Automated Knowledge Base Construction (2021) Conference paper

Reasoning with Transformer-based Models:

Deep Learning, but Shallow Reasoning

Chadi Helwe chadi.helwe@telecom-paris.fr

Chloé Clavel chloe.clavel@telecom-paris.fr
Fabian Suchanek suchanek@telecom-paris.fr
Télécom Paris, Institut Polytechnique de Paris, France

Abstract
Recent years have seen impressive performance of transformer-based models on different
natural language processing tasks. However, it is not clear to what degree the transform-
ers can reason on natural language. To shed light on this question, this survey paper
discusses the performance of transformers on different reasoning tasks, including mathe-
matical reasoning, commonsense reasoning, and logical reasoning. We point out successes
and limitations, of both empirical and theoretical nature.

1. Introduction
In recent years, language models have achieved impressive results on a variety of natural
language processing (NLP) tasks, such as recognizing textual entailment, machine reading
comprehension, and machine translation. Most of these language models are based on vari-
ants of the transformer architecture [Vaswani et al., 2017], for example BERT [Devlin et al.,
2019], T5 [Raffel et al., 2020a], and GPT-3 [Brown et al., 2020]. These models depend en-
tirely on the attention mechanism, and thus eliminate the need for recurrent computations
used by LSTMs [Hochreiter and Schmidhuber, 1997] and GRUs [Cho et al., 2014]. They can
easily learn long-range dependencies, and the computation can be parallelized efficiently.
Today’s models contain millions or even billions of parameters. The models are pre-trained
on large unlabeled corpora, and then later fine-tuned to tackle a specific NLP task. For
example, the pre-trained BERT can reply to questions such as the following:
Context: The iPhone is produced by [MASK].
Expected answer: Apple
Model answer: Apple
However, this performance is deceiving: If we introduce a trap word, the pre-trained BERT
model replies completely differently:
Context: Samsung. The iPhone is produced by [MASK].
Expected answer: Apple
Model answer: Samsung
Here, the BERT model got distracted by the additional word (a technique called misprim-
ing [Kassner and Schütze, 2020]). Thus, the question arises to what degree such models
really “understand” the natural language text, and to what degree they merely respond to
statistical cues. This question is of utmost importance, because if we start relying on such

1
 Helwe, Clavel, & Suchanek

language models, there is the danger that we obtain good responses only in common test
settings, and completely abstruse replies in less common settings.
In this survey paper, we shed light on this question by investigating some of the most
complex natural language tasks: those that involve reasoning. That is, we look at test data
sets that have explicitly been designed to test the limitations of transformer-based models,
and we investigate to what degree the models really “understand” these tasks. While several
survey papers have focused on transformer-based models and on their applications [Rogers
et al., 2020b, Qiu et al., 2020, Xia et al., 2020, Yates et al., 2021], the capabilities of
transformer-based models in reasoning tasks have so far not been surveyed. Our paper is
organized as follows. In Section 2, we describe some common pitfalls that all models need
to handle in order to reason on natural language text. Section 3 analyzes the performance
of transformer-based models on different reasoning tasks. In Section 4, we describe the
theoretical limitations of the transformer architecture, and show, in Sections 4.2-4.3, that
they impact natural language reasoning. We conclude in Section 5. The appendix contains
a detailed list of basic models (Appendix A) and challenging datasets (Appendix B), as well
as of the model performances (Appendix C).

2. Common Pitfalls for BERT-like Models

We discuss here some common pitfalls that any approach needs to handle in order to reason
on natural language. Our discussion focuses on BERT, but the phenomena may affect other
transformer-based models as well.

2.1 Negation

The pre-trained BERT model cannot differentiate between positive and negative state-
ments. As an example, take this sentence from the LAnguage Model Analysis (LAMA)
dataset [Petroni et al., 2019], where BERT performs well:
Context: Marcel Oopa died in the city of [MASK].
Expected answer: Paris
Model answer: Paris (-2.3), Lausanne (-3.3), Brussels (-3.3)
When Kassner and Schütze [2020] added the word “not”, BERT delivered the exact same
top-ranked result:
Context: Marcel Oopa did not die in the city of [MASK].
Expected answer: any city different from Paris
Model answer: Paris (-2.4), Helsinki (-3.5),Warsaw (-3.5)
This phenomenon was also confirmed by Ettinger [2020]. Kassner and Schütze [2020] show
that BERT can be fine-tuned to pay attention to the negation. Thus, it is essential to add
examples with negation to the training set. Niven and Kao [2019] point out that these
examples should be diverse enough to not rely only on the word “not”, and Hosseini et al.
[2021] propose an unlikelihood objective function to learn to differentiate between positive
and negative statements.

2
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

2.2 Mispriming
The ability to distinguish useful from distracting contexts is an essential building block for
any reasoning task. We have already seen an example of mispriming in the introduction.
Mispriming can, in principle, affect any task, and thus also reasoning in particular.
Interestingly, mispriming works only when the distracting word is of the same type as
the expected answer (companies, in our example). The pre-trained BERT is not easily
misled by primes of other types [Niven and Kao, 2019]. Misra et al. [2020] also show that
the problem of mispriming can be overcome by providing more context. In the following
sentence, for example, the mispriming fails:
Context: Samsung. The iPhone was produced by [MASK],
whose CEO was Steve Jobs
Expected answer: Apple
Model answer: Apple
This shows that, although there is some dependency on misprimes, their power decreases
when sentences provide more context.

2.3 Pattern Heuristics

Fine-tuned BERT models have a tendency to learn simple pattern-based heuristics. For
example, BERT can be trained to perform well on textual entailment in the MNLI dataset
[Williams et al., 2018]:
Premise: The actor and the professor mentioned the lawyer.
Hypothesis: The professor mentioned the lawyer.
Expected answer: Entailment
Model answer: Entailment
To better understand the performance of BERT, McCoy et al. [2019b] designed the HANS
(Heuristic Analysis for NLI Systems) dataset. HANS makes BERT fail as follows:
Premise: The doctors advised the presidents and the tourists.
Hypothesis: The presidents advised the tourists.
Expected answer: Non entailment
Model answer: Entailment
This shows that the model learned the “lexical overlap heuristic”, which assumes that a
premise entails all hypotheses constructed from words in the premise. This problem can be
addressed by adding more HANS-like examples to the training dataset.

2.4 Word Order

Different studies [Ettinger, 2020, Sankar et al., 2019, Pham et al., 2020, Gupta et al., 2021]
have shown that BERT-like models are unperturbed by grammatically incorrect sentences:
If presented with a sentence of randomly shuffled words, they will still reply correctly. This
insensitivity to order can also mislead textual entailment. For example, the pre-trained
BERT fine-tuned on the MNLI dataset fails to provide the correct answer in the following

3
 Helwe, Clavel, & Suchanek

case [McCoy et al., 2019b,a]:

Premise: The doctor visited the lawyer
Hypothesis: The lawyer visited the doctor
Expected answer: Non entailment
Model answer: Entailment
This issue can be solved by augmenting the training set with modified word order instances
with their respective labels or by fine-tuning the model on sensitive word ordering tasks
such as CoLA [Warstadt et al., 2019].

3. Types of Reasoning with Transformer-based Models

3.1 Horn Rule Reasoning
A rather simple way of logical reasoning is to infer a conclusion from a set of premises and
rules. Transformer-based models are able to perform such kind of reasoning [Clark et al.,
2020, Talmor et al., 2020, Betz et al., 2020] without any external knowledge, if both the
rules and the facts are mentioned explicitly in the text. They can even generate the proofs
[Saha et al., 2020, Tafjord et al., 2020, Gontier et al., 2020]. Here is an example from the
ParaRules dataset [Clark et al., 2020]:
Context: Fact 1: Erin is young.
Fact 2: Erin is not kind.
Fact 3: Peter is nice.
Rule 1: If someone is young and not kind then they are big.
Question: Is Erin big?
Expected answer: Conclusion: Erin is big.
Proof: (Fact 1 & Fact 2) → Rule 1 → Conclusion
In this task, the best model, a fine-tuned T5-11B, achieves an accuracy above 95% in proof
generation and question answering. A transformer-based model can thus solve the task
nearly perfectly.

3.2 Commonsense Reasoning

Commonsense reasoning is any reasoning task that requires background knowledge that
humans commonly have. For example, the instruction “Can you do a Napoleon for the
camera?” requires commonsense reasoning to realize that the word “Napoleon” expresses
a specific pose [Bender and Koller, 2020]. Several studies have shown that BERT learned a
certain amount of commonsense knowledge during pre-training [Petroni et al., 2019, Davi-
son et al., 2019, Bosselut et al., 2019, Zhou et al., 2020b, Cui et al., 2020]. Consider, for
example, the LAMA dataset [Petroni et al., 2019], which asks:
Context: Ravens can [MASK]
Expected answer: fly
Model answer: fly
The model (the pre-trained BERT-large) is able to recall such commonsense knowledge.

4
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

This good performance has prompted the research community to develop datasets that
specifically probe the commonsense reasoning of transformer models. Prominent datasets
are COSMOS QA [Huang et al., 2019], CommonsenseQA [Talmor et al., 2019], the Wino-
grad Schema Challenge [Levesque et al., 2012], SWAG [Zellers et al., 2018], ReCoRD [Zhang
et al., 2018], CODAH [Chen et al., 2019], and PIQA [Bisk et al., 2020]. Transformer-based
models can indeed achieve a high performance (often > 75%) on these datasets, but only
with additional methods. These include data augmentation techniques [Yang et al., 2020],
multi-task learning [Lourie et al., 2021], and fusing knowledge graphs into language models
[Xu et al., 2021]. The following is an example from the CommonsenseQA dataset [Talmor
et al., 2019]:
Question: Bats have many quirks, with the exception of ?
Expected Answer: Laying eggs
Model w/o knowledge graph fusing: Eating bugs
Model w/ knowledge graph fusing: Laying eggs

The above example shows that providing the model with information from a knowledge
graph helps the model to correctly answer the question. However, several studies [Forbes
et al., 2019, Zhou et al., 2020b, Lin et al., 2020, Boratko et al., 2020, Singh et al., 2021] show
that when the datasets are specifically changed to target the weaknesses of transformer-
based models (for example, by adversarial instances), the models fail. Here is an example
from the COM2SENSE dataset [Singh et al., 2021], which asks the model to judge whether
a given sentence is logically coherent or not:
Context: Expecting ten fish in the net, Sammy was
thrilled to see forty fish swimming in there.
Expected answer: Coherent
Model answer: Coherent
The authors created a counterpart to this question by modifying a few words:
Context: Expecting ten fish in the net, Sammy was
thrilled to see five fish swimming in there.
Expected answer: Incoherent
Model answer: Coherent
When the model (UnifiedQA-3B [Khashabi et al., 2020], a multi-task trained model) is
tricked this way, it fails to predict correctly. This shows that the model can fall prey to
relatively simple modifications, and does not really reason.

3.3 Event-based Commonsense Reasoning

Some commonsense reasoning tasks are concerned with the usual sequence of events. For
example, the TIMEDIAL dataset [Qin et al., 2021] evaluates temporal reasoning capabili-
ties in dialogs. The TORQUE dataset [Ning et al., 2020] asks temporal relation questions
such as which events have already finished, given a short passage of text. In a similar spirit,
the MCTACO dataset [Zhou et al., 2019] asks:

5
 Helwe, Clavel, & Suchanek

Context: Mr. Barco has refused US troops or advisers but has accepted US military aid.
Question: What happened after Mr. Barco accepted the military aid?
Choices: (A) the aid was denied, (B) he received the aid, (C) things started to progress

The best model is a fine-tuned BERT model that uses normalization to convert numerical
expressions such as “30 months” to “2.5 years”. It achieves an F1-score of 69.9% (while
human performance has an F1-score of 87.1%). In the same spirit, Zhou et al. [2021] de-
veloped TRACIE, a temporal reasoning textual entailment dataset that asks whether one
event preceded another one. The authors use distant supervision from Wikipedia, and
a symbolic reasoning model called SymTime. This approach predicts the end time of an
event by having two transformer models that predict the start time and the duration of this
event and symbolically compare them against the prediction of another start time event.
With this, the authors achieve an accuracy of about 71% (with variations for different sub-
tasks). Like the “normal” commonsense tasks, event-based tasks can be solved rather
well by transformer-based models. However, this works mainly when symbolic machinery
(such as date normalization and symbolic reasoning) or background knowledge (such as
Wikipedia) is added. Human performance, in the high nineties, remains unachieved.

3.4 Implicit Reasoning

We now turn to implicit reasoning tasks, where (different from the tasks in Section 3.1), the
rules and facts are not given explicitly. Many of these tasks can be solved by transformer-
based models. Here is an example from the SNLI dataset [Bowman et al., 2015]:
Premise: Three girls take cover under their umbrellas.
Hypothesis: Nobody has umbrellas.
Expected answer: Contradiction
Model answer: Contradiction
In this task, a RoBERTa-large model, trained with a few-shot learning method [Wang et al.,
2021a], achieves an accuracy of 93.1%. However, these datasets contain superficial cues that
the models can take advantage of [Schlegel et al., 2020, Huang and Zhu, 2021, Lin et al.,
2021]. To adequately evaluate the understanding of a model, several more challenging log-
ical reasoning tasks have been designed, which mostly take the form of machine reading
comprehension. LogiQA [Liu et al., 2020b], for example, is a multiple choice dataset trans-
lated from the National Civil Servants Examination of China:
Context: David knows Mr. Zhang’s friend Jack, and Jack knows David’s friend Ms. Lin.
Everyone of them who knows Jack has a master’s degree,
and everyone of them who knows Ms. Lin is from Shanghai.
Question: Who is from Shanghai and has a master’s degree?
Choices: (A) David (B) Jack (C) Mr. Zhang (D) Ms. Lin

The best language model is a pre-trained RoBERTa model [Liu et al., 2019] fine-tuned
on the training set and has an accuracy of 35.31% (while the best human performance is
96%) [Liu et al., 2020b]. Several other benchmarks in this vein also show bad performance:
ReClor [Yu et al., 2020], QuAIL [Rogers et al., 2020a], ConTRoL [Liu et al., 2020a], Strate-

6
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

gyQA [Geva et al., 2021], AR-LSAT [Zhong et al., 2021], and CLUTRR [Sinha et al., 2019].
This shows that transformer-based models are currently unable to build a representation
of a longer text and draw a logical conclusion from it. This weakness can be remedied to
some degree by adding symbolic representations on top of RoBERTa, such as graph-based
modules [Huang et al., 2021, Ouyang et al., 2021], or logical information [Wang et al.,
2021b]. Other approaches develop neuro-symbolic methods, which teach reasoning strate-
gies by gradient-based optimisation [Minervini et al., 2020], or combine probabilistic logic
programming with neural networks [Manhaeve et al., 2018]. Integrating logical information
into RoBERTa pushes the performance on the easier questions of ReClor to 81.4%. How-
ever, the more difficult questions of these datasets incur performances of 50%-60%. The
same is true for comparison-based tasks. The RICA dataset [Zhou et al., 2020a], for exam-
ple, asks:
Context: A prindag is smaller than a flurberg,
so a flurberg is [MASK] likely to contain a prindag.
Expected answer: more
Pre-trained and fine-tuned language models such as GPT-2 [Radford et al., 2019] and
RoBERTa achieve a dismal performance of 50% on unseen inferences. Thus, these models
are unable to learn comparisons between (fictitious) objects.

3.5 Mathematical Reasoning

Mathematical reasoning is the process of reasoning about different mathematical aspects

such as arithmetic operations, numerical comparison, counting, and sorting. The level of
complexity can range from solving simple mathematical equations to proving theorems. The
following is an example of a math problem that is not linguistically complex, taken from
the DeepMind mathematics dataset [Saxton et al., 2019]:
Context: Calculate -841880142.544 + 411127
Expected answer: -841469015.544
This task can be solved by GPT-3 [Henighan et al., 2020]. Along the same line, Lample and
Charton [2019] show that a transformer network can compute function integrals, and solve
differential equations. The next more complex problems are math word problems (MWP),
which consist of a short text that describes a mathematical problem (such as a one-unknown
math problem) and a question. This task requires a model to extract relevant information
from the text to perform mathematical reasoning to solve it. The most prominent MWP
datasets are MAWPS [Koncel-Kedziorski et al., 2016] and ASDiv-A [Miao et al., 2020] (both
for one-unknown arithmetic problems). However, these datasets can be solved by models
even when the order of the words is modified and when the questions are omitted, proving
that the models rely on heuristic patterns found in the problem narrative. To remove these
artifacts, Patel et al. [2021] developed the SVAMP dataset, which applies simple variations
to ASDiv-A. The following is an example:

7
 Helwe, Clavel, & Suchanek

Context: Jack had 8 pens and Mary had 5 pens.

Mary gave 3 pens to Jack.
Question: How many pens does Jack have now?
Expected answer: 8 + 3 = 11
On this dataset, a trained model achieves an accuracy of around 65%. Among the different
mathematical operators (+,-,/,*), the model accuracy ranges from 65.3% for divisions to
35.8% for multiplications. Also, the performance drops drastically when the equations have
more than two numbers or more than one operator.
An even more complicated dataset is MATH [Hendrycks et al., 2021], which consists of
competition problems in mathematics:
Context: Tom has a red marble, a green marble, a blue marble,
and three identical yellow marbles.
Question: How many different groups of two marbles can Tom choose?
Expected answer: There are two cases here: either Tom chooses
two yellow marbles (1 result), or he chooses two marbles of different
colors (( 42 )= 6 results). The total number of distinct pairs
of marbles Tom can choose is 1 + 6 = 7

Here, the best model, a fine-tuned GPT-2 model, achieves an accuracy of only 6.9%.
Another dataset at the boundary of what is currently feasible is the IsarStep bench-
mark [Li et al., 2021], which is concerned with mathematical proofs:
Context: 2b2 = a2 ⇒ [M issing P roposition] ⇒ ∃ c ∈ Z. a = 2c
Expected answer: a is even
The authors developed a hierarchical transformer model, which outperforms all the other
tested baselines with an accuracy of 22.8% for the top-1 prediction, and an accuracy of
35.2% for the top-10 predictions. Other mathematical theorem proving datasets in the
same spirit are HOList [Bansal et al., 2019] and MetaMathStep [Polu and Sutskever, 2020].
In conclusion, these tasks show that transformer-based models cannot be trained to “under-
stand” mathematical word problems and to “generate” mathematical proofs. In contrast to
simple mathematical problems (as the example we mentioned above), which are not linguis-
tically complex, such challenging tasks require more than huge transformer-based models
to achieve high performance.

3.6 Summary
In all of these reasoning tasks, transformer-based models rarely achieve human performance.
That is not surprising, given that they are general-purpose tools that feed mainly from
training data, and lack any symbolic machinery that is commonly considered essential for
such tasks. In fact, it is impressive that the models perform so well at all.
Among the different reasoning tasks, we find that when the transformer-based models
are explicitly given all the information required to perform deductive reasoning, such as
facts and rules, the models can easily learn logical reasoning. However, when this infor-
mation is stated only implicitly in the text or in the supervision, the models struggle. In

8
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

terms of commonsense reasoning, transformer-based models have a certain degree of com-

monsense knowledge learned during pre-training. However, they can be easily disrupted
with adversarial commonsense instances. They are also limited in logical reasoning over
events and physical commonsense.
We thus see that the strength of transformer-based models comes from two components:
simple patterns in the training data, combined with background knowledge from the pre-
training. This combination allows the models to perform well on tasks such as Horn Rule
Reasoning (where the model learns a pattern on the training data), simple commonsense
reasoning (where the answer was learned from the pretraining), and simple mathematical
calculations (where the model learns a pattern during training). However, when these el-
ements are absent, the models struggle. We have seen several commonsense datasets that
specifically avoid patterns, or use adversarial patterns. Here, the models fail. In particular,
textual understanding remains out of reach for now, if the tasks are sufficiently different
from each other to avoid patterns, and if fictional entities are used (for which no background
knowledge is available). Mathematical reasoning, too, falls into this category, if the tasks
do not follow a pattern.
This does not mean that the tasks would be unsolvable in general: Several studies [Huang
et al., 2021, Ouyang et al., 2021, Wang et al., 2021b, Yang et al., 2020, Lourie et al., 2021,
Xu et al., 2021] show that the addition of symbolic knowledge (such as date normalization,
quasi-logical reasoning, and graph-based modules) and the use of supplementary techniques
such as data augmentation, multi-task learning, and knowledge-base fusion improve the
performance. Such tools may thus hold the key to address even the harder reasoning
problems.

4. Impossible Reasoning Tasks

We have seen that transformer-based models can be trained and tuned to work on reasoning
tasks, and that some tasks require additional machinery. In this section, we turn to tasks
that BERT will never be able to solve without additional machinery, no matter the amount
of tuning.

4.1 Theoretical Limitations of Transformers

Hahn [2020] studied the theoretical limitations of transformers. The main limitations come
from the fact that self-attention does not have the same level of expressiveness as recurrent
models such as LSTMs. In particular, transformers cannot emulate a stack and finite-state
automata. Based on this insight, Hahn proved that transformer-based networks cannot
model two languages, known as Parity and Dyck-2. Parity is the set of bit strings where
the number of 1s is even. Dyck-2 is the language of strings that are balanced sequences
of round brackets “()” and square brackets “[]”. Hahn shows that for any transformer
network, we can find an integer N such that strings of these languages longer than N
cannot be recognized. That is: to recognize such strings, the number of heads and layers
of the model would have to increase with the input length N . Bhattamishra et al. [2020b]
have verified these results in practice, and showed that when the input length is bounded
during training, LSTM can generalize to longer instances, whereas transformer architectures
cannot. Other theoretical limitations are studied in Bhattamishra et al. [2020a]. These

9
 Helwe, Clavel, & Suchanek

limitations are of rather theoretical nature. And yet, they have a very concrete impact on
natural language reasoning. To show this, we designed two experiments: the light switch
task and the cake task. All our datasets and the code of the experiments can be found at
https://github.com/dig-team/FailBERT.

4.2 Light Switch Task

Our first task puts the Parity language into practice. The input is a word of the language
((a|b) and )*(a|b), with a =“I ate a pizza” and b =“I operated the light switch”, i.e.,
the input is a sentence that describes a sequence of these two activities. Assuming that
the light is off in the beginning, the goal is to determine whether the light is on in the end
(which corresponds to an odd number of switches). This is a simple reasoning task that a
school child can solve.
Context: I operated the light switch, and I ate a pizza, and I ate a pizza.
Expected answer: ON

We fine-tuned a pre-trained RoBERTa model for 50 iterations on 20k examples, each con-
taining up to 20 a’s and b’s. On the training and validation datasets, the model achieves
an F-score > 0.99. However, when testing on examples that contain more than 20 a’s and
b’s, we obtain on average a random precision of 0.50. This confirms that the theoretical
limitation of the transformer-based model has practical implications for natural language
reasoning.

4.3 Cake Task

Our next task puts the Dyck language into practice. The input to the task is a word of
the language S → |SS|aSa0 |bSb0 , where a =“I add a peanut layer to my cake”, a0 =“I eat
a peanut layer from my cake”, b =“I add a chocolate layer to my cake”, and b0 =“I eat a
chocolate layer from my cake” (with the conjunction “and” in suitable places). The goal is
to determine whether this sequence of steps is possible and the cake is gone. Again, this is
a simple reasoning task that a child can solve on a sheet of paper (or with suitable baking
tools).
Context: I add a peanut layer and I eat a peanut layer.
Expected answer: Yes

Context: I add a peanut layer and I eat a chocolate layer.

Expected answer: No

We fine-tuned a pre-trained RoBERTa model on 24k examples, each with up to 20 items,

and with nesting depths up to 15, for 50 iterations. Again, the model achieves an F-score
> 0.99 on the training and validation sets. However, when testing on examples that contain
more than 20 items, and on depths larger than 15, we obtain, as expected, on average
a dismal F-score of 0.55. This shows again that the theoretical limitation of BERT-like
models lead to very concrete limitations on natural language reasoning.

10
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

5. Conclusion
This survey paper has shown that transformer-based models can perform a shallow level
of reasoning on textual data but lack deeper reasoning capabilities. The first stumbling
stones are some common pitfalls for BERT-like models: word order, negation, shallow
patterns, and priming problems. The models have to be explicitly trained to deal with
these. We have then discussed several reasoning tasks – from the simple Horn rule reasoning
to the more complex commonsense, textual understanding, and mathematical tasks. On
these tasks, the performance of transformer-based models is significantly behind human
performance. One promising direction of research here is to add symbolical knowledge
to the system – an approach that has been pursued with success on some of the tasks.
However, we have also recalled that transformer-based models have theoretical limitations
in that they cannot model the languages Dyck-2 and Parity. We have shown on small
reasoning tasks that these theoretical limitations, too, can hinder reasoning on natural
language. Further research could explore how different types of positional encodings (such
as learned embeddings, sinusoidal embeddings, or CAPE [Likhomanenko et al., 2021]) and
different attention mechanisms (such as saturated attention [Merrill et al., 2021]) could help
the models overcome even these limitations.
Acknowledgements. This work was partially funded by ANR-20-CHIA-0012-01 (“NoRDF”).

References
Kshitij Bansal, Sarah Loos, Markus Rabe, Christian Szegedy, and Stewart Wilcox. Holist:
An environment for machine learning of higher order logic theorem proving. In Interna-
tional Conference on Machine Learning, 2019.

Emily M Bender and Alexander Koller. Climbing towards nlu: On meaning, form, and
understanding in the age of data. In Annual Meeting of the Association for Computational
Linguistics, 2020.

Gregor Betz, Christian Voigt, and Kyle Richardson. Critical thinking for language models.
arXiv preprint arXiv:2009.07185, 2020.

Satwik Bhattamishra, Kabir Ahuja, and Navin Goyal. On the ability of self-attention
networks to recognize counter languages. In Conference on Empirical Methods in Natural
Language Processing, 2020a.

Satwik Bhattamishra, Kabir Ahuja, and Navin Goyal. On the practical ability of recur-
rent neural networks to recognize hierarchical languages. In International Conference on
Computational Linguistics, 2020b.

Yonatan Bisk, Rowan Zellers, Ronan LeBras, Jianfeng Gao, and Yejin Choi. Piqa: Reason-
ing about physical commonsense in natural language. In AAAI Conference on Artificial
Intelligence, 2020.

11
 Helwe, Clavel, & Suchanek

Michael Boratko, Xiang Li, Tim O’Gorman, Rajarshi Das, Dan Le, and Andrew McCallum.
Protoqa: A question answering dataset for prototypical common-sense reasoning. In
Conference on Empirical Methods in Natural Language Processing, 2020.

Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz,
and Yejin Choi. Comet: Commonsense transformers for automatic knowledge graph
construction. In Annual Meeting of the Association for Computational Linguistics, 2019.

Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large
annotated corpus for learning natural language inference. In Conference on Empirical
Methods in Natural Language Processing, 2015.

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla
Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini
Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya
Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric
Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner,
Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models
are few-shot learners. In Advances in Neural Information Processing Systems, 2020.

Michael Chen, Mike D’Arcy, Alisa Liu, Jared Fernandez, and Doug Downey. Codah: An
adversarially-authored question answering dataset for common sense. In Workshop on
Evaluating Vector Space Representations for NLP, 2019.

Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi
Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations us-
ing rnn encoder–decoder for statistical machine translation. In Conference on Empirical
Methods in Natural Language Processing, 2014.

Peter Clark, Oyvind Tafjord, and Kyle Richardson. Transformers as soft reasoners over
language. arXiv preprint arXiv:2002.05867, 2020.

Leyang Cui, Sijie Cheng, Yu Wu, and Yue Zhang. Does bert solve commonsense task via
commonsense knowledge? arXiv preprint arXiv:2008.03945, 2020.

Joe Davison, Joshua Feldman, and Alexander Rush. Commonsense knowledge mining from
pretrained models. In Conference on Empirical Methods in Natural Language Processing
and the International Joint Conference on Natural Language Processing, 2019.

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of
deep bidirectional transformers for language understanding. In Conference of the North
American Chapter of the Association for Computational Linguistics, 2019.

Allyson Ettinger. What bert is not: Lessons from a new suite of psycholinguistic diagnostics
for language models. Transactions of the Association for Computational Linguistics, 8:
34–48, 2020.

Maxwell Forbes, Ari Holtzman, and Yejin Choi. Do neural language representations learn
physical commonsense? arXiv preprint arXiv:1908.02899, 2019.

12
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant.
Did aristotle use a laptop? a question answering benchmark with implicit reasoning
strategies. Transactions of the Association for Computational Linguistics, 2021.

Nicolas Gontier, Koustuv Sinha, Siva Reddy, and Chris Pal. Measuring systematic gener-
alization in neural proof generation with transformers. Advances in Neural Information
Processing Systems, 2020.

Ashim Gupta, Giorgi Kvernadze, and Vivek Srikumar. Bert & family eat word salad:
Experiments with text understanding. In AAAI Conference on Artificial Intelligence,
2021.

Ivan Habernal, Henning Wachsmuth, Iryna Gurevych, and Benno Stein. The argument
reasoning comprehension task: Identification and reconstruction of implicit warrants. In
Conference of the North American Chapter of the Association for Computational Lin-
guistics, 2018.

Michael Hahn. Theoretical limitations of self-attention in neural sequence models. Trans-

actions of the Association for Computational Linguistics, 2020.

Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang,
Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the
math dataset. arXiv preprint arXiv:2103.03874, 2021.

Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson,
Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for
autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020.

Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation,
1997.

Arian Hosseini, Siva Reddy, Dzmitry Bahdanau, R Devon Hjelm, Alessandro Sordoni, and
Aaron Courville. Understanding by understanding not: Modeling negation in language
models. In Conference of the North American Chapter of the Association for Computa-
tional Linguistics, 2021.

Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Cosmos QA: Machine
reading comprehension with contextual commonsense reasoning. In Conference on Em-
pirical Methods in Natural Language Processing and the International Joint Conference
on Natural Language Processing, 2019.

Shanshan Huang and Kenny Q Zhu. Statistically profiling biases in natural language rea-
soning datasets and models. arXiv preprint arXiv:2102.04632, 2021.

Yinya Huang, Meng Fang, Yu Cao, Liwei Wang, and Xiaodan Liang. Dagn: Discourse-
aware graph network for logical reasoning. In Conference of the North American Chapter
of the Association for Computational Linguistics, 2021.

13
 Helwe, Clavel, & Suchanek

Nora Kassner and Hinrich Schütze. Negated and misprimed probes for pretrained lan-
guage models: Birds can talk, but cannot fly. In Annual Meeting of the Association for
Computational Linguistics, 2020.

Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark,
and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system.
In Conference on Empirical Methods in Natural Language Processing, 2020.

Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi.
Mawps: A math word problem repository. In Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language Technologies, 2016.

Guillaume Lample and François Charton. Deep learning for symbolic mathematics. In
International Conference on Learning Representations, 2019.

Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge.
In Conference on the Principles of Knowledge Representation and Reasoning, 2012.

Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed,
Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence
pre-training for natural language generation, translation, and comprehension. In Annual
Meeting of the Association for Computational Linguistics, 2019.

Wenda Li, Lei Yu, Yuhuai Wu, and Lawrence Paulson. Isarstep: a benchmark for high-level
mathematical reasoning. In International Conference on Learning Representations, 2021.

Tatiana Likhomanenko, Qiantong Xu, Ronan Collobert, Gabriel Synnaeve, and Alex Ro-
gozhnikov. Cape: Encoding relative positions with continuous augmented positional
embeddings. arXiv preprint arXiv:2106.03143, 2021.

Bill Yuchen Lin, Seyeon Lee, Rahul Khanna, and Xiang Ren. Birds have four legs?! nu-
mersense: Probing numerical commonsense knowledge of pre-trained language models.
In Conference on Empirical Methods in Natural Language Processing, 2020.

Jieyu Lin, Jiajie Zou, and Nai Ding. Using adversarial attacks to reveal the statistical bias
in machine reading comprehension models. arXiv preprint arXiv:2105.11136, 2021.

Hanmeng Liu, Leyang Cui, Jian Liu, and Yue Zhang. Natural language inference in context–
investigating contextual reasoning over long texts. In AAAI Conference on Artificial
Intelligence, 2020a.

Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. Logiqa:
A challenge dataset for machine reading comprehension with logical reasoning. arXiv
preprint arXiv:2007.08124, 2020b.

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy,
Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized
bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.

14
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

Nicholas Lourie, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Unicorn on rain-
bow: A universal commonsense reasoning model on a new multitask benchmark. In AAAI
Conference on Artificial Intelligence, 2021.

Robin Manhaeve, Sebastijan Dumančić, Angelika Kimmig, Thomas Demeester, and Luc
De Raedt. Deepproblog: Neural probabilistic logic programming. Advances in Neural
Information Processing Systems, 2018.

R Thomas McCoy, Junghyun Min, and Tal Linzen. Berts of a feather do not generalize
together: Large variability in generalization across models with similar test set perfor-
mance. In Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks
for NLP, 2019a.

Tom McCoy, Ellie Pavlick, and Tal Linzen. Right for the wrong reasons: Diagnosing
syntactic heuristics in natural language inference. In Annual Meeting of the Association
for Computational Linguistics, 2019b.

William Merrill, Yoav Goldberg, Roy Schwartz, and Noah A Smith. On the power of
saturated transformers: A view from circuit complexity. arXiv preprint arXiv:2106.16213,
2021.

Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. A diverse corpus for evaluating and
developing english math word problem solvers. In Annual Meeting of the Association for
Computational Linguistics, 2020.

Pasquale Minervini, Sebastian Riedel, Pontus Stenetorp, Edward Grefenstette, and Tim
Rocktäschel. Learning reasoning strategies in end-to-end differentiable proving. In Inter-
national Conference on Machine Learning, 2020.

Kanishka Misra, Allyson Ettinger, and Julia Rayz. Exploring bert’s sensitivity to lexical
cues using tests from semantic priming. In Conference on Empirical Methods in Natural
Language Processing, 2020.

Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. Torque:
A reading comprehension dataset of temporal ordering questions. In Conference on Em-
pirical Methods in Natural Language Processing, 2020.

Timothy Niven and Hung-Yu Kao. Probing neural network comprehension of natural lan-
guage arguments. In Annual Meeting of the Association for Computational Linguistics,
2019.

Siru Ouyang, Zhuosheng Zhang, and Hai Zhao. Fact-driven logical reasoning. arXiv preprint
arXiv:2105.10334, 2021.

Arkil Patel, Satwik Bhattamishra, and Navin Goyal. Are nlp models really able to solve
simple math word problems? In Conference of the North American Chapter of the
Association for Computational Linguistics, 2021.

15
 Helwe, Clavel, & Suchanek

Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang
Wu, and Alexander Miller. Language models as knowledge bases? In Conference on Em-
pirical Methods in Natural Language Processing and the International Joint Conference
on Natural Language Processing, 2019.

Thang M Pham, Trung Bui, Long Mai, and Anh Nguyen. Out of order: How important
is the sequential order of words in a sentence in natural language understanding tasks?
arXiv preprint arXiv:2012.15180, 2020.

Stanislas Polu and Ilya Sutskever. Generative language modeling for automated theorem
proving. arXiv preprint arXiv:2009.03393, 2020.

Lianhui Qin, Aditya Gupta, Shyam Upadhyay, Luheng He, Yejin Choi, and Manaal Faruqui.
Timedial: Temporal commonsense reasoning in dialog. arXiv preprint arXiv:2106.04571,
2021.

Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. Pre-
trained models for natural language processing: A survey. Science China Technological
Sciences, 2020.

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.
Language models are unsupervised multitask learners. OpenAI blog, 2019.

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael
Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning
with a unified text-to-text transformer. Journal of Machine Learning Research, 2020a.

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael
Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning
with a unified text-to-text transformer. Journal of Machine Learning Research, 2020b.

Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. Getting closer to
ai complete question answering: A set of prerequisite real tasks. In AAAI Conference on
Artificial Intelligence, 2020a.

Anna Rogers, Olga Kovaleva, and Anna Rumshisky. A primer in bertology: What we know
about how bert works. Transactions of the Association for Computational Linguistics,
2020b.

Swarnadeep Saha, Sayan Ghosh, Shashank Srivastava, and Mohit Bansal. Prover: Proof
generation for interpretable reasoning over rules. In Conference on Empirical Methods in
Natural Language Processing, 2020.

Chinnadhurai Sankar, Sandeep Subramanian, Christopher Pal, Sarath Chandar, and Yoshua
Bengio. Do neural dialog systems use the conversation history effectively? an empirical
study. In Annual Meeting of the Association for Computational Linguistics, 2019.

David Saxton, Edward Grefenstette, Felix Hill, and Pushmeet Kohli. Analysing mathe-
matical reasoning abilities of neural models. In International Conference on Learning
Representations, 2019.

16
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

Viktor Schlegel, Marco Valentino, André Freitas, Goran Nenadic, and Riza Theresa Batista-
Navarro. A framework for evaluation of machine reading comprehension gold standards.
In Language Resources and Evaluation Conference, 2020.

Shikhar Singh, Nuan Wen, Yu Hou, Pegah Alipoormolabashi, Te-Lin Wu, Xuezhe Ma, and
Nanyun Peng. Com2sense: A commonsense reasoning benchmark with complementary
sentences. arXiv preprint arXiv:2106.00969, 2021.

Koustuv Sinha, Shagun Sodhani, Jin Dong, Joelle Pineau, and William L Hamilton. Clutrr:
A diagnostic benchmark for inductive reasoning from text. In Conference on Empirical
Methods in Natural Language Processing and the International Joint Conference on Nat-
ural Language Processing, 2019.

Oyvind Tafjord, Bhavana Dalvi Mishra, and Peter Clark. Proofwriter: Generating im-
plications, proofs, and abductive statements over natural language. arXiv preprint
arXiv:2012.13048, 2020.

Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsenseqa:
A question answering challenge targeting commonsense knowledge. In Conference of the
North American Chapter of the Association for Computational Linguistics, 2019.

Alon Talmor, Oyvind Tafjord, Peter Clark, Yoav Goldberg, and Jonathan Berant. Leap-of-
thought: Teaching pre-trained models to systematically reason over implicit knowledge.
arXiv preprint arXiv:2006.06609, 2020.

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances
in Neural Information Processing Systems, 2017.

Cunxiang Wang, Shuailong Liang, Yue Zhang, Xiaonan Li, and Tian Gao. Does it make
sense? and why? a pilot study for sense making and explanation. In Annual Meeting of
the Association for Computational Linguistics, 2019.

Sinong Wang, Han Fang, Madian Khabsa, Hanzi Mao, and Hao Ma. Entailment as few-shot
learner. arXiv preprint arXiv:2104.14690, 2021a.

Siyuan Wang, Wanjun Zhong, Duyu Tang, Zhongyu Wei, Zhihao Fan, Daxin Jiang, Ming
Zhou, and Nan Duan. Logic-driven context extension and data augmentation for logical
reasoning of text. arXiv preprint arXiv:2105.03659, 2021b.

Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. Neural network acceptability
judgments. Transactions of the Association for Computational Linguistics, 2019.

Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for
sentence understanding through inference. In Conference of the North American Chapter
of the Association for Computational Linguistics, 2018.

Patrick Xia, Shijie Wu, and Benjamin Van Durme. Which* bert? a survey organizing
contextualized encoders. In Conference on Empirical Methods in Natural Language Pro-
cessing, 2020.

17
 Helwe, Clavel, & Suchanek

Yichong Xu, Chenguang Zhu, Ruochen Xu, Yang Liu, Michael Zeng, and Xuedong Huang.
Fusing context into knowledge graph for commonsense reasoning. In Annual Meeting of
the Association for Computational Linguistics, 2021.

Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras,
Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. G-daug: Generative
data augmentation for commonsense reasoning. In Conference on Empirical Methods in
Natural Language Processing, 2020.

Andrew Yates, Rodrigo Nogueira, and Jimmy Lin. Pretrained transformers for text ranking:
Bert and beyond. In ACM International Conference on Web Search and Data Mining,
2021.

Weihao Yu, Zihang Jiang, Yanfei Dong, and Jiashi Feng. Reclor: A reading comprehension
dataset requiring logical reasoning. In International Conference on Learning Representa-
tions, 2020.

Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. Swag: A large-scale adversarial
dataset for grounded commonsense inference. In Conference on Empirical Methods in
Natural Language Processing, 2018.

Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag:
Can a machine really finish your sentence? In Annual Meeting of the Association for
Computational Linguistics, 2019.

Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin
Van Durme. Record: Bridging the gap between human and machine commonsense read-
ing comprehension. arXiv preprint arXiv:1810.12885, 2018.

Wanjun Zhong, Siyuan Wang, Duyu Tang, Zenan Xu, Daya Guo, Jiahai Wang, Jian Yin,
Ming Zhou, and Nan Duan. Ar-lsat: Investigating analytical reasoning of text. arXiv
preprint arXiv:2104.06598, 2021.

Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. “going on a vacation” takes
longer than “going for a walk”: A study of temporal commonsense understanding. In
Conference on Empirical Methods in Natural Language Processing and the International
Joint Conference on Natural Language Processing, 2019.

Ben Zhou, Kyle Richardson, Qiang Ning, Tushar Khot, Ashish Sabharwal, and Dan Roth.
Temporal reasoning on implicit events from distant supervision. In North American
Chapter of the Association for Computational Linguistics, 2021.

Pei Zhou, Rahul Khanna, Bill Yuchen Lin, Daniel Ho, Xiang Ren, and Jay Pujara. Can bert
reason? logically equivalent probes for evaluating the inference capabilities of language
models. arXiv preprint arXiv:2005.00782, 2020a.

Xuhui Zhou, Yue Zhang, Leyang Cui, and Dandan Huang. Evaluating commonsense in
pre-trained language models. In AAAI Conference on Artificial Intelligence, 2020b.

18
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio
Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual ex-
planations by watching movies and reading books. In IEEE international conference on
computer vision, 2015.

Appendix

This appendix describes the basic transformer-based models (Appendix A), the datasets
mentioned above that are particularly challenging (Appendix B), and the performances of
the models (Appendix C).

Appendix A. Models
A.1 The Transformer Model
The transformer model [Vaswani et al., 2017] is a neural network architecture that is based
entirely on the attention mechanism. Thereby, it eliminates the need for recurrent computa-
tion used by LSTMs [Hochreiter and Schmidhuber, 1997] and GRUs [Cho et al., 2014]. Also,
it easily learns long-range dependencies, and it allows the computation to be performed in
parallel. The transformer achieved state of the art results in machine translation.

A.2 BERT
BERT [Devlin et al., 2019] is a pre-trained language mode that consists of a stack of trans-
former blocks. BERT was pre-trained on two large corpora: The Books Corpus [Zhu et al.,
2015] (800M words) and Wikipedia (2500M words). BERT was pre-trained on two tasks:
Masked Language Modeling (MLM) and Next Sentence Prediction (NSP).
The task of MLM consists of training the model to predict a masked word given the other
words in a sentence. The dataset is constructed by choosing 15% of its tokens to be masked,
and by replacing 80% of them with the [MASK] token, 10% with a random token, and 10%
with the original token. The BERT model is trained to predict the masked word based on
the context of the sentences. The task of NSP consists of training the model to learn the
relationship between two sentences by taking as input two sentences and predicting if one
sentence follows the other.
The BERT-base model consists of 12 layers of transformer blocks with a hidden size of
768. It has 110M parameters. The BERT-large model consists of 24 layers of transformer
blocks with a hidden size of 1024, and has 340M parameters.

A.3 RoBERTa
RoBERTa [Liu et al., 2019] is an improved BERT model, which achieves better results than
BERT on different NLP tasks. The model was pre-trained longer and on a larger dataset
than BERT, by including three more datasets, namely the CommonCrawl News dataset of
63 million articles, the Web text corpus, and the Stories Corpus from Common Crawl. The
authors pre-trained the model on longer sequences, removed the NSP task, and introduced
dynamic masking (a masking technique to dynamically change the masked tokens after each
training epoch). Both variants of RoBERTa, RoBERTa-base and RoBERTa-large, have an

19
 Helwe, Clavel, & Suchanek

architecture that is similar to the one of BERT-base and BERT-large, respectively, but use
more parameters.

A.4 BART
BART [Lewis et al., 2019] is a denoising autoencoder for pre-training sequence-to-sequence
models. The model is composed of an encoder and a decoder. The encoder is a bidirectional
encoder such as BERT, and the decoder is GPT, an autoregressive decoder. Different
pre-training objectives were tested, such as token masking, token infilling, and sentence
permutations. The effectiveness of such pre-training objectives depends on the end tasks.
The BART-base model consists of 6 encoders and 6 decoders. It has 140M parameters. The
BART-large model consists of 12 encoders and 12 decoders, and has 400M parameters.

A.5 GPT-N Models

GPT is a neural language model that is pre-trained to predict the next word given all the
previous words. The model consists of a stack of transformer decoders. GPT exists in
different versions: GPT-1 was the original model. GPT-2 is a GPT model with 1.5 billion
parameters, while GPT-3 can have up to 175 billion parameters.

A.6 T5
T5 is a text-to-text transfer transformer [Raffel et al., 2020b]. It uses a unified architecture
that can be trained on a variety of NLP problems. Each problem is formulated as a text-to-
text approach. It consists of an encoder-decoder architecture that is similar to the BERT
model. However, T5 uses a causal self-attention and a fill-in-the-blank denoising pre-training
objective. There are different T5 models with different sizes: The smallest version of T5
consists of 12 layers of transformer blocks with a hidden size of 512. It has 60M parameters.
The largest T5 model consists of 24 layers of transformer blocks with a hidden size of 1024.
It has 11B parameters.

Appendix B. Datasets
B.1 ParaRules
ParaRules [Clark et al., 2020] is a dataset that serves to evaluate deductive reasoning ca-
pabilities in language models. It consists of 40K synthetic questions. These instances were
generated for 2K paraphrased facts, which were acquired by crowdworkers. Here is an ex-
ample:
Context: Harry can do magic.
Muggles cannot do magic.
If a person can do magic then they can vanish.
Mr Dursley is a Muggle.
Question: Can Harry vanish ?
Expected answer: True

20
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

B.2 ProtoQA

ProtoQA [Boratko et al., 2020] is a question-answer dataset that is designed to evaluate

commonsense reasoning capabilities in prototypical situations. A prototypical situation is
represented as a question that can have multiple common answers. Here is an example with
its possible answers:
Question: Name a profession where you might be fired if you lost your voice
Expected answers: Radio host, Teacher

The dataset is split into 9762 questions for training, 52 for validation, and 102 for testing.

B.3 COM2SENSE

The COM2SENSE dataset [Singh et al., 2021] was designed to evaluate the common-
sense reasoning capabilities in language models. The dataset includes 4K natural language
true/false statements, with each sample paired with its complementary counterpart. The
task consists of asking a model to judge whether a given sentence is logically coherent or
not:
Context: Expecting ten fish in the net, Sammy was
thrilled to see forty fish swimming in there.
Expected answer: Coherent
The authors created a counterpart to this question by modifying a few words:
Context: Expecting ten fish in the net, Sammy was
thrilled to see five fish swimming in there.
Expected answer: Incoherent

B.4 CODAH

The CODAH dataset [Chen et al., 2019] was designed to target the weaknesses of the state-
of-the-art language models. The dataset was adversarially-constructed by allowing crowd
workers to receive feedback from a pre-trained model and use this information to create
challenging commonsense questions. The dataset consists of 2801 questions. The following
is an example from the dataset:
Context: A man on his first date wanted to break the ice. He
Choices: (A) drank all of his water.
(B) threw the ice at the wall.
(C) looked at the menu.
(D) made a corny joke.

21
 Helwe, Clavel, & Suchanek

B.5 CATS
The CATs dataset [Zhou et al., 2020b] reframes 6 different commonsense reasoning bench-
marks to evaluate pre-trained transformer-based models on word-level and sentence-level
tasks. These 6 different benchmarks are Sense Making [Wang et al., 2019], the Winograd
Schema Challenge [Levesque et al., 2012], SWAG [Zellers et al., 2018], HellaSwag [Zellers
et al., 2019], Sense Making with Reasoning [Wang et al., 2019], and the Argument Rea-
soning Comprehension Task [Habernal et al., 2018]. Also, they created a new task called
Conjunction Acceptability to evaluate logical commonsense-knowledge in language models.
Here is an example from CATs:
Choices: (A) Money can be used for buying cars.
(B) Money can be used for buying stars.
Expected Answer: (A)
Here, the model has to differentiate between statements that make sense and statements
that don’t.

B.6 PIQA
The PIQA dataset [Bisk et al., 2020] is a benchmark to evaluate the physical commonsense
capabilities of language models. It consists of a set of questions, where each question has
two possible answers, but only one is correct. The training set has around 16000 instances,
while the validation set and the testing sets have around 2000 and 3000 examples, respec-
tively. The following is an instance of the dataset:
Context: To make a hard shelled taco,
Choices: (A) put seasoned beef, cheese, and lettuce onto the hard shell.
(B) put seasoned beef, cheese, and lettuce into the hard shell.

B.7 TIMEDIAL
TIMEDIAL [Qin et al., 2021] is a dataset to test temporal commonsense reasoning in di-
alogs. It consists of 1.1K dialogs represented as multiple-choice cloze tasks. This task
requires deep reasoning capabilities, such as performing different arithmetic operations over
temporal expressions with a need for commonsense reasoning. Here is an example:
Context: A: How long do you want the house? All summer ?
B: No, just for six weeks.
A: I’m afraid I can only rent it for two months.
B: My holiday is only, [MASK] but I think my brother
and his family would take it for the other two weeks .
Choices: (A) six decades
(B) 45 days
(C) six weeks
(D) two months

22
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

B.8 TORQUE

The TORQUE dataset [Ning et al., 2020] is a reading comprehension dataset concerning
temporal ordering. It consists of 21K questions, split into 80% for training, 5% for valida-
tion, and 15% for testing. Here is an example:
Context: Heavy snow is causing disruption to transport across the UK,
with heavy rainfall bringing flooding to the south-west of England.
Rescuers searching for a woman trapped in a landslide at her home
in Looe, Cornwall, said that had found a body.
Question: What events have already finished?
Expected answers: searching, trapped, landslide, said, found

B.9 MCTACO

The MCTACO dataset [Zhou et al., 2019] was designed to evaluate temporal commonsense
in transformer-based models. The dataset consists of 13K questions, split into 30% for
development and 70% for testing. Here is an example:
Context: Mr. Barco has refused US troops or advisers but has accepted US military aid.
Question: What happened after Mr. Barco accepted the military aid?
Choices: (A) The aid was denied
(B) He received the aid
(C) Things started to progress

In the above example, two answers are correct to the same questions.

B.10 TRACIE

TRACIE [Zhou et al., 2021] is a temporal reasoning textual entailment dataset. It consists
of 5.5K instances, split into 20% for training and 80% for testing. Each instance has a
hypothesis that is querying either about the start time of an event or about the end time
of an event. Here is an example:

Premise: Tom needed to get braces. He was afraid of them.

The dentist assured him everything would be fine.
Tom had them on for awhile. Once removed he felt it was worth it.
Hypothesis: Tom avoids foods he can’t eat with braces
starts before the braces are removed.
Expected answer: Entailment

23
 Helwe, Clavel, & Suchanek

B.11 RICA

RICA [Zhou et al., 2020a] is a dataset of cloze questions that can be used to assess common-
sense reasoning capabilities. To build this dataset, the authors first created commonsense
axioms such as ”Larger objects can contain smaller objects” and then translated them into
commonsense statements. RICA consists of 16000 commonsense statements, split into 80%
for training, 10% for validation, and 10% for testing. The task is to guess the comparator,
which is masked in the input sentence, as here:
Context: A prindag is smaller than a flurberg, so a flurberg
is [MASK] likely to contains a prindag.
Expected answer: more

B.12 LogiQA

LogiQA [Liu et al., 2020b] is a multiple-choice machine reading comprehension dataset.

This task assesses the logical deductive ability of language models for the case where the
correct answer to a question is not explicitly included in the passage. The corpus includes
8678 paragraph-question pairs translated from the National Civil Servants Examination of
China. Each question has one correct answer from a choice of four possible answers, as
here:
Context: David knows Mr. Zhang’s friend Jack, and Jack knows David’s friend Ms. Lin.
Everyone of them who knows Jack has a master’s degree,
and everyone of them who knows Ms. Lin is from Shanghai.
Question: Who is from Shanghai and has a master’s degree?
Choices: (A) David (B) Jack (C) Mr. Zhang (D) Ms. Lin

The dataset is split into 80% for training, 10% for validation, and 10% for testing.

B.13 ReCLOR

ReCLOR [Yu et al., 2020] is a multiple-choice machine reading comprehension dataset that
tests logical reasoning. The corpus consists of questions retrieved from standardized exams
such as LSAT and GMAT. It consists of 6138 paragraph-question pairs. Here is an example:

24
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

Context: Heavy rains during Centralia’s corn planting season prevented some farmers
there from planting corn. It is now the planting season for soybeans,
another of Centralia’s principal crops, and those fields originally intended
for corn are dry enough for planting. Nonetheless, even though soybean prices
are unusually high at present, the farmers will leave most of these fields empty
rather than plant them with soybeans, since
Question: Which of the following most logically completes the passage below ?
Choices: (A) some Centralian farmers anticipate serious financial losses due
to the extremely wet spring planting season.
(B) the extensive rains have led to an increase in the price of corn.
(C) chemicals that were used to prepare the fields for corn
planting would stunt the growth of soybeans.
(D) many centralian farmers grow both corn and soybeans.

To adequately evaluate a model without allowing it to take advantage of artifacts in the

corpus, the authors split the testing set into two sets: the EASY set where the instances
are biased and the HARD set where they are not.

B.14 AR-LSAT

AR-LSAT [Zhong et al., 2021] is a machine reading comprehension dataset that can be
used to evaluate logical reasoning capabilities. The dataset was constructed by selecting
the analytical reasoning section of 90 LSAT exams from 1991 to 2016. It consists of 2046
multiple-choice questions. Here is an example:
Context: A professor must determine the order in which five of her students
— Fernando, Ginny, Hakim, Juanita, and Kevin —
will perform in an upcoming piano recital.
Each student performs one piece, and no two performances overlap.
The following constraints apply:
Ginny must perform earlier than Fernando.
Kevin must perform earlier than Hakim and Juanita.
Hakim must perform either immediately before or immediately after Fernando
Question: If Juanita performs earlier than Ginny, then which one of the following could be true?
Choices: (A) Fernando performs fourth.
(B) Ginny performs second.
(C) Hakim performs third.
(D) Juanita performs third.
(E) Kevin performs second.

B.15 QuAIL

QuAIL [Rogers et al., 2020a] is a machine reading comprehension dataset. It assesses verbal
reasoning capabilities across 4 different domains: fiction, news, blogs, and user stories. The

25
 Helwe, Clavel, & Suchanek

corpus consists of 15K questions for 800 passages. The testing dataset comprises 15% of
the questions, and different approaches were evaluated. Due to the size of the passages, we
cannot show an example here.

B.16 StrategyQA

StrategyQA [Geva et al., 2021] is a boolean QA benchmark that can be used to evaluate
a model’s reasoning capabilities. The model has to perform implicit decomposition of the
question into reasoning steps in order to answer a question correctly. Here is an example:
Question: Did Aristotle use a laptop?
Implicit Reasoning Steps: 1. When did Aristotle live?
2. When was the laptop invented?
3. Is #2 before #1?
Expected answers: No
The dataset is composed of 2780 instances, where each instance consists of a strategy
question, a decomposition into reasoning steps, and Wikipedia paragraphs that answer
each reasoning step.

B.17 ConTROL

ConTRoL [Liu et al., 2020a] is a dataset of 8325 context-hypothesis pairs to evaluate a

models’ contextual reasoning capabilities over long texts. It is a passage-level textual en-
tailment task that consists of context-hypothesis pairs. Here is an example:
Premise: Ten new television shows appeared during the month of September.
Five of the shows were sitcoms, three were hour-
long dramas, and two were news-magazine shows.
By January, only seven of these new shows were still on the air.
Five of the shows that remained were sitcoms.
Hypothesis: At least one of the shows that were cancelled was an hour-long drama.
Expected answer: Entailment

B.18 CLUTRR

CLUTRR [Sinha et al., 2019] is a benchmark dataset to evaluate the inductive reasoning
capabilities of models. The task requires a model to infer the kinship between characters
in short stories. Here is an example:
Context: Kristin and her son Justin went to visit her mother Carol
on a nice Sunday afternoon. They went out for a movie
together and had a good time.
Question: How is Carol related to Justin ?
Expected answer: Carol is the grandmother of Justin.

26
Reasoning with Transformer-based Models: Deep Learning, but Shallow Reasoning

CLUTRR is a synthetic dataset. For each experiment, 5000 instances were generated for
training and 100 for testing.

B.19 SVAMP

SVAMP [Patel et al., 2021] is a dataset that was created by varying instances of ASDiv-A
(a dataset of one-unknown arithmetics problems). It contains 1000 tasks. To solve these
tasks, a model needs a certain level of reasoning capability. It also has to be sensitive to
the question. Here is an example:
Context: Jack had 8 pens and Mary had 5 pens.
Mary gave 3 pens to Jack.
Question: How many pens does Jack have now?
Expected answer: 8 + 3 = 11

B.20 MATH

MATH [Hendrycks et al., 2021] is a dataset that consists of 12500 competition mathematics
problems. It is split into 7500 problems for training and 5000 for testing. Each instance is a
description of the problem with a question, the step-by-step solution, and the final answer.
Here is an example from the dataset:
Context: Tom has a red marble, a green marble, a blue marble,
and three identical yellow marbles.
How many different groups of two marbles can Tom choose?
Expected answer: There are two cases here:
either Tom chooses two yellow marbles (1 result),
or he chooses two marbles of different colors (( 24 )= 6 results).
The total number of distinct pairs of marbles Tom can choose is
1+6=7

B.21 IsarSTEP

IsarStep [Li et al., 2021] is a mathematical reasoning benchmark. It was built by collecting
formal proofs written in Isabelle from the Archive of Formal Proofs and from the standard
library of Isabelle/HOL. In this task a model needs to predict√ the missing intermediate
proposition in a proof. Here is an example for the proof that 2 is not a rational number,
where the missing intermediate proposition is a is even:
Context: 2b2 = a2 ⇒ [M issing P roposition] ⇒ ∃ c ∈ Z. a = 2c
Expected answer: a is even
The dataset is split into 820K examples for training, 5000 for validation, and 5000 for
testing.

27
 Helwe, Clavel, & Suchanek

B.22 HOList
HOList [Bansal et al., 2019] is an automated theorem proving dataset for higher-order logic.
The benchmark includes 29465 theorems and their proofs, split into 60% for training, 20%
for validation, and 20% for testing. Two tasks can be evaluated in HOList: (1) proving
each theorem in the dataset, and (2) predicting the tactic and the arguments of the tactic
that were used in the human proof. A tactic can be a previously proven theorem or a list
of previously proven theorems.

B.23 MetaMathStep
MetaMathStep [Polu and Sutskever, 2020] is a benchmark for automated theorem proving.
The dataset evaluates the capabilities of a language model to generate a proof for a given
statement. The dataset contains 3 million proof steps for around 38000 theorems, which
are split into 36K for training, 1K for validation, and 1K for testing.

Appendix C. Model Performances on Selected Datasets

Type of Reasoning Dataset Model Performance Metric

ParaRules RoBERTa 98.8% Accuracy
Horn Reasoning (Section: 3.1) ParaRules PROVER (Based on RoBERTa) 98.4% Accuracy
ParaRules PRoofWriter (Based on T5) 99.1% Accuracy
Commonsense Reasoning (Section: 3.2) ProtoQA GPT-2 71.2% Accuracy
CODAH BERT 69.5% Accuracy
CATS RoBERTa 67% Accuracy
PIQA RoBERTa 77% Accuracy
MCTACO BERT 69.9% F1-Score
Event-based Commonsense Reasoning (Section: 3.3) TRACIE SymTime (Based on T5) 80.6% F1-Score
TORQUE RoBERTa 75.2% F1-Score
LogiQA RoBERTa 35.31% Accuracy
LogiQA DAGN 39.32% Accuracy
LogiQA FOCAL REASONER (Based on RoBERTa) 40.25% Accuracy
ReClor EASY RoBERTa 75.50% Accuracy
ReClor HARD RoBERTa 54.3% Accuracy
ReCLor EASY DAGN (Based on RoBERTa) 75.91% Accuracy
ReClor HARD DAGN (Based on RoBERTa) 44.46% Accuracy
ReCLor EASY FOCAL REASONER (Based on RoBERTa) 77.05% Accuracy
ReClor HARD FOCAL REASONER (Based onRoBERTa) 44.64% Accuracy
Implicit Reasoning (Section: 3.4)
ReClor EASY RoBERTa with Logical Data Augmentation 81.4% Accuracy
ReClor HARD RoBERTa with Logical Data Augmentation 62.5% Accuracy
AR-LSAT RoBERTa 23.1% Accuracy
QuAIL BERT 55.9% Accuracy
StrategyQA RoBERTa 66% Accuracy
ConTRoL BART 60.95% Accuracy
CLUTTR GAT 77% Accuracy
RICA GPT-2 50% Accuracy
RICA RoBERTa 50% Accuracy
Mathematical Reasoning (Section: 3.5) SVAMP RoBERTa Embeddings + Graph2Tree 65% Accuracy
MATH GPT-2 6.9% Accuracy
IsarStep Hierarchical Transformer 22.8% Accuracy

Table 1: Model Performances on Selected Datasets

28