable to fine-tuning across scales and tasks, in: Proceedings of the 60th Annual
Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2022,
pp. 61–68.
[28] X. L. Li, P. Liang, Prefix-tuning: Optimizing continuous prompts for generation, arXiv
preprint arXiv:2101.00190 (2021).
[29] X. Liu, Y. Zheng, Z. Du, M. Ding, Y. Qian, Z. Yang, J. Tang, GPT understands, too, AI Open
(2023).
[30] E. Kıcıman, R. Ness, A. Sharma, C. Tan, Causal reasoning and large language models:
Opening a new frontier for causality, arXiv preprint arXiv:2305.00050 (2023).
[31] L. Yang, O. Clivio, V. Shirvaikar, F. Falck, A critical review of causal inference benchmarks
for large language models, in: AAAI 2024 Workshop on “Are Large Language Models
Simply Causal Parrots?”, 2023.
[32] B. Yu, Y. Li, J. Wang, Detecting causal language use in science findings, in: Proceedings of
the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th
International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019,
pp. 4664–4674.
[33] P. Sumner, S. Vivian-Griffiths, J. Boivin, A. Williams, C. A. Venetis, A. Davies, J. Ogden,
L. Whelan, B. Hughes, B. Dalton, et al., The association between exaggeration in health
related science news and academic press releases: retrospective observational study, BMJ
349 (2014).
159
[34] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of deep bidirectional
transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018).
[35] T. Schick, H. Schütze, It’s not just size that matters: Small language models are also
few-shot learners, in: Proceedings of the 2021 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language Technologies, 2021,
pp. 2339–2352.
160