COLD: Causal reasOning in cLosed Daily activities

Abhinav Joshi∗ Areeb Ahmad∗ Ashutosh Modi

Department of Computer Science and Engineering
Indian Institute of Technology Kanpur (IIT Kanpur)
Kanpur, India
{ajoshi,areeb,ashutoshm}@cse.iitk.ac.in

Abstract
Large Language Models (LLMs) have shown state-of-the-art performance in a vari-
ety of tasks, including arithmetic and reasoning; however, to gauge the intellectual
capabilities of LLMs, causal reasoning has become a reliable proxy for validating
a general understanding of the mechanics and intricacies of the world similar to hu-
mans. Previous works in natural language processing (NLP) have either focused on
open-ended causal reasoning via causal commonsense reasoning (CCR) or framed
a symbolic representation-based question answering for theoretically backed-up
analysis via a causal inference engine. The former adds an advantage of real-world
grounding but lacks theoretically backed-up analysis/validation, whereas the latter
is far from real-world grounding. In this work, we bridge this gap by proposing the
COLD (Causal reasOning in cLosed Daily activities) framework, which is built
upon human understanding of daily real-world activities to reason about the causal
nature of events. We show that the proposed framework facilitates the creation
of enormous causal queries (∼ 9 million) and comes close to the mini-turing test,
simulating causal reasoning to evaluate the understanding of a daily real-world
task. We evaluate multiple LLMs on the created causal queries and find that causal
reasoning is challenging even for activities trivial to humans. We further explore
(the causal reasoning abilities of LLMs) using the backdoor criterion to determine
the causal strength between events.

1 Introduction
In recent times, Large Language Models (LLMs) have shown remarkable generalization capabilities
[Devlin et al., 2019, Radford et al., 2019, Brown et al., 2020]. Consequently, the ability to perform
causal reasoning (often considered a core feature of intelligence [Penn and Povinelli, 2007, Pearl
and Mackenzie, 2018]) has sparked research interest in the context of LLMs, aiming to answer if
causal reasoning is possible with LLMs [Weber et al., 2020, Jin et al., 2023, 2024, Cohrs et al., 2023,
Romanou et al., 2023, Yang et al., 2023, Mitchell et al., 2023, Vashishtha et al., 2023, Stolfo et al.,
2023]. On a broader level, there are two lines of work; first, that treats the causal reasoning via
learning relationships between the events that are grounded in the real world [Gordon et al., 2012,
Ho et al., 2022, Zečević et al., 2023, Zhang et al., 2023, Wang et al., 2023]. Second line of work
relies on a causal inference engine and establishes relationships between variables via symbolic
representation [Jin et al., 2023, 2024]. The former relies on understanding real-world events but lacks
formal definitions that adhere to the causal inference theory. The latter solves the issue using a causal
inference engine but uses symbolic representations not grounded in the world, making the causal
queries more like a test for the understanding of causal theory. Though the first line of work includes
real-world events, the causal queries are often limited and could be answered by memorizing the
causal relationships between the events. Recent findings that include rigorous analysis using a causal
∗
Equal Contribution

38th Conference on Neural Information Processing Systems (NeurIPS 2024).

inference engine claim LLMs to be “Causal Parrots” [Zečević et al., 2023], i.e., the LLMs tend to
pick up (memorize) patterns in the training data to perform well on the causal reasoning benchmarks.
Moreover, some initial findings by Tang et al. [2023] suggest that LLMs perform significantly better
when semantics are consistent with commonsense but struggle to solve symbolic tasks, pointing
towards semantic representation to be better for proper validation of LLMs, leading to a conclusion
that an in-depth analysis using real-world events is necessary.
In this work, we bridge the gap between the two approaches by proposing the COLD (Causal
reasOning in cLosed Daily activities) framework, based on the human understanding of real-
world daily activities capturing commonsense (for example, “making coffee,” “boarding an airplane,”
etc.), that adheres to causal theory literature. It is more natural to frame real-life reasoning-based
queries via language; consequently, we follow the literature on Causal Commonsense Reasoning
(CCR), which studies the relationships between real-world events (described via natural language).
CCR is a non-trivial task of estimating the cause-
and-effect relationship between events that are Intention to perform a task

studied under the umbrella of commonsense

reasoning [Kuipers, 1984, Gordon et al., 2012,
Go to
Airport
U
(traveling by an airplane,
planting a tree, going
grocery shopping, etc. )

Zhang et al., 2022b, Wang et al., 2023, Chun

Check-
et al., 2023, Du et al., 2022]. The events in CCR in
Luggage
Search
generally refer to actions taking place in an ac- for the
boardin

tivity in the real world. For example, consider Temporal Link

g gate Wait
for
luggage
the activity of “traveling by an airplane” Causal Link
. Take
back
luggage
given in Fig. 1, where the occurrence of all the Cause Event
Leave
Airport

events is confounded by a universal variable U Effect Event(s)

(“intention to perform a task”). Moreover, there

are a few events that cause one another. For ex- Figure 1: U denotes the unobserved variables, con-
ample, the event “checking in luggage” (E1 ) founding all events present in a real-world activ-
caused the occurrence of events like “waiting ity. In an activity, some events cause other events
at the luggage belt” (E2 ) after the flight, to happen. For example, in “traveling by an air-
i.e., in an alternate universe where one does not plane”, the event of “check-in luggage” causes
checks in luggage and goes with the cabin bags, events like “taking back luggage.”
will never wait for their luggage after the flight
has landed. Moreover, some of the events have no causal impact, like “find the boarding gate”
(E3 ) has no causal relationship with “checking in luggage” (E1 ). More formally,
∆(E1 →E2 ) = P(E2 |do(E1 )) − P(E2 |do(¬E1 ))
(1)
∆(E1 →E3 ) = P(E3 |do(E1 )) − P(E3 |do(¬E1 ))
where do(.) denotes the do operator [Pearl, 2012] showing the intervention on E1 , and ∆ is the
causal estimand capturing the causal strength between two events, i.e., ∆(E1 → E2 ) is expected
to be higher when compared to ∆(E1 → E3 ). Note, CCR excludes the causal events that are
beyond the reach of commonsense knowledge, for example, does “planting trees” have a direct
impact on the “rainy season”? or does “providing free education” improve the “economic
condition of the country/state”; does “carpooling” directly impact “air pollution”, etc.
A noteworthy point concerning causality is that though the logical temporal (or prototypical) order of
these events provides a weak signal about causal relationships, temporal precedence does not always
imply causation (§2). For example, one could erroneously argue that “boarding a plane” is also
the cause of “waiting at the luggage belt” since without “boarding a plane,” one cannot
wait for the luggage belt.
For building a causal reasoning framework (based on CCR) around real-life daily activities, one
would require a few primary features readily available: 1) Clear distinction between the events,
i.e., the events should encapsulate describing a particular step in an activity, 2) Causal Dependency
between the variables/events, i.e., there should be some events causing other events to occur. 3)
Causal independence of events with the rest of the world, i.e., the occurrence of events should be
independent of events that are not part of the activity (i.e., the covariates are balanced out). We found
that “Scripts” [Schank, 1975, Schank and Abelson, 1975] provide a concrete medium that satisfies all
these requirements. Scripts are defined as a sequence of events describing a prototypical activity, such
as going to a restaurant, and hence capture commonsense knowledge about the world. [Schank and
Abelson, 1975, Modi et al., 2016, Wanzare et al., 2016, Ostermann et al., 2018, Modi, 2016, 2017,
Modi et al., 2017, Modi and Titov, 2014]. Moreover, different people have similar understandings of

2
 Causal Graph for the Activity “Going Grocery Shopping”

Pre activity
The pre-world and post-world activities involve various

world
events that take place outside the context of daily
get in car
activities, making them independent of the events
happening during the activity. turn-on car go by walk go by bus
make list
drive to grocery
carry list
Steps performed in daily life activities
(capturing commonsense knowledge)
park car

Going Grocery Shopping

Flying in an airplane get groceries

Going on a Train
enter store

Riding on a Bus
get list from car
Taking a Bath
bring vehicle
from parking
go to checkout
pack groceries
present coupons
check list put conveyor
(if anything is left) get wallet cashier
scan/weight
take shop cart
take bags
pay
use bags
Post activity

return cart to corral get receipt

put bags in cart
world

Begin Start Start

again work Colliders denoting Conditional Dependence
another
relax activity Independent Event (may or may not be present in an activity
and is not a cause or effect of any of the other events)

Figure 2: Left: the figure represents the closed nature of daily real-world activities (capturing
commonsense, commonly understood by humans), start and end given the context of the task, i.e.,
the pre-activity world and post-activity world activities marginalize out the dependence of event
occurring during the activity with the rest of the world. Right: Causal Graph for “going grocery
shopping.” Notice the collider (red nodes) makes the independent set of nodes (highlighted in
different colors) unconditionally independent in the causal graph. In contrast, when given a condition
on a collider (“put bags in cart”, the two clusters (yellow and blue) become dependent (if collider
is observed, both yellow and blue clusters may have been observed as well).

the activity in the form of scripts that inherently balance out the covariates present in the real world,
i.e., all the activities have the same starting and ending point and account for common exogenous
and endogenous variables, providing a suitable platform to establish a cause-and-effect relationship
between the events. In other words, for an activity like “flying in an airplane,” or “going
grocery shopping” (also see Fig. 2, left) the events that happened before starting the activity and
after completing the activity are marginalized out using a common understanding of these activities by
different humans and hence will have no causal relations with any of the exogenous events during the
activity. Creating a causal graph for script knowledge, i.e., establishing relationships between events
taking place during the activity, provides a perfect platform for creating causal queries, thus providing
a medium to establish CCR between events. In a nutshell, we make the following contributions:
• We propose COLD (Causal reasOning in cLosed Daily activities), a CCR framework based on
Script knowledge (daily activities involving commonsense) that provides a closed system to test the
understanding of causal inference grounded in the real-world. The proposed framework adheres
to SUTVA (Stable Unit Treatment Value Assumption) [Cox, 1958, Rubin, 1980] by design (§3).
COLD consists of activity-specific observational graphs (created via crowd-sourcing) and causal
graphs. Further, COLD facilitates creating an enormous number of causal queries (e.g., 2, 887, 950
per activity) via causal query triplets from the causal graph. This comes close to the mini-Turing
test [Pearl and Mackenzie, 2018], where the story becomes the understanding of the daily activity,
and the sampled enormous causal queries help in the exhaustive and rigorous evaluation of LMs.
• We devise various design mechanisms for estimating causal strength analytically and show how the
representations learned by language models can be validated.
• Via detailed experimentation on the widely used open-weight language models, including encoder-
only models (RoBERTa-MNLI) and autoregressive models (gpt-neo-125M, gpt-neo-1.3B, gemma-
2b, gpt-neo-2.7B, phi-2, gpt-j-6B, Llama-2-7b-chat-hf, Mistral-7B-v0.1, gemma-7b, and Meta-
Llama-3-8B) we estimate the causal reasoning capability of the learned representations. We release
the framework, model code, and results via https://github.com/Exploration-Lab/COLD.

3
 Observational “Causal Query Triplets”
Crowdsource Human Written ESDs Algorithm-1 for creating
Graph ( )
Workers “Causal Query Triplets”
1. Ge t a cake mix Premise: “go to store and buy cake mix”
1. Prehe at the Oven
2. Mix in the extra ingredients
2. Prepare Dry Ingredie nts (floor, Question: Which of the following is an effect?
3. Prepare the cake pan. baking powder, salt) in a bowl
4. Prehe at the oven 3. Cream Butte r and Sugar
5. Put the mix in the pans 4. Add Eggs and Vanilla
Choice 1: “come home with the ingredients”
6. Put the cake batte r in the5. oven
Combine w et and dry Choice 2: “go to kitchen”
7. Ta ke it out of the oven Ingre dients
1. Purc hase cake m ix the c ake
6. Bake
2. Prehe at oven7. Frost the ca ke
3. Grease pan 8. De corate the cake
4. add ingredients Premise: “turn off oven”
5. Mix Question: Which of the following is a cause?
6. Pour into prepared pan
7. Bake cake for re quired time
8. Remove cake f rom oven Choice 1: “go to store, buy ingredients”
“Baking a Cake” 9. Turn cake out onto cake plate Causal Graph Choice 2: “first heat oven”
10. Apply ic ing or gla ze
( )

Real-World Commonsense Validating Causal Reasoning

Activities Knowledge Abilities
GPT-neo Phi
GPT-J Llama
Gemma Mistral

Figure 3: The proposed COLD framework for evaluating LLMs for causal reasoning. The human-
written Event Sequence Descriptions (ESDs) are obtained from crowdsource workers and include a
telegrammic-style sequence of events when performing an activity. The Observational Graph and the
Causal Graph for an activity are used to create causal query triplets (details in Algorithm 1), shown
towards the right. Using counterfactual reasoning, “going to the kitchen” is possible without going to
the market (if the ingredients are already available), making “come home with the ingredients.” a
more plausible effect among the given choices. Similarly, in the second example, the event “going to
market” has no direct relation with the event “heating the oven”.
2 Background

The Mini Turing Test proposed by Pearl and Mackenzie [2018] is designed in a question-answering
format to validate the understanding of causal knowledge about a simple story. The primary feature of
a mini-Turing test is the enormous number of causal queries that can be framed using the underlying
causal graph, which governs the occurrence of events in the story. Due to the enormous number of
causal queries, passing the mini-Turing via memorization becomes combinatorially heavy, and hence,
the authors argue that it can only be beaten if one has access to the underlying causal graph governing
the occurrence of events (i.e., one has the ability to reason causally about the events). In this work,
though, we only consider a more straightforward case of choice-based causal triplets; we realize the
number of causal queries that can be created is enormous and helps validate the causal reasoning
abilities coming close to the mini-Turing test.
d-separation: Establishing the independence of variables becomes non-trivial when dealing with
complex interactions among multiple variables. d-separation [Pearl, 1988] facilitates the determina-
tion of conditional independence between two sets of nodes X and Y in a graphical model G given
another set of nodes Z. d-separation asserts that X and Y, given the set Z, are d-separated if all paths
for every node in X and every node in Y are blocked by conditioning on Z, denoted as X G Y | Z.
|=

A path p is blocked by a set of nodes Z [Pearl et al., 2016], if and only if: 1) p contains a chain of
nodes A → B → C or a fork A ← B → C such that the middle node B is in Z OR 2) p contains a
collider A → B ← C such that the collision node B or its descendant is not in Z.
Backdoor Criterion: A set of variables W satisfies the backdoor criterion relative to T and Y if the
following are true:
(A) W blocks all backdoor paths from T to Y i.e. blocking confounding or non-causation
association paths
(B) W doesn’t contain any descendants of T
Then, W satisfies the backdoor criterion [Pearl et al., 2016, Neal, 2020]. We make use of the backdoor
criterion to estimate the causal estimand, capturing the relationship between the causal events. (Refer
App. C for more detail)

3 COLD (Causal reasOning in cLosed Daily activities)

We propose COLD (Causal reasOning in cLosed Daily activities) framework for testing causal
reasoning abilities of natural language understanding systems such as LLMs. Fig. 3 gives an overview
of the creation process. We use crowd-sourced data of script knowledge to create observational

4
graphs which is further used along with manual intervention to create causal graphs. Subsequently, an
algorithm is used to create an enormous number of causal queries (causal triplets), which are further
used to test LLMs for causal reasoning. Next, we explain each of the steps in more detail.
Task Formulation: COLD is motivated by Causal Commonsense Reasoning (CCR), which we
define as the task of finding the strength of the cause-and-effect relationship between two events (E1
and E2 ) given in an activity a ∈ A, where A is the set of all activities. For example, for an activity
like “going in an airplane”, the central question is to determine the causal relationship between
two events that occur during the activity (events like “checking in luggage” and “waiting for
luggage”). Since reasoning about a sequence of events is tedious (and sometimes confusing [Do
et al., 2011]), researchers often rely on a more plausible cause rather than defining a definite causal
event. For instance, COPA dataset [Gordon et al., 2012] provides a premise event and a corresponding
causal query question along with two choices (see Table 1 for an example); a system is required to
predict which of the two choices is most plausible cause/effect as required by the question.
Creating a Closed Causal System
Given the nature of Script knowledge (satisfying the criterion of balanced covariates, §1), we use a
script corpus called DeScript [Wanzare et al., 2016] for creating the observational graphs. DeScript is
a corpus with a telegram-style sequential description of an activity in English (e.g., baking a cake,
taking a bath, etc.). DeScript is created via crowd-sourcing. For a given activity, crowd-workers write
a point-wise and sequential short description of various events involved in executing the activity (this
one complete description is called an ESD (Event Sequence Description)). DeScript collects data
for a set of 40 daily activities (100 ESDs each) varying in complexity and background knowledge.
Additionally, for a given activity, semantically similar events from different ESDs are manually
aligned by human annotators (for more details, refer to Wanzare et al. [2016]). These alignments
were later used by Joshi et al. [2023b,a] to create a DAG representing the overall activity. In our work,
(a)
we use these DAGs as the observational distribution of an activity (Go , where a ∈ A, where A is the
set of all activities). These DAGs provide a medium for generating enormous trajectories (scales from
1.6e + 16 to 1.3e + 27, also see Table 2), that are coming directly from human annotations (alignment
as well as the ESDs), providing us a proxy to represent the understanding of daily activities.
Observational Distribution (Go ): Note that the graphs Go , approximately represent (almost) all
possible ways in which an ESD can be written for an activity, providing the true observational
distribution, i.e., how the combinations of events will look like while performing the activity in the
real world (see App. A.3 for examples).
Causal Graphs (Gc ): To reason about the causal relationships between the events (nodes of Go ),
we would need the underlying causal graph that shows the cause of occurrence of various activities
(directly or indirectly). We construct the causal graphs manually by reasoning about the independence
of various events in the activity. Fig. 2 shows the pictorial representation of one of the created causal
graphs for the activity “going grocery shopping.” Notice that various sets of events in the graph
create clusters, denoting independence between various events. For example, nodes related to make
list cause the events that involve the presence of a list and do not cause events like going via car
(as some of the population will not create a list for shopping). Similarly, the mode of transportation
(car/bus/walk) is independent of the events performed inside the store.
Causal Query Triplets: The obtained Causal Graph (Gc ), for an activity provides a medium to
reason about causal links between the events. Notice in Fig. 2, how the red nodes (colliders) help
separate out the independent event clusters. For example, the nodes ‘get list from car’ and
‘check list (if anything is left)’ being colliders, separate out the making list-related events
with the rest of the graph. Similarly, the node ‘put bags in cart’ separates out the ‘take shop
cart’ and ‘take bags’. Another interesting property represented in the obtained causal graph is the
conditional dependence between various clusters. For example, the cluster related to ‘get in car’ is
unconditionally independent of ‘make list’. However, if we condition on the collider (‘get list
from car’), they become dependent, i.e., if ‘get list from car’ is observed, it means that the
person will have created the list as well as went by car for the grocery shopping (similarly for node
‘put bags in cart’). d-separation (§2) provides an easier way to establish conditional/unconditional
independence between the set of nodes. For creating the dataset of causal queries (similar to other
datasets like COPA [Gordon et al., 2012]), we need a triplet of three events (premise p, Choice-1
c1 , Choice-2 c2 ) associated with a question about ‘cause’ or ‘effect’ relationship, i.e., given the
premise which of the two choices is the cause/effect? (Table 1). We call these triplets Causal Query

5
Table 1: The table shows examples of causal query triplets created using the causal graphs (Gc ) and
observational graphs (Go ) in Algorithm 1. The top row is taken from the COPA dataset [Gordon et al.,
2012] for the purpose of comparison. Note the examples in the table show the samples taken from
the instance version.
Dataset Premise Choice-1 (1) Choice-2 (2) question answer
The man turned on the faucet. The toilet filled with water. Water flowed from the spout. effect 2
C OPA

The girl found a bug in her cereal. She poured milk in the bowl. She lost her appetite. effect 2
The hamburger meat browned. The cook froze it. The cook grilled it. cause 2
buy proper ingredients. go home with ingredients. wait for the timer to go off. effect 1
measure ingredients in designated measuring cups. whisk after each addition.stir to combine. clean up the mess. effect 1
COLD
(Cake)

bake until cake is ready. set timer. carefully remove cake from pan. cause 1
turn off oven. go to store, buy ingredients. first heat oven. cause 2
preheat oven to 350 degrees. turn off oven. prepare the microwave oven and utensils effect 1
pay total. get receipt. place cart into cart corral. effect 1
(Shopping)

get the bill for groceries. pay for the grocery. return cart to store. cause 1
COLD

pay for it . start at the non-cold side of the store. go to shelf and get the food items. cause 2
go back to the car.put the bags in the car. take your car and drive to grocery shop. bring items to checkout. cause 1
take the full cart to the checkout lane. watch prices as the checker scans. go down aisles. effect 1
check train schedules. choose a destination. go to the car. cause 1
(Train)

you board your train and find your seat. find your seat or compartment. wait for train. effect 1
COLD

get off train. go out of the station. take all the luggage out of train. effect 1
go out of the station. put carry on luggage in overhead bin. get off at your correct stop. cause 2
arrive at destination. when train reaches destination, exit train. walk to the train platform. effect 1
go to garden center. transport it home. choose type of tree. effect 1
fill hole with dirt and fertilizer. get tree. dig hole big enough for tree to grow. cause 2
COLD
(Tree)

place the tree at the top of the hole. cover the roots with dirt. get a tree. effect 1
fill in dirt around the tree gently. take it home. place the tree in the hole. cause 2
place tree sapling into hole. pack dirt back in. find place for tree. effect 1
pull signal for stop. bus stops at destination. stand up and go to door. effect 1
board the bus when it arrives. while boarding, pay the driver the required fee pull signal for stop. effect 1
COLD
(Bus)

step on bus. take available seat. wait for the bus to arrive. effect 1
buy bus ticket. when it arrives get on. when your stop approaches, pull cord. cause 1
find seat on bus and sit. find out what bus to take. the bus arrives at the departure station. cause 2

Table 2: The table provides details of the observational graph (Go ) for 5 activities. The Causal Query
Triplets represent the total number of triplets generated via Algorithm 1. The instance version shows
the number of samples present in the instance version (including different text instances describing
the same event) of the created dataset. Table 1 shows a small sample taken for the 5 activities. Overall,
the huge number of samples highlights the exhaustive nature of evaluation that can be done for LLMs.
Activity Nodes Compact Trajectories Total Trajectories Causal Query Triplets Instance version (Num. samples)
Baking a Cake 28 177030 1.3e + 27 864 2887950
Riding on a Bus 20 13945 1.3e + 17 334 834046
Going Grocery Shopping 33 626096 3.1e + 26 1984 3739184
Going on a Train 26 133799 4.9e + 22 950 1213114
Planting a Tree 23 4466 1.6e + 16 260 846046
Total Dataset Samples - - - 4392 9,520,340

Triplets that are used to frame a causal query between the events. App. B, Algorithm 1 presents the
mechanism to create a dataset of causal query triplets. We start by constructing the set of possible
triplets, and sort the nodes in every triplet using the topological order preset in the observational
graph (a DAG). Further, using d-separation, we figure out the triplets that have one node d-separated
from the other nodes. The d-separated node becomes the wrong choice. The premise and correct
choice are determined from the remaining choices, leading to a ‘cause’ and ‘effect’ query based on
the topological order, i.e., the event occurring before (temporally) in an activity becomes ‘cause.’ The
other event becomes the plausible ‘effect.’ Note that the temporal precedence is generally assumed
essential for defining causation, and it is one of the most important clues that is used to distinguish
causal from other types of associations [Mill, 1898, Hill, 1965, Pearl and Verma, 1995]. For our
framework, we also consider the topologically sorted order obtained from the observational graphs
and use the temporal order to define the causal query triplets, i.e., the cause events will always precede
the effect events (see App. B, Algorithm 1, where Go helps determine the temporal order of events).
Table 1 shows a sample of the created dataset in comparison to the COPA dataset. Note that each node
(premise, Choice-1, Choice-2) in Gc also has multiple texts (instances of text describing the same
event) written by different crowd-sourced workers. We can further use these text instances to enrich
the created dataset by considering all the available instances to create all possible combinations. This
strategy increases the scale of the created dataset by a huge margin. Overall, we found a dataset
created from triplets using Algorithm 1 results in 4,392 tuples, which, after using the text instances,
increases to 9,520,340, bringing it close to ‘mini turning test” [Pearl and Mackenzie, 2018].
Adherence to SUTVA: In causal literature, a fundamentally acknowledged Stable Unit Treatment
Value Assumption (SUTVA) [Cox, 1958, Rubin, 1980]) requires that for each unit (e.g., sequence of
events), there is only one version of the non-treatment, i.e., for an event in the sequence, there lie only

6
 Table 3: The table provides evaluation results of language models over the created causal triplets.
Triplets Model Name cake shopping train tree bus
gpt-neo-125M 50.71 50.01 49.99 50.13 50.15
gpt-neo-1.3B 44.77 45.69 42.52 45.67 42.89
gemma-2b 53.76 52.19 60.57 60.71 53.64

causal triplets
gpt-neo-2.7B 50.00 50.01 50.00 50.01 50.00
phi-2 85.14 83.65 77.29 82.24 71.74
gpt-j-6B 49.59 50.02 50.29 49.92 49.93
Llama-2-7b-chat-hf 77.92 72.41 73.48 72.40 68.21
Mistral-7B-v0.1 77.64 69.38 68.46 72.43 69.37
gemma-7b 81.47 82.26 77.24 80.78 70.29
Meta-Llama-3-8B 80.79 76.46 76.08 78.21 67.39

two versions occurring and not occurring. SUTVA plays a vital role in causal inference by ensuring
that each unit’s treatment assignment has a consistent impact, facilitating the accurate estimation of
treatment effects. Our framework closely adheres to SUTVA assumptions (details in A.1).
Comparison with Other Causal Datasets: We briefly compare the created dataset with the existing
set of causal reasoning datasets in App. Table 5. The created dataset serves as a middle ground,
having both real-world groundings as well as an underlying causal graph to create an exhaustive set
of causal queries.

4 Experiments and Results

COLD provides a causal query dataset for evaluating LMs for causal understanding. In particular,
we consider the “Causal Query Triplets" (Table 2) coming from compact trajectories as a base and
sample the instance version coming from the same skeleton. Since it is not possible to evaluate
all the possible causal queries that could be created using our framework, we use 10K samples
for each activity to report our findings. For a fair comparison between various models and better
reproducibility, we freeze the sampled causal query triplets and compare the success rate over the
frozen samples. We evaluate via two methods. First, as done in previous work Jin et al. [2024, 2023],
Chen et al. [2024], we first experiment with various LLMs using a prompt-based evaluation scheme;
second, we propose other mechanisms (based on causal theory, e.g., Average Treatment Effect) that
could be used to perform an in-depth analysis of evaluating causal relationships between events.
Causal Reasoning Evaluation of LLMs via Prompts: We start with the prompt-based evaluation of
recent open-weight LLMs (gpt-neo-125M, gpt-neo-1.3B, gpt-neo-2.7B [Black et al., 2021], gemma-
2b [Team et al., 2024], phi-2 [Javaheripi et al., 2023], gpt-j-6B [Wang and Komatsuzaki, 2021],
gemma-7b [Team et al., 2024], Llama-2-7b-chat-hf [Touvron et al., 2023], Mistral-7B-v0.1 [Jiang
et al., 2023], and Meta-Llama-3-8B [Dubey et al., 2024]) We frame the prompt as a multi-choice
question-answering (MCQA) objective [Robinson and Wingate, 2023]. The prompt is intentionally
structured so that the LLM is intended to predict a single choice token (Such as “ A”, “ B”, etc.).
Robinson and Wingate [2023] highlight the advantages of MCQA-based evaluation over cloze
evaluation [Brown et al., 2020](where the LLMs are expected to generate the entire answer in a cloze
test), leading to a significant boost in various tasks, including commonsense-based tasks. App. E, Fig.
5 presents various prompt templates for autoregressive experiments, and App. E Fig. 6 shows a few
qualitative examples for the framed causal query templates. Table 3 shows the success rate obtained
for various LLMs. The success rate corresponds to the percentage of queries where the LLM predicts
the desired choice. We observe that reasoning causally about simple daily activities is challenging
when a rigorous test is framed, validating the dependencies between the events. Overall, for the more
common activities like baking a cake and going grocery shopping, the LLMs perform better
when compared to activities like boarding a bus or planting a tree. We also experimented
with another version of the dataset, where incorrect choice may correspond to temporally plausible
but causally implausible events. The results drop significantly in this case; details and results are
provided in the App. F.1.
Evaluation using Average Treatment Effect (ATE) (∆): Computing the Average Treatment Effect
(∆) helps establish the strength of causal links given a context (Eq. 1). In our setup, to estimate

7
P (y|do(x)) (i.e., the causal estimand) from statistical estimands (obtained from observational dis-
tribution), we make certain reasonable assumptions about the underlying process that governs the
relation among variables/events and then utilize the implications of these assumptions. For any activ-
ity taking place, the causal relationships between two events E1 and E2 may have a causal link along
with a non-causal link through a set of confounders z. We define the confounder z = {ti |ti ∈ T },
where T denotes all the trajectories (sequence of events) from the start of the activity till the event
E1 . The temporal nature of events makes this assumption suitable since the occurrence of E1 and E2
can be confounded by all the events preceding E1 . Note the possibility of unobserved confounders
(events that are not explicitly mentioned but may be affecting the mentioned events) in our case is
removed due to two reasons: 1) Keeping a closed system representation with a large number of diverse
scripts (written by humans) helps cover the set of most generic and diverse events either implicitly
or explicitly as a part of the activity, and 2) The causal reasoning goal is restricted to figuring out
the causal effect between the events that are present explicitly. Assuming the unmentioned events
have insignificant effects, we can establish that there will not be any unobserved confounders. This
assumption makes the observed confounders satisfy the backdoor criterion [Pearl, 1993] that make
sufficient adjustment sets. By using the backdoor criterion (App. C), the interventional distributions
are estimated as follows:
X
P (E2 |do(E1 )) = p∗ (E2 |E1 , z = ti )p∗ (z = ti ) (2)
ti ∈T

Note that the true observational distribution, i.e. p∗ (E2 |E1 , z) and p∗ (z) both are unknown and
have to be approximated (p̂(E2 |E1 , z) and p̂(z)). Further, we describe ways p̂ can be estimated via
multiple design mechanisms. Due to space limitation, we only describe the p̂ estimation via language
models below and move the statistical analysis using the original trajectories and observational graphs
to the App. D.2.
ATE using Language Models Since pre-trained LMs capture world knowledge [Devlin et al., 2019,
Brown et al., 2020, Li et al., 2023, Nanda et al., 2023, Karvonen, 2024], these provide a suitable
proxy for establishing relationships between these events. For our experiments, we consider a simple
reasoning capability of language models, i.e., to reason about the temporal order of various events,
i.e., given an event, what is the likelihood of the occurrence of another event? We further ask if this
can be used to estimate the causal relationship between the events (a similar strategy is used by Zhang
et al. [2022b] for zero-shot causal estimation). It is worth noting that for these activities about daily
activities, one way to find causes is to establish the temporal likelihood of the events. For each of the
multiple language models, we frame the temporal prediction differently.
Encoder-only Models: For BERT-based models trained for mask token prediction, we model the
temporal prediction using the probability assigned to the mask tokens “before” and “after” [Zhang
et al., 2022b]. Given two events E1 and E2 , the temporal link is predicted using a prompt like E1
<mask> E2 , and the scores corresponding to the before and after tokens are collected. App. D,
Fig. 7, the top row highlights the prompt template used for BERT-based models. For encoder-only
experiments, we consider RoBERTa MNLI [Liu et al., 2019].
Decoder-only Models: For other language models that are autoregressive in nature, we modify the
prompt to predict the temporal order as the last token. We again use the MCQA-based prompting
style to frame the temporal order query by providing “before" and “after" as the options in the prompt.
App. D, Fig. 7, the bottom row highlights the prompt template used for Decoder-only Models.
Interventions: We utilize the SUTVA assumption in the proposed framework to devise an intervention
over a trajectory in natural language form. App. D, Fig. 8 shows the style of intervention made by an
event (E1 ) taking place (do(E1 )) or not taking place do(¬E1 ).
Given the above strategies, LMs can be used to evaluate p̂(E2 |¬E1 , z = t), by feeding the prompt
that contains E1 , E2 and the z = t) and predict the temporal nature between E1 and E2 , given a
trajectory z = t. Further, applying the backdoor criterion for multiple trajectories T , we obtain
1 X
pM (E2 |do(E1 )) = p̂(E2 |E1 , z = t)
|T |
t∈T
(3)
1 X
pM (E2 |do(¬E1 )) = p̂(E2 |¬E1 , z = t)
|T |
t∈T

8
Table 4: Accuracy over the causal triplets for various ∆ estimates. The blue text denotes the
improvements by backdoor adjustments over the temporal scheme for multiple language models.
Bold text represents the best-performing method for a particular activity.
p̂ estimation Scheme ATE Cake Shopping Train Tree Bus
Original Trajectories ∆o 28.20 34.30 31.10 30.10 30.40
- ∆n 30.40 30.10 29.80 28.60 25.4
Observational Graphs
- ∆t 40.90 47.10 40.30 37.60 40.10
RoBERTa MNLI 46.80 54.00 45.50 52.70 43.00
gpt-neo-125M 47.70 55.50 55.50 53.60 48.20
gpt-neo-1.3B 47.40 45.40 53.30 43.40 52.90
gemma-2b 43.80 41.70 52.20 49.70 49.80

Temporal
gpt-neo-2.7B 50.10 48.90 52.40 47.60 53.70
Language Models phi-2 60.30 59.20 56.90 70.30 49.40
gpt-j-6B 49.50 46.40 56.00 62.70 56.00
Llama-2-7b-chat-hf 38.90 42.10 51.00 40.70 47.80
Mistral-7B-v0.1 50.90 54.40 64.50 60.50 62.30
gemma-7b 46.8 54.00 45.50 52.70 43.00
Meta-Llama-3-8B 58.20 54.10 55.6 55.00 64.00
∆M (RoBERTa MNLI) 59.20 54.40 56.30 57.50 53.30
∆M (gpt-neo-125M) 59.20 55.10 50.50 52.10 45.50
Backdoor Adjustments

∆M (gpt-neo-1.3B) 51.30 50.70 55.00 43.90 49.00

∆M (gemma-2b) 44.50 45.30 52.60 63.50 43.90
∆M (gpt-neo-2.7B) 49.10 51.30 51.50 54.00 51.40
Language Models ∆M (phi-2) 57.00 66.00 62.10 57.10 45.80
∆M (gpt-j-6B) 51.30 45.60 50.50 49.10 46.00
∆M (Llama-2-7b-chat-hf) 62.60 64.60 68.50 70.50 63.80
∆M (Mistral-7B-v0.1) 63.90 71.40 73.70 61.30 67.00
∆M (gemma-7b) 72.80 77.80 73.60 71.90 62.40
∆M (Meta-Llama-3-8B) 66.00 70.20 68.40 62.00 63.40

which can further be used to estimate the causal strength between the events E1 and E2 .
∆M = pM (E2 |do(E1 )) − pM (E2 |do(¬E1 ))

Using the multiple ∆ estimates defined above, we estimate the causal strength between the events
available for an activity. We follow the scheme presented in the App. B, Algorithm 2 to compute the
performance in terms of success rates.
Temporal Scheme: In this scheme, we validate if temporal ordering knowledge of LLMs could be
directly used to estimate the causal estimand. We make use of templates shown in the App. Fig. 7.
The causal estimand is estimated via the difference in logit values when intervening over an event, i.e.,
does the predicted probability take into account the context of events not happening? Surprisingly,
we found that temporal ordering does provide a suitable proxy for estimating causal strength between
the events. We further extend this approximation to incorporate the backdoor adjustments in the ∆.
Backdoor Adjustments: For the experiments with Language models, we apply the backdoor
adjustment to estimate the causal estimand ∆M . App. Fig. 8 shows the prompt template used
to determine the relationship between the events. The prompt template takes a trajectory, ti , that
contains all the events till the event E1 in sequential order of occurrence, further, an added prompt
determines the intervention (do(E1 ) or do(¬E1 )) and the causal estimand is estimated using the logit
values associated with the predicted token. App. B, Algorithm 3 provides the designed scheme to
compute unbiased causal estimands. We essentially flip the options and generate the scores associated
with options ‘A’ and ‘B’ for increase and decrease, respectively (more details in the App. C.)
Table 4 shows a comparison between various design choices. We observe that when using LLMs for p̂
estimation, the backdoor adjustment increases the performance over the temporal estimation scheme
by a significant margin. The understanding of these activities is generic, and LLMs do provide a
suitable set of sequences when prompted to generate a list of steps to complete the activity. For
example, when prompted with ‘Generate the sequential steps in a telegrammic style to
perform the activity “going grocery shopping"’, almost all the models we tested provide

9
a valid set of steps for the given activity. However, when prompted with causal queries, the lower
performance signifies the lack of understanding of the underlying causal mechanism. The constructed
dataset helps to rigorously validate the understanding of the activity through an enormous number
of causal query triplets. The results show that although the LLMs can explain the activity in detail,
including generating correct steps for performing tasks, causally reasoning about the set of events
remains challenging.
Human Study: We conducted a small-scale human validation study over the created causal query
dataset and asked 5 graduate students to answer 100 randomly sampled causal query triplets (20 per
activity). We record an average performance of 92.20%. (More details about the human study are
provided in the App. B)

5 Related Work

Causal reasoning has been an active research area in the ML community [Spirtes et al., 2000a,
Peters et al., 2017, Schölkopf et al., 2021]. Some of the initial works highlight the causal nature of
events present in text [Schank, 1975] as ‘causal chains’. Multiple works have considered creating
benchmarks/datasets that capture causal relationships between the events described in the text (see
App. Table 5). More recently, with the rapid growth of LLMs on reasoning/understanding tasks,
attention has shifted to validating these general-purpose models capturing causal reasoning [Jin
et al., 2023, Zečević et al., 2023, Willig et al., 2023a, Liu et al., 2023, Willig et al., 2023b, Zhang
et al., 2022a, Jin et al., 2024]. App. A.2 Table 5 shows a broad overview of the existing causal
Dataset/Benchmarks presented in the NLP community. In this work, the primary focus is to bridge
the gap between various lines of work that consider natural language to learn/validate/reason about
causal relationships between events.

6 Limitations and Future Directions

One of the primary limitations of our work is the limited set of activities. Though the frameworks
support generating exhaustive/enormous causal queries, finding general commonsense reasoning
activities/tasks that are well understood by humans remains challenging. Moreover, creating a causal
graph for an activity increases as we move toward more long-term tasks. However, as a general
test of causal intelligence, our framework provides a suitable platform to validate the reasoning
capabilities more rigorously. In the future, it would be interesting to sample trajectories from the
observational distribution Goa to create a training dataset and check if causal reasoning ability can
be acquired by language modeling objectives (including other variants like presented in Lampinen
et al. [2023]). We leave this detailed analysis for future endeavors. The proposed algorithm for causal
triplet generation generates the simplest variant of causal queries in the form of causal triplets (also
referred to as Pairwise Causal Discovery (PCD) task by [Chen et al., 2024]). More complicated
causal queries can be generated, such as considering cases with common confounders, long/short
causal chain dependency, etc. Moreover, taking formal definitions. (i.e., using the formal causal
inference language) causal queries inspired from Jin et al. [2023, 2024] can be framed for a more
rigorous analysis. Being at the initial state, we stick to the simple causal queries that provide two
choices, and the task is to choose the more plausible cause. The creation of underlying causal graphs
provides endless possibilities for creating varied versions of causal queries. In this work, we only
consider an unconditional version of d-separation. In the future, the same causal graphs could be used
to define more datasets for covering other rungs of the ‘causal ladder’ [Pearl and Mackenzie, 2018].

7 Conclusion

In this paper, we proposed the COLD (Causal reasOning in cLosed Daily activities) framework for
generating causal queries that can be used to rigorously evaluate LLMs. We performed extensive
experimentation with LLMs for the task of Causal Commonsense Reasoning. Results indicate that
LLMs are still far from a complete understanding of daily commonsensical activities and fail to
answer causal queries when analyzed in an exhaustive manner. We believe this framework will
provide a good platform for future research in understanding the causal reasoning abilities of LLMs.

10
Acknowledgments
We would like to thank the anonymous reviewers and the meta-reviewer for their insightful comments
and suggestions. We would like to thank Google Deepmind India for helping us with the conference
travel support.

References
Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman,
Hannah Rashkin, Doug Downey, Wen tau Yih, and Yejin Choi. Abductive commonsense reasoning.
In International Conference on Learning Representations, 2020. URL https://openreview.
net/forum?id=Byg1v1HKDB. [Cited on page 24.]

Sid Black, Gao Leo, Phil Wang, Connor Leahy, and Stella Biderman. GPT-Neo: Large Scale
Autoregressive Language Modeling with Mesh-Tensorflow, March 2021. URL https://doi.
org/10.5281/zenodo.5297715. [Cited on page 7.]

Alexander Bondarenko, Magdalena Wolska, Stefan Heindorf, Lukas Blübaum, Axel-Cyrille

Ngonga Ngomo, Benno Stein, Pavel Braslavski, Matthias Hagen, and Martin Potthast. CausalQA:
A benchmark for causal question answering. In Nicoletta Calzolari, Chu-Ren Huang, Hansaem
Kim, James Pustejovsky, Leo Wanner, Key-Sun Choi, Pum-Mo Ryu, Hsin-Hsi Chen, Lucia Do-
natelli, Heng Ji, Sadao Kurohashi, Patrizia Paggio, Nianwen Xue, Seokhwan Kim, Younggyun
Hahm, Zhong He, Tony Kyungil Lee, Enrico Santus, Francis Bond, and Seung-Hoon Na, editors,
Proceedings of the 29th International Conference on Computational Linguistics, pages 3296–
3308, Gyeongju, Republic of Korea, October 2022. International Committee on Computational
Linguistics. URL https://aclanthology.org/2022.coling-1.291. [Cited on page 24.]

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared 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 M. Ziegler,
Jeffrey Wu, Clemens Winter, Christopher 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. CoRR, abs/2005.14165,
2020. URL https://arxiv.org/abs/2005.14165. [Cited on pages 1, 7, and 8.]

Sirui Chen, Bo Peng, Meiqi Chen, Ruiqi Wang, Mengying Xu, Xingyu Zeng, Rui Zhao, Shengjie
Zhao, Yu Qiao, and Chaochao Lu. Causal evaluation of language models, 2024. [Cited on pages 7
and 10.]

Changwoo Chun, SongEun Lee, Jaehyung Seo, and Heuiseok Lim. CReTIHC: Designing causal
reasoning tasks about temporal interventions and hallucinated confoundings. In Houda Bouamor,
Juan Pino, and Kalika Bali, editors, Findings of the Association for Computational Linguistics:
EMNLP 2023, pages 10334–10343, Singapore, December 2023. Association for Computational
Linguistics. doi: 10.18653/v1/2023.findings-emnlp.693. URL https://aclanthology.org/
2023.findings-emnlp.693. [Cited on page 2.]

Kai-Hendrik Cohrs, Emiliano Diaz, Vasileios Sitokonstantinou, Gherardo Varando, and Gustau
Camps-Valls. Large language models for constrained-based causal discovery. In AAAI 2024
Workshop on ”Are Large Language Models Simply Causal Parrots?”, 2023. URL https://
openreview.net/forum?id=NEAoZRWHPN. [Cited on page 1.]

D. R. Cox. Planning of Experiments. Wiley, New York„ 1958. [Cited on pages 3, 6, and 23.]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of
deep bidirectional transformers for language understanding. In Jill Burstein, Christy Doran, and
Thamar Solorio, editors, Proceedings of the 2019 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language Technologies, Volume 1
(Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota, June 2019. Association for
Computational Linguistics. doi: 10.18653/v1/N19-1423. URL https://aclanthology.org/
N19-1423. [Cited on pages 1 and 8.]

11
Quang Do, Yee Seng Chan, and Dan Roth. Minimally supervised event causality identification. In
Regina Barzilay and Mark Johnson, editors, Proceedings of the 2011 Conference on Empirical
Methods in Natural Language Processing, pages 294–303, Edinburgh, Scotland, UK., July 2011.
Association for Computational Linguistics. URL https://aclanthology.org/D11-1027. [Cited
on pages 5, 24, 25, and 30.]
Li Du, Xiao Ding, Kai Xiong, Ting Liu, and Bing Qin. e-CARE: a new dataset for explor-
ing explainable causal reasoning. In Smaranda Muresan, Preslav Nakov, and Aline Villav-
icencio, editors, Proceedings of the 60th Annual Meeting of the Association for Computa-
tional Linguistics (Volume 1: Long Papers), pages 432–446, Dublin, Ireland, May 2022. As-
sociation for Computational Linguistics. doi: 10.18653/v1/2022.acl-long.33. URL https:
//aclanthology.org/2022.acl-long.33. [Cited on pages 2, 23, and 24.]
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha
Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn,
Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston
Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron,
Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris
McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton
Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David
Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes,
Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip
Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme
Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu,
Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov,
Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah,
Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu
Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph
Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Kartikeya Upasani,
Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz
Malik, Kuenley Chiu, Kunal Bhalla, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence
Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas
Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri,
Marcin Kardas, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis,
Min Si, Mitesh Kumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov,
Nikolay Bogoychev, Niladri Chatterji, Olivier Duchenne, Onur Çelebi, Patrick Alrassy, Pengchuan
Zhang, Pengwei Li, Petar Vasic, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan,
Punit Singh Koura, Puxin Xu, Qing He, Qingxiao Dong, Ragavan Srinivasan, Raj Ganapathy,
Ramon Calderer, Ricardo Silveira Cabral, Robert Stojnic, Roberta Raileanu, Rohit Girdhar, Rohit
Patel, Romain Sauvestre, Ronnie Polidoro, Roshan Sumbaly, Ross Taylor, Ruan Silva, Rui Hou,
Rui Wang, Saghar Hosseini, Sahana Chennabasappa, Sanjay Singh, Sean Bell, Seohyun Sonia
Kim, Sergey Edunov, Shaoliang Nie, Sharan Narang, Sharath Raparthy, Sheng Shen, Shengye Wan,
Shruti Bhosale, Shun Zhang, Simon Vandenhende, Soumya Batra, Spencer Whitman, Sten Sootla,
Stephane Collot, Suchin Gururangan, Sydney Borodinsky, Tamar Herman, Tara Fowler, Tarek
Sheasha, Thomas Georgiou, Thomas Scialom, Tobias Speckbacher, Todor Mihaylov, Tong Xiao,
Ujjwal Karn, Vedanuj Goswami, Vibhor Gupta, Vignesh Ramanathan, Viktor Kerkez, Vincent
Gonguet, Virginie Do, Vish Vogeti, Vladan Petrovic, Weiwei Chu, Wenhan Xiong, Wenyin Fu,
Whitney Meers, Xavier Martinet, Xiaodong Wang, Xiaoqing Ellen Tan, Xinfeng Xie, Xuchao Jia,
Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen
Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe
Papakipos, Aaditya Singh, Aaron Grattafiori, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya
Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alex
Vaughan, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Anam Yunus, Andrei
Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew
Ryan, Ankit Ramchandani, Annie Franco, Aparajita Saraf, Arkabandhu Chowdhury, Ashley
Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin
Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu,
Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt
Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Changhan Wang, Changkyu Kim, Chao
Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Damon

12
 Civin, Dana Beaty, Daniel Kreymer, Daniel Li, Danny Wyatt, David Adkins, David Xu, Davide
Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le,
Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery, Eleonora Presani, Emily
Hahn, Emily Wood, Erik Brinkman, Esteban Arcaute, Evan Dunbar, Evan Smothers, Fei Sun, Felix
Kreuk, Feng Tian, Firat Ozgenel, Francesco Caggioni, Francisco Guzmán, Frank Kanayet, Frank
Seide, Gabriela Medina Florez, Gabriella Schwarz, Gada Badeer, Georgia Swee, Gil Halpern,
Govind Thattai, Grant Herman, Grigory Sizov, Guangyi, Zhang, Guna Lakshminarayanan, Hamid
Shojanazeri, Han Zou, Hannah Wang, Hanwen Zha, Haroun Habeeb, Harrison Rudolph, Helen
Suk, Henry Aspegren, Hunter Goldman, Ibrahim Damlaj, Igor Molybog, Igor Tufanov, Irina-
Elena Veliche, Itai Gat, Jake Weissman, James Geboski, James Kohli, Japhet Asher, Jean-Baptiste
Gaya, Jeff Marcus, Jeff Tang, Jennifer Chan, Jenny Zhen, Jeremy Reizenstein, Jeremy Teboul,
Jessica Zhong, Jian Jin, Jingyi Yang, Joe Cummings, Jon Carvill, Jon Shepard, Jonathan McPhie,
Jonathan Torres, Josh Ginsburg, Junjie Wang, Kai Wu, Kam Hou U, Karan Saxena, Karthik
Prasad, Kartikay Khandelwal, Katayoun Zand, Kathy Matosich, Kaushik Veeraraghavan, Kelly
Michelena, Keqian Li, Kun Huang, Kunal Chawla, Kushal Lakhotia, Kyle Huang, Lailin Chen,
Lakshya Garg, Lavender A, Leandro Silva, Lee Bell, Lei Zhang, Liangpeng Guo, Licheng Yu,
Liron Moshkovich, Luca Wehrstedt, Madian Khabsa, Manav Avalani, Manish Bhatt, Maria
Tsimpoukelli, Martynas Mankus, Matan Hasson, Matthew Lennie, Matthias Reso, Maxim Groshev,
Maxim Naumov, Maya Lathi, Meghan Keneally, Michael L. Seltzer, Michal Valko, Michelle
Restrepo, Mihir Patel, Mik Vyatskov, Mikayel Samvelyan, Mike Clark, Mike Macey, Mike Wang,
Miquel Jubert Hermoso, Mo Metanat, Mohammad Rastegari, Munish Bansal, Nandhini Santhanam,
Natascha Parks, Natasha White, Navyata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier,
Nikolay Pavlovich Laptev, Ning Dong, Ning Zhang, Norman Cheng, Oleg Chernoguz, Olivia
Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro
Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani,
Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Raghotham Murthy,
Raghu Nayani, Rahul Mitra, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Rohan
Maheswari, Russ Howes, Ruty Rinott, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara
Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Verma, Seiji Yamamoto, Sharadh
Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha,
Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe,
Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan
Govindaprasad, Sumit Gupta, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury,
Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Kohler, Thomas Robinson, Tianhe
Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi,
Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vítor Albiero, Vlad Ionescu,
Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang,
Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaofang Wang, Xiaojian Wu, Xiaolan Wang,
Xide Xia, Xilun Wu, Xinbo Gao, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang,
Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yuchen Hao, Yundi Qian, Yuzi He, Zach Rait,
Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang, and Zhiwei Zhao. The llama 3 herd
of models, 2024. URL https://arxiv.org/abs/2407.21783. [Cited on page 7.]

Jesse Dunietz, Lori Levin, and Jaime Carbonell. The BECauSE corpus 2.0: Annotating causality
and overlapping relations. In Nathan Schneider and Nianwen Xue, editors, Proceedings of the
11th Linguistic Annotation Workshop, pages 95–104, Valencia, Spain, April 2017. Association for
Computational Linguistics. doi: 10.18653/v1/W17-0812. URL https://aclanthology.org/
W17-0812. [Cited on page 24.]

Jörg Frohberg and Frank Binder. CRASS: A novel data set and benchmark to test counterfactual
reasoning of large language models. In Nicoletta Calzolari, Frédéric Béchet, Philippe Blache,
Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, Hitoshi Isahara, Bente Maegaard,
Joseph Mariani, Hélène Mazo, Jan Odijk, and Stelios Piperidis, editors, Proceedings of the
Thirteenth Language Resources and Evaluation Conference, pages 2126–2140, Marseille, France,
June 2022. European Language Resources Association. URL https://aclanthology.org/2022.
lrec-1.229. [Cited on page 24.]

Andrew Gordon, Zornitsa Kozareva, and Melissa Roemmele. SemEval-2012 task 7: Choice of
plausible alternatives: An evaluation of commonsense causal reasoning. In Eneko Agirre, Johan
Bos, Mona Diab, Suresh Manandhar, Yuval Marton, and Deniz Yuret, editors, *SEM 2012: The

13
 First Joint Conference on Lexical and Computational Semantics – Volume 1: Proceedings of
the main conference and the shared task, and Volume 2: Proceedings of the Sixth International
Workshop on Semantic Evaluation (SemEval 2012), pages 394–398, Montréal, Canada, 7-8 June
2012. Association for Computational Linguistics. URL https://aclanthology.org/S12-1052.
[Cited on pages 1, 2, 5, 6, 23, and 24.]

Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid Ó Séaghdha, Sebastian
Padó, Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. SemEval-2010 task 8:
Multi-way classification of semantic relations between pairs of nominals. In Katrin Erk and Carlo
Strapparava, editors, Proceedings of the 5th International Workshop on Semantic Evaluation,
pages 33–38, Uppsala, Sweden, July 2010. Association for Computational Linguistics. URL
https://aclanthology.org/S10-1006. [Cited on page 24.]

Austin Bradford Hill. The environment and disease: association or causation?, 1965. [Cited on
pages 6 and 24.]

Matthew Ho, Aditya Sharma, Justin Chang, Michael Saxon, Sharon Levy, Yujie Lu, and William Yang
Wang. Wikiwhy: Answering and explaining cause-and-effect questions, 2022. [Cited on page 1.]

Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Caio César Teodoro Mendes, Weizhu
Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, Suriya Gunasekar, Piero Kauffmann,
Yin Tat Lee, Yuanzhi Li, Anh Nguyen, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah,
Michael Santacroce, Harkirat Singh Behl, Adam Taumann Kalai, Xin Wang, Rachel Ward, Philipp
Witte, Cyril Zhang, and Yi Zhang. Phi-2: The surprising power of small language models. Microsoft
Research Blog, 2023. [Cited on page 7.]

Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot,
Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier,
Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas
Wang, Timothée Lacroix, and William El Sayed. Mistral 7b, 2023. URL https://arxiv.org/
abs/2310.06825. [Cited on page 7.]

Zhijing Jin, Jiarui Liu, Zhiheng Lyu, Spencer Poff, Mrinmaya Sachan, Rada Mihalcea, Mona Diab,
and Bernhard Schölkopf. Can large language models infer causation from correlation?, 2023.
[Cited on pages 1, 7, 10, and 24.]

Zhijing Jin, Yuen Chen, Felix Leeb, Luigi Gresele, Ojasv Kamal, Zhiheng Lyu, Kevin Blin, Fer-
nando Gonzalez Adauto, Max Kleiman-Weiner, Mrinmaya Sachan, and Bernhard Schölkopf.
Cladder: Assessing causal reasoning in language models, 2024. [Cited on pages 1, 7, 10, and 24.]

Abhinav Joshi, Areeb Ahmad, Umang Pandey, and Ashutosh Modi. From scripts to rl environments:
Towards imparting commonsense knowledge to rl agents. In Proceedings of the 2023 International
Conference on Autonomous Agents and Multiagent Systems, pages 2801–2803, 2023a. [Cited on
page 5.]

Abhinav Joshi, Areeb Ahmad, Umang Pandey, and Ashutosh Modi. Scriptworld: Text based
environment for learning procedural knowledge. In Edith Elkind, editor, Proceedings of the
Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 5095–
5103. International Joint Conferences on Artificial Intelligence Organization, 8 2023b. doi:
10.24963/ijcai.2023/566. URL https://doi.org/10.24963/ijcai.2023/566. Main Track.
[Cited on page 5.]

Adam Karvonen. Emergent world models and latent variable estimation in chess-playing language
models, 2024. [Cited on page 8.]

Benjamin Kuipers. Commonsense reasoning about causality: Deriving behavior from struc-
ture. Artificial Intelligence, 24(1):169–203, 1984. ISSN 0004-3702. doi: https://doi.org/
10.1016/0004-3702(84)90039-0. URL https://www.sciencedirect.com/science/article/
pii/0004370284900390. [Cited on page 2.]

14
Yash Kumar Lal, Nathanael Chambers, Raymond Mooney, and Niranjan Balasubramanian. TellMe-
Why: A dataset for answering why-questions in narratives. In Chengqing Zong, Fei Xia, Wen-
jie Li, and Roberto Navigli, editors, Findings of the Association for Computational Linguis-
tics: ACL-IJCNLP 2021, pages 596–610, Online, August 2021. Association for Computational
Linguistics. doi: 10.18653/v1/2021.findings-acl.53. URL https://aclanthology.org/2021.
findings-acl.53. [Cited on page 24.]
Andrew Kyle Lampinen, Stephanie C.Y. Chan, Ishita Dasgupta, Andrew Joo Hun Nam, and Jane X
Wang. Passive learning of active causal strategies in agents and language models. In Thirty-seventh
Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/
forum?id=BRpi8YAfac. [Cited on page 10.]
Kenneth Li, Aspen K Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wat-
tenberg. Emergent world representations: Exploring a sequence model trained on a synthetic
task. In The Eleventh International Conference on Learning Representations, 2023. URL
https://openreview.net/forum?id=DeG07_TcZvT. [Cited on page 8.]
Xiao Liu, Da Yin, Chen Zhang, Yansong Feng, and Dongyan Zhao. The magic of IF: Investigating
causal reasoning abilities in large language models of code. In Anna Rogers, Jordan Boyd-Graber,
and Naoaki Okazaki, editors, Findings of the Association for Computational Linguistics: ACL 2023,
pages 9009–9022, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.
18653/v1/2023.findings-acl.574. URL https://aclanthology.org/2023.findings-acl.574.
[Cited on page 10.]
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. [Cited on page 8.]
John Stuart Mill. A System of Logic, Ratiocinative and Inductive: Being a Connected View of the
Princilples of Evidence and the Methods of Scientific Investigation. Longmans, Green, 1898.
[Cited on pages 6 and 24.]
Paramita Mirza, Rachele Sprugnoli, Sara Tonelli, and Manuela Speranza. Annotating causality
in the TempEval-3 corpus. In Oleksandr Kolomiyets, Marie-Francine Moens, Martha Palmer,
James Pustejovsky, and Steven Bethard, editors, Proceedings of the EACL 2014 Workshop on
Computational Approaches to Causality in Language (CAtoCL), pages 10–19, Gothenburg, Sweden,
April 2014. Association for Computational Linguistics. doi: 10.3115/v1/W14-0702. URL https:
//aclanthology.org/W14-0702. [Cited on page 24.]
Melanie Mitchell, Alessandro B. Palmarini, and Arseny Moskvichev. Comparing humans, gpt-4, and
gpt-4v on abstraction and reasoning tasks, 2023. [Cited on page 1.]
Ashutosh Modi. Event Embeddings for Semantic Script Modeling. In Proceedings of the 20th SIGNLL
Conference on Computational Natural Language Learning, 2016. doi: 10.18653/v1/K16-1008.
URL https://aclanthology.org/K16-1008. [Cited on page 2.]
Ashutosh Modi. Modeling Common Sense Knowledge via Scripts. PhD thesis, Saarland University,
2017. [Cited on page 2.]
Ashutosh Modi and Ivan Titov. Inducing neural models of script knowledge. In Proceedings of the
Eighteenth Conference on Computational Natural Language Learning, pages 49–57, 2014. [Cited
on page 2.]
Ashutosh Modi, Tatjana Anikina, Simon Ostermann, and Manfred Pinkal. InScript: Narrative
texts annotated with script information. In Proceedings of the Tenth International Conference
on Language Resources and Evaluation (LREC’16), 2016. URL https://aclanthology.org/
L16-1555. [Cited on page 2.]
Ashutosh Modi, Ivan Titov, Vera Demberg, Asad Sayeed, and Manfred Pinkal. Modeling Semantic
Expectation: Using Script Knowledge for Referent Prediction. Transactions of the Association
for Computational Linguistics, 2017. doi: 10.1162/tacl_a_00044. URL https://aclanthology.
org/Q17-1003. [Cited on page 2.]

15
Nasrin Mostafazadeh, Alyson Grealish, Nathanael Chambers, James Allen, and Lucy Vanderwende.
CaTeRS: Causal and temporal relation scheme for semantic annotation of event structures. In
Martha Palmer, Ed Hovy, Teruko Mitamura, and Tim O’Gorman, editors, Proceedings of the
Fourth Workshop on Events, pages 51–61, San Diego, California, June 2016. Association for
Computational Linguistics. doi: 10.18653/v1/W16-1007. URL https://aclanthology.org/
W16-1007. [Cited on page 24.]

Neel Nanda, Andrew Lee, and Martin Wattenberg. Emergent linear representations in world models
of self-supervised sequence models. In Yonatan Belinkov, Sophie Hao, Jaap Jumelet, Najoung Kim,
Arya McCarthy, and Hosein Mohebbi, editors, Proceedings of the 6th BlackboxNLP Workshop:
Analyzing and Interpreting Neural Networks for NLP, pages 16–30, Singapore, December 2023.
Association for Computational Linguistics. doi: 10.18653/v1/2023.blackboxnlp-1.2. URL https:
//aclanthology.org/2023.blackboxnlp-1.2. [Cited on page 8.]

Brady Neal. Introduction to causal inference. Course Lecture Notes (draft), 2020. [Cited on pages 4
and 27.]

Simon Ostermann, Ashutosh Modi, Michael Roth, Stefan Thater, and Manfred Pinkal. MCScript: A
Novel Dataset for Assessing Machine Comprehension Using Script Knowledge. In LREC, 2018.
[Cited on page 2.]

Judea Pearl. Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan
kaufmann, 1988. [Cited on page 4.]

Judea Pearl. Comment: Graphical models, causality and intervention. 8(3):266–269, August 1993.
doi: 10.1214/ss/1177010894. [Cited on page 8.]

Judea Pearl. The do-calculus revisited. In Proceedings of the Twenty-Eighth Conference on Uncer-
tainty in Artificial Intelligence, UAI’12, page 3–11, Arlington, Virginia, USA, 2012. AUAI Press.
ISBN 9780974903989. [Cited on page 2.]

Judea Pearl and Dana Mackenzie. The Book of Why: The New Science of Cause and Effect. Basic
Books, Inc., USA, 1st edition, 2018. ISBN 046509760X. [Cited on pages 1, 3, 4, 6, 10, and 25.]

Judea Pearl and Thomas S Verma. A theory of inferred causation. In Studies in Logic and the
Foundations of Mathematics, volume 134, pages 789–811. Elsevier, 1995. [Cited on pages 6
and 24.]

Judea Pearl, Madelyn Glymour, and Nicholas P Jewell. Causal inference in statistics: A primer. John
Wiley & Sons, 2016. [Cited on pages 4, 25, and 27.]

Derek Penn and Daniel Povinelli. Causal cognition in human and nonhuman animals: A comparative,
critical review. Annual Review of Psychology, 58:97–118, 02 2007. doi: 10.1146/annurev.psych.
58.110405.085555. [Cited on page 1.]

Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. Elements of Causal Inference: Foun-
dations and Learning Algorithms. Adaptive Computation and Machine Learning. MIT Press,
Cambridge, MA, 2017. ISBN 978-0-262-03731-0. URL https://mitpress.mit.edu/books/
elements-causal-inference. [Cited on page 10.]

Lianhui Qin, Antoine Bosselut, Ari Holtzman, Chandra Bhagavatula, Elizabeth Clark, and Yejin Choi.
Counterfactual story reasoning and generation. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun
Wan, editors, Proceedings of the 2019 Conference on Empirical Methods in Natural Language
Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-
IJCNLP), pages 5043–5053, Hong Kong, China, November 2019. Association for Computational
Linguistics. doi: 10.18653/v1/D19-1509. URL https://aclanthology.org/D19-1509. [Cited
on page 24.]

Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language
models are unsupervised multitask learners. 2019. URL https://api.semanticscholar.org/
CorpusID:160025533. [Cited on page 1.]

16
Hannah Rashkin, Maarten Sap, Emily Allaway, Noah A. Smith, and Yejin Choi. Event2Mind:
Commonsense inference on events, intents, and reactions. In Iryna Gurevych and Yusuke Miyao,
editors, Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers), pages 463–473, Melbourne, Australia, July 2018. Association for
Computational Linguistics. doi: 10.18653/v1/P18-1043. URL https://aclanthology.org/
P18-1043. [Cited on page 24.]
Joshua Robinson and David Wingate. Leveraging large language models for multiple choice question
answering. In The Eleventh International Conference on Learning Representations, 2023. URL
https://openreview.net/forum?id=yKbprarjc5B. [Cited on pages 7 and 29.]
Angelika Romanou, Syrielle Montariol, Debjit Paul, Leo Laugier, Karl Aberer, and Antoine Bosselut.
CRAB: Assessing the strength of causal relationships between real-world events. In Houda
Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical
Methods in Natural Language Processing, pages 15198–15216, Singapore, December 2023.
Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.940. URL https:
//aclanthology.org/2023.emnlp-main.940. [Cited on pages 1 and 24.]
Donald B. Rubin. Randomization analysis of experimental data: The fisher randomization test
comment. Journal of the American Statistical Association, 75(371):591–593, 1980. [Cited on
pages 3, 6, and 23.]
Maarten Sap, Ronan LeBras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah
Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. Atomic: An atlas of machine common-
sense for if-then reasoning, 2019a. [Cited on page 24.]
Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. Social IQa: Common-
sense reasoning about social interactions. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun
Wan, editors, Proceedings of the 2019 Conference on Empirical Methods in Natural Language
Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-
IJCNLP), pages 4463–4473, Hong Kong, China, November 2019b. Association for Computational
Linguistics. doi: 10.18653/v1/D19-1454. URL https://aclanthology.org/D19-1454. [Cited
on page 24.]
Roger C. Schank. The structure of episodes in memory. In DANIEL G. BOBROW and
ALLAN COLLINS, editors, Representation and Understanding, pages 237–272. Morgan
Kaufmann, San Diego, 1975. ISBN 978-0-12-108550-6. doi: https://doi.org/10.1016/
B978-0-12-108550-6.50014-8. URL https://www.sciencedirect.com/science/article/
pii/B9780121085506500148. [Cited on pages 2 and 10.]
Roger C. Schank and Robert P. Abelson. Scripts, Plans, and Knowledge. In Proceedings of the 4th
International Joint Conference on Artificial Intelligence, IJCAI, 1975. [Cited on page 2.]
Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner,
Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the
IEEE, 109(5):612–634, 2021. doi: 10.1109/JPROC.2021.3058954. [Cited on page 10.]
Shikhar Singh, Nuan Wen, Yu Hou, Pegah Alipoormolabashi, Te-lin Wu, Xuezhe Ma, and Nanyun
Peng. COM2SENSE: A commonsense reasoning benchmark with complementary sentences.
In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors, Findings of the Asso-
ciation for Computational Linguistics: ACL-IJCNLP 2021, pages 883–898, Online, August
2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.findings-acl.78. URL
https://aclanthology.org/2021.findings-acl.78. [Cited on page 24.]
P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction, and Search. MIT press, 2nd edition,
2000a. [Cited on page 10.]
Peter Spirtes, Clark N Glymour, and Richard Scheines. Causation, prediction, and search. MIT
press, 2000b. [Cited on page 28.]
Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam
Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska,
Aitor Lewkowycz, Akshat Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W.

17
Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda
Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Johan
Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmüller, Andrew M. Dai, Andrew La,
Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna
Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes,
Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut
Erdem, Ayla Karakaş, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartłomiej Bojanowski,
Batuhan Özyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk
Ekmekci, Bill Yuchen Lin, Blake Howald, Bryan Orinion, Cameron Diao, Cameron Dour, Catherine
Stinson, Cedrick Argueta, Cesar Ferri, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta
Baral, Chiyu Wu, Chris Callison-Burch, Christopher Waites, Christian Voigt, Christopher D
Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel,
Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman,
Dan Roth, C. Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Moseguí González, Danielle
Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David
Drakard, David Jurgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz
Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho
Mollo, Diyi Yang, Dong-Ho Lee, Dylan Schrader, Ekaterina Shutova, Ekin Dogus Cubuk, Elad
Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodolà,
Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A Chi, Ethan Dyer, Ethan
Jerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar,
Fernando Martínez-Plumed, Francesca Happé, Francois Chollet, Frieda Rong, Gaurav Mishra,
Genta Indra Winata, Gerard de Melo, Germán Kruszewski, Giambattista Parascandolo, Giorgio
Mariani, Gloria Xinyue Wang, Gonzalo Jaimovitch-Lopez, Gregor Betz, Guy Gur-Ari, Hana
Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar,
Henry Francis Anthony Shevlin, Hinrich Schuetze, Hiromu Yakura, Hongming Zhang, Hugh Mee
Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jaehoon
Lee, Jaime Fernández Fisac, James B Simon, James Koppel, James Zheng, James Zou, Jan Kocon,
Jana Thompson, Janelle Wingfield, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason
Phang, Jason Wei, Jason Yosinski, Jekaterina Novikova, Jelle Bosscher, Jennifer Marsh, Jeremy
Kim, Jeroen Taal, Jesse Engel, Jesujoba Alabi, Jiacheng Xu, Jiaming Song, Jillian Tang, Joan
Waweru, John Burden, John Miller, John U. Balis, Jonathan Batchelder, Jonathan Berant, Jörg
Frohberg, Jos Rozen, Jose Hernandez-Orallo, Joseph Boudeman, Joseph Guerr, Joseph Jones,
Joshua B. Tenenbaum, Joshua S. Rule, Joyce Chua, Kamil Kanclerz, Karen Livescu, Karl Krauth,
Karthik Gopalakrishnan, Katerina Ignatyeva, Katja Markert, Kaustubh Dhole, Kevin Gimpel,
Kevin Omondi, Kory Wallace Mathewson, Kristen Chiafullo, Ksenia Shkaruta, Kumar Shridhar,
Kyle McDonell, Kyle Richardson, Laria Reynolds, Leo Gao, Li Zhang, Liam Dugan, Lianhui Qin,
Lidia Contreras-Ochando, Louis-Philippe Morency, Luca Moschella, Lucas Lam, Lucy Noble,
Ludwig Schmidt, Luheng He, Luis Oliveros-Colón, Luke Metz, Lütfi Kerem Senel, Maarten
Bosma, Maarten Sap, Maartje Ter Hoeve, Maheen Farooqi, Manaal Faruqui, Mantas Mazeika,
Marco Baturan, Marco Marelli, Marco Maru, Maria Jose Ramirez-Quintana, Marie Tolkiehn,
Mario Giulianelli, Martha Lewis, Martin Potthast, Matthew L Leavitt, Matthias Hagen, Mátyás
Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael Andrew
Yee, Michael Cohen, Michael Gu, Michael Ivanitskiy, Michael Starritt, Michael Strube, Michał
Sw˛edrowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Mimee Xu, Mirac
Suzgun, Mitch Walker, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini,
Mukund Varma T, Nanyun Peng, Nathan Andrew Chi, Nayeon Lee, Neta Gur-Ari Krakover,
Nicholas Cameron, Nicholas Roberts, Nick Doiron, Nicole Martinez, Nikita Nangia, Niklas
Deckers, Niklas Muennighoff, Nitish Shirish Keskar, Niveditha S. Iyer, Noah Constant, Noah
Fiedel, Nuan Wen, Oliver Zhang, Omar Agha, Omar Elbaghdadi, Omer Levy, Owain Evans,
Pablo Antonio Moreno Casares, Parth Doshi, Pascale Fung, Paul Pu Liang, Paul Vicol, Pegah
Alipoormolabashi, Peiyuan Liao, Percy Liang, Peter W Chang, Peter Eckersley, Phu Mon Htut,
Pinyu Hwang, Piotr Miłkowski, Piyush Patil, Pouya Pezeshkpour, Priti Oli, Qiaozhu Mei, Qing
Lyu, Qinlang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahel Habacker, Ramon
Risco, Raphaël Millière, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe
Raymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan Le Bras, Rosanne
Liu, Rowan Jacobs, Rui Zhang, Russ Salakhutdinov, Ryan Andrew Chi, Seungjae Ryan Lee, Ryan
Stovall, Ryan Teehan, Rylan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou,
Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel Stern Schoenholz,

18
 Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey,
Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan,
Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane
Gu, Shubh Pachchigar, Shubham Toshniwal, Shyam Upadhyay, Shyamolima Shammie Debnath,
Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-
Hwan Lee, Spencer Torene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon,
Stella Biderman, Stephanie Lin, Stephen Prasad, Steven Piantadosi, Stuart Shieber, Summer
Misherghi, Svetlana Kiritchenko, Swaroop Mishra, Tal Linzen, Tal Schuster, Tao Li, Tao Yu,
Tariq Ali, Tatsunori Hashimoto, Te-Lin Wu, Théo Desbordes, Theodore Rothschild, Thomas Phan,
Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Titus Tunduny, Tobias Gerstenberg,
Trenton Chang, Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera
Demberg, Victoria Nyamai, Vikas Raunak, Vinay Venkatesh Ramasesh, vinay uday prabhu,
Vishakh Padmakumar, Vivek Srikumar, William Fedus, William Saunders, William Zhang, Wout
Vossen, Xiang Ren, Xiaoyu Tong, Xinran Zhao, Xinyi Wu, Xudong Shen, Yadollah Yaghoobzadeh,
Yair Lakretz, Yangqiu Song, Yasaman Bahri, Yejin Choi, Yichi Yang, Yiding Hao, Yifu Chen,
Yonatan Belinkov, Yu Hou, Yufang Hou, Yuntao Bai, Zachary Seid, Zhuoye Zhao, Zijian Wang,
Zijie J. Wang, Zirui Wang, and Ziyi Wu. Beyond the imitation game: Quantifying and extrapolating
the capabilities of language models. Transactions on Machine Learning Research, 2023. ISSN
2835-8856. URL https://openreview.net/forum?id=uyTL5Bvosj. [Cited on page 24.]

Alessandro Stolfo, Zhijing Jin, Kumar Shridhar, Bernhard Schoelkopf, and Mrinmaya Sachan. A
causal framework to quantify the robustness of mathematical reasoning with language models.
In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Proceedings of the 61st
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),
pages 545–561, Toronto, Canada, July 2023. Association for Computational Linguistics. doi:
10.18653/v1/2023.acl-long.32. URL https://aclanthology.org/2023.acl-long.32. [Cited
on page 1.]

Xiaojuan Tang, Zilong Zheng, Jiaqi Li, Fanxu Meng, Song-Chun Zhu, Yitao Liang, and Muhan
Zhang. Large language models are in-context semantic reasoners rather than symbolic reasoners,
2023. [Cited on page 2.]

Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak,
Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot,
Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex
Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson,
Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy,
Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan,
George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian
Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau,
Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine
Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej
Mikuła, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar
Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona
Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith,
Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De,
Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed,
Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff
Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral,
Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and
Kathleen Kenealy. Gemma: Open models based on gemini research and technology, 2024. URL
https://arxiv.org/abs/2403.08295. [Cited on page 7.]

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay
Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cris-
tian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu,
Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn,
Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel
Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee,
Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra,

19
 Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi,
Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh
Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen
Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic,
Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models,
2023. URL https://arxiv.org/abs/2307.09288. [Cited on page 7.]

Aniket Vashishtha, Abbavaram Gowtham Reddy, Abhinav Kumar, Saketh Bachu, Vineeth N Bala-
subramanian, and Amit Sharma. Causal inference using llm-guided discovery, 2023. [Cited on
page 1.]

Ben Wang and Aran Komatsuzaki. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model.
https://github.com/kingoflolz/mesh-transformer-jax, May 2021. [Cited on page 7.]

Zhaowei Wang, Quyet V. Do, Hongming Zhang, Jiayao Zhang, Weiqi Wang, Tianqing Fang,
Yangqiu Song, Ginny Wong, and Simon See. COLA: Contextualized commonsense causal
reasoning from the causal inference perspective. In Anna Rogers, Jordan Boyd-Graber, and
Naoaki Okazaki, editors, Proceedings of the 61st Annual Meeting of the Association for Com-
putational Linguistics (Volume 1: Long Papers), pages 5253–5271, Toronto, Canada, July
2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.288. URL
https://aclanthology.org/2023.acl-long.288. [Cited on pages 1, 2, 24, and 28.]

Lilian D. A. Wanzare, Alessandra Zarcone, Stefan Thater, and Manfred Pinkal. A Crowdsourced
Database of Event Sequence Descriptions for the Acquisition of High-quality Script Knowledge.
In Proceedings of the Tenth International Conference on Language Resources and Evaluation
(LREC’16), 2016. URL https://aclanthology.org/L16-1556. [Cited on pages 2, 5, and 28.]

Noah Weber, Rachel Rudinger, and Benjamin Van Durme. Causal inference of script knowledge.
In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, Proceedings of the 2020
Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7583–7596,
Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.
emnlp-main.612. URL https://aclanthology.org/2020.emnlp-main.612. [Cited on page 1.]

Moritz Willig, Matej Zečević, Devendra Singh Dhami, and Kristian Kersting. Probing for correlations
of causal facts: Large language models and causality, 2023a. URL https://openreview.net/
forum?id=UPwzqPOs4-. [Cited on page 10.]

Moritz Willig, Matej Zečević, Jonas Seng, and Florian Peter Busch. Causal concept identification in
open world environments. In Martin Mundt, Keiland W. Cooper, Devendra Singh Dhami, Adéle
Ribeiro, James Seale Smith, Alexis Bellot, and Tyler Hayes, editors, Proceedings of The First
AAAI Bridge Program on Continual Causality, volume 208 of Proceedings of Machine Learning
Research, pages 52–58. PMLR, 07–08 Feb 2023b. URL https://proceedings.mlr.press/
v208/willig23a.html. [Cited on page 10.]

Linying Yang, Oscar Clivio, Vik Shirvaikar, and Fabian 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. URL https://openreview.net/forum?id=mRwgczYZFJ. [Cited
on page 1.]

Matej Zečević, Moritz Willig, Devendra Singh Dhami, and Kristian Kersting. Causal parrots: Large
language models may talk causality but are not causal. Transactions on Machine Learning Research,
2023. ISSN 2835-8856. URL https://openreview.net/forum?id=tv46tCzs83. [Cited on
pages 1, 2, 10, and 24.]

Cheng Zhang, Stefan Bauer, Paul Bennett, Jiangfeng Gao, Wenbo Gong, Agrin Hilmkil, Joel Jennings,
Chao Ma, Tom Minka, Nick Pawlowski, and James Vaughan. Understanding causality with large
language models: Feasibility and opportunities, 2023. [Cited on page 1.]

Honghua Zhang, Liunian Harold Li, Tao Meng, Kai-Wei Chang, and Guy Van den Broeck. On the
paradox of learning to reason from data. In International Joint Conference on Artificial Intelligence,
2022a. URL https://api.semanticscholar.org/CorpusID:248986434. [Cited on page 10.]

20
Jiayao Zhang, Hongming Zhang, Weijie Su, and Dan Roth. ROCK: Causal inference principles
for reasoning about commonsense causality. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song,
Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International
Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research,
pages 26750–26771. PMLR, 17–23 Jul 2022b. URL https://proceedings.mlr.press/v162/
zhang22am.html. [Cited on pages 2, 8, 23, and 29.]

21
Appendix

Table of Contents

A COLD Framework Details 23

A.1 Adherence to SUTVA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
A.2 Comparison with previous Causal Reasoning Datasets/Benchmarks . . . . . . . . . 23
A.3 Observational Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

B Algorithms in the COLD Framework 24

C Backdoor Adjustments 26

D Experiments and Results 28

D.1 Compute Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
D.2 Evaluation using Average Treatment Effect (ATE) . . . . . . . . . . . . . . . . . . 28

E Prompt Templates for Language Model based Experiments 29

F Additional Results 30
F.1 Temporally Plausible Choices in Causal Triplets . . . . . . . . . . . . . . . . . . . 30

List of Tables
5 Comparison with existing Causal Benchmarks/Datasets in NLP . . . . . . . . . . . 24
6 Human performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
7 Results for Causal and Causal Temporal Triplets . . . . . . . . . . . . . . . . . . . 32

List of Figures
4 Causal Graphical Model of Events. E1 temporally precedes E2 , and z is trajectory
variable, which assumes a values t where t ∈ All trajectories from start to E1 . . . 26
5 Input prompt formats for the MCQA-based evaluation of autoregressive open-weight
models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6 Qualitative examples for the MCQA-based evaluation of autoregressive open-weight
models (e.g., llama(-2), GPT-J, etc.). . . . . . . . . . . . . . . . . . . . . . . . 30
7 Input prompt formats for the ∆ estimation (temporal) via Language models. . . . . 30
8 Input prompt formats for computing the causal estimand using the autoregressive
open-weight models (using backdoor criterion) . . . . . . . . . . . . . . . . . . . 31
9 The “observational graph" for the activity Baking a Cake. . . . . . . . . . . . . 33
10 The “observational graph" for the activity Going Grocery Shopping. . . . . . . 34
11 The “observational graph" for the activity Going on a Train. . . . . . . . . . . 35
12 The “observational graph" for the activity Planting a Tree. . . . . . . . . . . . 36
13 The “observational graph" for the activity Riding on a Bus. . . . . . . . . . . . 37

22
A COLD Framework Details
The COLD framework consists of Observational Distributions represented in the form of DAGs (Go )
along with the corresponding causal graphs (Gc ) governing the dependency of occurrence between the
events. Table 2 highlights the total number of causal queries that can be created using the framework.
Table 1 shows a qualitative comparison with the COPA dataset Gordon et al. [2012] and the triplet
samples coming from the COLD framework.

A.1 Adherence to SUTVA

In causal literature, a fundamentally acknowledged Stable Unit Treatment Value Assumption (SUTVA)
[Cox, 1958, Rubin, 1980]) requires that for each unit (e.g., sequence of events), there is only one
version of the non-treatment, i.e., for an event in the sequence, there lie only two versions occurring
and not occurring. SUTVA plays a vital role in causal inference by ensuring that each unit’s treatment
assignment has a consistent impact, facilitating the accurate estimation of treatment effects. Although,
in the past, researchers have created various datasets that capture the causal relationship between
real-world events [Gordon et al., 2012, Du et al., 2022], the problem of achieving the SUTVA
assumption has remained challenging. For example, given events (taken from the COPA dataset
[Gordon et al., 2012]) E1 : “The teacher assigned homework to students” and E2 : “The
students groaned,” it becomes challenging to define ¬E1 since there are enormous possibilities
that may have occurred at the same time (in place of E1 ) that negates E1 , making it difficult to
define an event of not having done something. Recent work by Zhang et al. [2022b], proposes to use
multiple alterations of events for capturing ¬E1 , violating the SUTVA assumption. In this work, we
highlight that if we define a closed system, capturing a commonsense activity, it facilitates adherence
to SUTVA assumptions as closely as possible. For example, in the activity of “going via an
airplane,” one would have either “checked-in the luggage” (E1 ) or “skipped checking-in
luggage" due to smaller bags (¬E1 ). Moreover, developing a causal setup with observations has
always been a challenging problem in the wild and often requires few assumptions, as the strong
causal link can only be established in an ideal world where randomized controlled trials (RCTs) are
feasible. In our framework, adhering to SUTVA comes naturally where, in a trajectory, an occurrence
of an event can be intervened to obtain an alternate trajectory, reaching an ideal setup facilitating
causal reasoning in daily commonsensical activities.

A.2 Comparison with previous Causal Reasoning Datasets/Benchmarks

Table 5 shows a broad overview of the existing causal Dataset/Benchmarks presented in the NLP
community. We find that most of the existing set of work relies on real-world events to reason about
causality in NLP, where human annotators are asked to reason causally between the nature of events.
However, most of these datasets/benchmarks try to establish a connection using a simple question
prompt, which may not be enough to construct the underlying causal graph. Moreover, most of the
real-world grounding-based methods remain open-ended due to the events taking place in the wild,
making it difficult to consider constructing a causal graph where multiple variables play a role. More
recently, with increased research attention on the causal reasoning abilities of LLMs, researchers
have tried framing causal queries based on a causal inference engine, requiring the underlying causal
graphs. However, when constructing causal queries from prompting LLMs, natural language is used
to verbalize the causal concepts in the form of symbolic variables that may not have a real-world
grounding.2 Moreover, the created causal queries are difficult for a human with little or no knowledge
of causal inference concepts.3 Table 5 shows a comparison of all these features in detail, where
COLD satisfies all the features.
We realize this is a first-of-its-kind framework built over real-world events and contains the underlying
causal graph. Having both the Observational Distribution (representing the enormous event sequences
present in daily activity) and the manually created underlying causal graph helps facilitate an in-depth
analysis of the causal reasoning abilities of LLMs. Moreover, the same framework can further
be extended in various ways: 1) Extending the number of activities: In the current version of the
framework, we only consider 5 daily activities to provide an in-depth analysis. In the future, those
can be extended to incorporate more such activities. 2) Extending the scope of activities: The tasks
2
https://huggingface.co/datasets/causalnlp/corr2cause
3
https://huggingface.co/datasets/causalnlp/CLadder

23
Table 5: Comparison of causal experimental settings used in prior LLM evaluation benchmarks. The
real-world grounding plays a crucial role in evaluating LLMs, which is not present in the symbolic benchmarks.
Datasets/Benchmarks Real-World Causal Graph Symbolic Exhaustive # Samples
SemEval2021 Task8 [Hendrickx et al., 2010] ✓ ✗ ✗ ✗ 1331
EventCausality [Do et al., 2011] ✓ ✗ ✗ ✗ 414
COPA [Gordon et al., 2012] ✓ ✗ ✗ ✗ 1000
Causal-TimeBank [Mirza et al., 2014] ✓ ✗ ✗ ✗ -
CaTeRS [Mostafazadeh et al., 2016] ✓ ✗ ✗ ✗ 320 stories (1.6K sent)
BECauSE [Dunietz et al., 2017] ✓ ✗ ✗ ✗ -
Event2Mind [Rashkin et al., 2018] ✓ ✗ ✗ ✗ 25K event phrases
ATOMIC [Sap et al., 2019a] ✓ ✗ ✗ ✗ 877K
SocialIQA [Sap et al., 2019b] ✓ ✗ ✗ ✗ 37K
TimeTravel [Qin et al., 2019] ✓ ✗ ✗ ✗ 81.4K
Abductive (ART) [Bhagavatula et al., 2020] ✓ ✗ ✗ ✗ 20K narr, 200K expl
Com2Sense [Singh et al., 2021] ✓ ✗ ✗ ✗ 4k sentence pairs.
TellMeWhy [Lal et al., 2021] ✓ ✗ ✗ ✗ 30K questions
CRASS [Frohberg and Binder, 2022] ✓ ✗ ✗ ✗ 274 PCT
e-CARE [Du et al., 2022] ✓ ✗ ✗ ✗ 20K CR questions
CausalQA [Bondarenko et al., 2022] ✓ ✗ ✗ ✗ 1.1 Million
COLA [Wang et al., 2023] ✓ ✗ ✗ ✗ 1,360 event pairs
CRAB [Romanou et al., 2023] ✓ ✗ ✗ ✗ 2.7K pairs
CausalJudgement [Srivastava et al., 2023] ✓ ✓ ✗ ✗ -
Corr2Cause [Jin et al., 2023] ✗ ✓ ✓ ✓ 200K
CausalParrots [Zečević et al., 2023] ✓ ✓ ✓ ✗ -
CLadder [Jin et al., 2024] ✗ ✓ ✓ ✗ 10K
COLD (ours) ✓ ✓ ✓ ✓ ∼9.52 Million

used in the activities are generic and capture commonsense; for validating domain-specific causal
reasoning abilities, the framework could be extended to domain activities, for example, cooking a
specific recipe where adding different ingredients causes a variation in taste. 3) Extending the type of
causal queries: while constructing the causal queries, we considered the simplest task of finding the
more plausible cause/effect given two options as events, keeping only unconditional d−separation as
the primary condition. The framework can directly be extended, keeping causal queries inspired from
Jin et al. [2023, 2024].
On the analysis front, we realize that the possibilities of an in-depth analysis increase by a significant
margin. In this work, we shed light on a few mechanisms for validating causal reasoning abilities
via zero-shot CCR (compared to previous works that rely on training and further testing on similar
datasets). We specifically focused on open-weight models for better applicability in the future and
proposed a few mechanisms for estimating the causal relationships between the events. This opens
up several possible avenues for an in-depth analysis of LLMs.

A.3 Observational Graphs

Fig. 9, Fig. 10, Fig. 11, Fig. 12, and Fig. 13 shows the “observational graphs" for the activity
Baking a Cake, Going Grocery Shopping, Going on a Train, Planting a Tree, and Riding
on a Bus respectively.

B Algorithms in the COLD Framework

In this section, we provide insights into the Algorithms used in the COLD framework. We start with
Algorithm 1, which creates causal query triplets given the observational graphs Go and along with the
Causal Graphs Gc .
Remark: Temporal precedence is generally assumed essential for defining causation, and it is one
of the most important clues that is used to distinguish causal from other types of associations [Mill,
1898, Hill, 1965, Pearl and Verma, 1995]. For our framework, we also consider the topologically
sorted order obtained from the observational graphs and use the temporal order to define the causal
query triplets, i.e., the cause events will always precede the effect events.
Creating Causal Query Triplets: The Algorithm 1 is designed to sample all the possible causal
query triplets to construct a dataset for validating causal reasoning ability over an activity. Provided
the observational graphs Go and the Causal Graphs Gc for an activity, we first sample all the possible
node triplets in the graph. Later, we iterate over the set of triplets and check if one of the nodes in

24
Table 6: Human validation done for a small sample of 100 causal query triplets. Overall we find
that humans do perform well in causal reasoning about these daily activities.
Human Annotators cake shopping train tree bus Average
Subject 1 95 95 90 100 90 94
Subject 2 100 100 90 95 90 95
Subject 3 100 100 85 85 70 88
Subject 4 100 100 95 90 85 94
Subject 5 100 85 95 90 80 90
Average 99.00 96.00 91.00 92.00 83.00 92.20

the triplet (ni , nj , nk ) is d-separated. Further, the d-separated node becomes the wrong choice. The
remaining two events become premise and correct choices, depending on their temporal order. For
example, if ni is the node that is d-separated from nj and nk , we check if nj and nk have a causal
link between them in Go . If nj and nk are found to have a causal link, we create two triplets using
the temporal ordering between nj and nk . The temporal link nj → nk leads to an ‘effect’ query
where nj becomes ‘premise’ and nk becomes ‘correct choice,’ and a ‘cause’ query where nj becomes
‘correct choice’ and nk becomes the ‘premise.’ Note in Algorithm 1 [Store tuple], we only show one
such instance for brevity; the implementation will consist of another mirror instance (i.e., for every
[Store tuple] both ‘cause’ and ‘effect’ question triplets are stored to the dataset).
The understanding of these activities is generic, and LLMs do provide a suitable set of sequences
when prompted to generate a list of steps to complete the activity. The constructed dataset helps
to rigorously validate the understanding of the activity through an enormous number of causal
query triplets. The results show that although the LLMs can explain the activity in detail, including
generating correct steps for performing tasks, causally reasoning about the set of events remains
challenging.
Human validation: To get a rough estimate of the human performance on the created causal
reasoning queries, we also perform a small-scale human study, where the annotators are given a set of
randomly chosen 100 causal queries. The human subjects were graduate students of computer science
who were given a brief tutorial about counterfactual reasoning. Table 6 shows the obtained results. We
would like to mention that validating human performance is challenging due to the nature of the causal
reasoning task. The nature of counterfactual reasoning requires the human/algorithm to assume a
similar alternate world/universe with only a particular happening or not happening to approximate the
causal strength between the events. These imaginations can be expressed in statements as highlighted
by Pearl and Mackenzie [2018], Pearl et al. [2016], containing an “if” statement in which the “if”
portion is untrue or unrealized (aka counterfactual). The “if” portion of a counterfactual is called the
hypothetical condition, or more often, the antecedent, making it challenging (cognitively heavy) to
conduct a human evaluation. Please note that the study is performed using only a small sample of
100 causal query triplets out of thousands of queries, and the presented results only provide a rough
estimate that may not generalize for a larger number of queries. Hence, a comparison of human
study results and LLMs is not fair, and the presented human performance estimates may not be true
representative of the entire population. Interactions with human subjects also revealed that they tend
to confuse temporality and causality (similar findings were reported by Do et al. [2011]).
Evaluating over Causal Query Triplets: Algorithm 2 makes use of the causal estimands ∆ to
compare the causal strength between the premise event and the choice events. We consider the
causal estimand computed between the premise and the available set of choices and predict the label
corresponding to the high ∆ values. For a given causal query from the created causal query triplet
dataset D, where each data entry Di corresponds to (p, c1 , c2 , q, l), i.e., premise event, choice 1,
choice 2, question and the label respectively. As the task is to predict the more plausible cause/effect
of the given premise event, we create two event pairs, (p, c1 ) and (p, c2 ), and compute the causal
estimand ∆ for both the pairs using the temporal or the backdoor scheme (described below in
Algorithm 3). Note that the order of events given to ∆M is in E1 and E2 format, i.e. ∆M (E1 , E2 ).
Using the temporal precedence (highlighted as remark above), the cause event will always precede the
effect event temporally. Hence, for a causal query with the question as ‘cause’, the causal estimand
is estimated as ∆M (ci1 , pi ), ∆M (ci2 , pi ) and ∆M (pi , ci1 ), ∆M (pi , ci2 ) when the question is ‘effect.’
Further, based on the estimated ∆M scores, the more plausible cause/effect is predicted.

25
Algorithm 1 Creating Causal Query Triplets
Gc : Causal Graphical Model; Go : Observational Graph;
ds (G, x, y): True iff (x, y) are d-separated unconditionally in any DAG G;
dc (G, x, y): True iff (x, y) are d-connected unconditionally in any DAG G;
genSamples(G): Generates all possible node triplets (ni , nj , nk ) from DAG G such that i ≤ j ≤ k,
where (i, j, k) are the respective indices of nodes in a topologically sorted list of nodes;
A(G, x, y): True iff x is an ancestor of y in DAG G.
⊕: Exclusive OR operator;
p: premise; c1 : choice 1; c2 : choice 2; q: question; l: answer (label)
(p, c1 , c2 , q, l) ∈ D
D=[ ] [Empty Dataset]
S ∼ genSamples(Go ) [Generate Samples]
for (ni , nj , nk ) in S
if ds (Gc , ni , nj ) then
if dc (Gc , ni , nk ) ⊕ dc (Gc , nj , nk )
q = ‘cause’
l = arg max[dc (Gc , ni , nk ), dc (Gc , nj , nk )]
p, c1 , c2 ← nk , ni , nj
if A(Gc , cl , p) then APPEND(D, [(p, c1 , c2 , q, l)) [Store tuple]
else if ds (Gc , nj , nk ) then
if dc (Gc , ni , nj ) ⊕ dc (Gc , ni , nk )
q = ‘effect’
l = arg max[dc (Gc , ni , nk ), dc (Gc , ni , nj )]
p, c1 , c2 ← ni , nj , nk
if A(Gc , p, cl ) then APPEND(D, (p, c1 , c2 , q, l)) [Store tuple]
return D

Computing ∆M : Algorithm 3 depicts the process of computing an unbiased estimate for the causal
estimand. The causal strength is computed between two events E1 and E2 where E1 is assumed to be
preceding E2 temporally. To make an unbiased estimate based on the provided options, we consider
normalizing the obtained probability scores by flipping the options and providing the same query
prompt to the Language Model.
sM (E1 , E2 , ϕ) + sM (E1 , E2 , ϕf )
fM (E1 , E2 , ϕ) ← ,
sM (E1 , E2 , ϕ) + sM (E1 , E2 , ϕf ) + s̃M (E1 , E2 , ϕ) + s̃M (E1 , E2 , ϕf )
where ϕ denotes the prompt template as shown in Figure 8 (top) and ϕf denotes the same prompt with
flipped options, Figure 8 (bottom). The overall equation helps normalize the prediction probabilities
of the ‘Increase’ option by using the probabilities of the ‘Decrease’ option. Finally, these normalized
scores are computed for multiple trajectories ti in the backdoor adjustment scheme to compute the
causal estimands pM (E2 | do(E1 )) and pM (E2 | do(¬E1 )) that help estimate the causal strength
∆M between the events E1 and E2 .

E1 E2

Figure 4: Causal Graphical Model of Events. E1 temporally precedes E2 , and z is trajectory variable,
which assumes a values t where t ∈ All trajectories from start to E1

C Backdoor Adjustments
A set of variables W satisfies the backdoor criterion relative to T and Y if the following are true

(A) W blocks all backdoor paths from T to Y i.e. blocking confounding or non-causation
association paths
(B) W doesn’t contain any descendants of T

26
Algorithm 2 Evaluating Causal Query Triplets
Tn : n unique trajectories from Start node (start of the activity) to node E1
1 X
pM (E2 |do(E1 )) = p̂(E2 |E1 , z = t)
|Tn | t∈T
n
1 X
pM (E2 |do(¬E1 )) = p̂(E2 |¬E1 , z = t)
|Tn | t∈T
n
∆M : Returns Average treatment effect (pM (E2 |do(E1 )) − pM (E2 |do(¬E1 ))) to determine the causal
effect of event E1 on event E2
p: premise; c1 : choice 1; c2 : choice 2; q: question; l: label
D: Set of all causal queries
Di : Causal query (p, c1 , c2 , q, l), where (p, c1 , c2 , q, l) ∈ D
for Di in D do
if q i == ‘cause’ then
prediction ← arg max ∆M (ci1 , pi ), ∆M (ci2 , pi )
 
i
else if q == ‘effect’ then 
prediction ← arg max ∆M (pi , ci1 ), ∆M (pi , ci2 )

end if
η ← η + 1(prediction == li )
end for
return η/|D|

Algorithm 3 Computing Causal Estimand

E1 , E2 : Events in an given activity
Tn : Set of n trajectories (temporally ordered sequence of events)
sM (E1 , E2 , ϕ) ← score of token A (associated with option ‘increase’) using prompt ϕ of model M.
sM (E1 , E2 , ϕf ) ← score of token B (associated with option ‘increase’) in prompt ϕf (flipped options in
prompt ϕ) of model M.
s̃M (E1 , E2 , ϕ) ← score of token B (associated with option ‘decrease’) using prompt ϕ of model M.
s̃M (E1 , E2 , ϕf ) ← score of token A (associated with option ‘decrease’) in prompt ϕf (flipped options in
prompt ϕ)of model M
sM (E1 , E2 , ϕ) + sM (E1 , E2 , ϕf )
Norm. Score fM (E1 , E2 , ϕ) ←
sM (E1 , E2 , ϕ) + sM (E1 , E2 , ϕf ) + s̃M (E1 , E2 , ϕ) + s̃M (E1 , E2 , ϕf )
p̂(E2 |E1 , z = t) ← fM ((t, E1 ), E2 , ϕ)
p̂(E2 |¬E1 , z = t) ← fMP ((t, ¬E1 ), E2 , ϕ)
pM (E2 |do(E1 ))← |T1n | t∈Tn p̂(E2 |E1 , z = t)
pM (E2 |do(¬E1 ))← |T1n | t∈Tn p̂(E2 |¬E1 , z = t)
P

∆M ← pM (E2 |do(E1 )) − pM (E2 |do(¬E1 ))

return ∆M

Then, W satisfies the backdoor criterion [Pearl et al., 2016, Neal, 2020].
Adhering to the above conditions of the backdoor criterion, it is reasonable to assume that the
trajectory t (temporally ordered sequence of events) till E1 will contain the events that confound
the event E1 and event E2 (condition A). Every event trajectory till E1 will temporally precede
E1 (condition B). Hence, the trajectory variable will satisfy backdoor criteria in the proposed closed
system. The domain of the trajectory variable is a set of all trajectories till E1 . Therefore, conditioning
on t closes all paths that induce non-causal associations. The generic representation of an approximate
causal graphical model involving E1 , E2 , and t is shown in Figure 4, and can be formulated as:

1 X
pM (E2 |do(E1 )) = p̂(E2 |E1 , z = t)
|T |
t∈T

Where T is a set of all trajectories from the start of the activity till the event E1 (excluding E1 ).

27
D Experiments and Results
D.1 Compute Resources

We perform all the experiments using a machine with 5 NVIDIA A100 GPUs. We use only the
open-weight models with frozen parameters to present the results for better reproducibility in the
future.

D.2 Evaluation using Average Treatment Effect (ATE)

Establishing Causal Relationships: To validate the causal reasoning ability, the MCQA-based
approach can be further extended to estimate the causal estimation and denote the causal strengths
between the events. Establishing cause-and-effect relationships can be achieved through various
statistical analyses. The strength of cause-and-effect relationships is approximated by statistically
analyzing events’ behavior using observational data (PC-Algorithm, Spirtes et al. [2000b]). Moreover,
some of the recent works [Wang et al., 2023] highlight the role of context in determining the causal
relationships between the events. To extend our analysis of causal reasoning abilities in the proposed
framework, we use the backdoor adjustments in LLMs as explained in the main paper. Moreover, we
also perform an interesting analysis of the observational graphs for estimating ∆ statistically.
1) Through Original Trajectories: DeScript Wanzare et al. [2016] collects data by considering
∼ 100 ESDs written by different crowd-sourced workers. We use the original Trajectories (ESDs)
To written by humans present in the DeScript dataset. These ESDs provide the original flow in the
graph directly coming from crowdsourced workers. We consider these as the original trajectories
To . Applying the backdoor criterion (Eq. 2) over these trajectories To . An interventional distribution
similar to the previous section is computed considering the likelihood of occurrence of E2 under each
treatment (E1 and ¬E1 ) for only these trajectories To . These estimations are further used to compute
the treatment effect using the Eq. 1. We denote the causal risk difference (∆) computed with To as
∆o .
2) Through Observational Graphs: The observational graphs provide a proxy for the underlying
knowledge about the activity, covering all possible sets of events, i.e., starting from the start node,
one can trace multiple trajectories that will essentially define the way of performing the activity. For
every pair of connected events (ei , ej ), the edge between them represents the probable transition from
ei to ej with some non-zero probability. However, a noteworthy point is that the transition probability
between two connected events (ei , ej ) can vary depending on the design choices/transition function
T (ei , ej ) → (0, 1]. We define this transition function in two ways: 1) Uniform Node Transition
(Tn ): The transition probability from current node ei to next probable events ej ∈ Eij would be
uniform that is T (ei , ej ) = |o1i | , where |oi | represents number of outgoing edges from event ei . (i.e.,
assuming after an event, the choice of the next event is uniform from the possible events). 2) Uniform
Trajectory Transition (Tt ): Another way to take the set of events in an activity (trajectory) is by
considering all the possible paths being equally probable, i.e., across the entire population the same
activity will be represented with one of the possible trajectories. Hence we can define the transition
function with each trajectory ti = (estart , e2 , ..., eend ) (sequence of events from starting to ending)
having the same probability, i.e.:
p(ti ) = p(tj )∀ti , tj ∈ T
Y
p(ti ) = Tt (el → em )
l,m∈ti

Further, given a transition function T , computation of p̂(E2 |E1 , z = ti ) becomes straightforward

as p(E2 |E1 ), since the next course of trajectories after E1 will be decided given E1 has occurred.
Analytically, it can be computed by counting every trajectory (i.e., Tij ) from E1 that leads to E2
X Y
p(E2 |E1 ) = T (el → em ) (4)
t∈Tij l,m∈t
"k=M #
X
k
p(E2 |E1 ) = T (E1 → E2 ) (5)
k=1

28
For estimating the probability p̂(E2 |¬E1 , z = ti ), we make use of the observational graph by
considering all the parent nodes of E1 and compute the probability of reaching E2 from the parent
(i.e., last event of trajectory ti ) avoiding the occurrence of E1 .
X Y
p̂(E2 |¬E1 , z = ti ) = T (el → em ) (6)
t∈Tij l,m∈t
E1 ∈t
/

also, p(z = ti ) can be computed as product of each transition in ti , i.e

Y
p(z = ti ) = T (ei → ej ) (7)
ij

Computations from Equations 5, 6, and 7 are used in the backdoor adjustment defined in Equation 2
for estimating the interventional likelihood of occurrence of E2 under each treatment (E1 and ¬E1 )
and causal risk difference (∆). Note depending on the choice of transition function (Tn or Tt ), we
obtain two deltas ∆n and ∆t , respectively.

E Prompt Templates for Language Model based Experiments

We present the various prompt templates used to estimate the temporal link between the events in
Figure 7. For BERT-based models, we use the MLM-trained models for predicting the masked token
given a sentence (Previously, a similar approach was adopted by Zhang et al. [2022b]). In contrast,
for autoregressive models, we frame the prompt as a question-answer objective, taking inspiration
from [Robinson and Wingate, 2023], where a multiple-choice-based question is framed to predict
the answer in the form of the option IDs. The prompt is intentionally structured so that the LLM is
intended to predict a single token (Such as “A”, “B”, etc.). Robinson and Wingate [2023] highlights
the advantages of MCQA-based evaluation over cloze evaluation (where the LLMs are expected to
generate the entire answer in a cloze test), leading to a significant boost in various tasks, including
commonsense-based tasks.
For our prompt-based evaluation experiments over the generated causal triplets, we follow the same
MCQA-based strategy and frame the prompts accordingly for a fair evaluation. Figure 5 presents
various prompt templates for autoregressive experiments, and Figure 6 shows a few qualitative
examples for the framed causal query templates.

Consider the activity of activity name.

[ in-context examples (if few-shot/in-context learning experiment) ]
Which of the following events (given as options A or B) is a plausible question
(cause/effect) of the event premise?
A. choice1
B. choice2
Answer: A
The following are multiple choice questions about activity name. You should directly
answer the question by choosing the correct option.
[ in-context examples (if few-shot/in-context learning experiment) ]
Which of the following events (given as options A or B) is a plausible question
(cause/effect) of the event premise?
A. choice1
B. choice2
Answer: A

Figure 5: Input prompt formats for the MCQA-based evaluation of autoregressive open-weight
models (e.g., llama(-2), GPT-J, etc.). The black text is the templated input. The orange text is
the input from the created causal query triplets, where the activity name denotes the description
of the activity like baking a cake. The next-token prediction probabilities of the option IDs at the
red text is used as the observed prediction distribution.

29
 Consider the activity of baking a cake.
Which of the following events (given as options A or B) is a plausible effect of the event
preheat oven to 350 degrees.?
A. turn off oven.
B. prepare the microwave oven and required utensils
Answer: A
The following are multiple choice questions about going on a train. You should directly
answer the question by choosing the correct option.
Which of the following events (given as options A or B) is a plausible cause of the event
get the bill for groceries. ?
A. pay the cashier for your items.
B. place cart into cart corral.
Answer: A

Figure 6: Qualitative examples for the MCQA-based evaluation of autoregressive open-weight

models (e.g., llama(-2), GPT-J, etc.).

In terms of ’before’ and ’after’, the event: “first event text" would have happened
<mask_token> the event: “second event text"
Consider the activity of activity name.
Question: Determine the temporal order.
The following events took place: 1. first event text, 2. second event text
Did the first event occur ’before’ or ’after’ the second event? (choose from the given
options)
A: before
B: after
Answer: A

Figure 7: Input prompt formats for the ∆ estimation via Language models. The first row shows the
prompt template used for BERT-based language models, where the mask token is predicted. The
second row shows the template for autoregressive open-weight models (e.g., llama(-2), GPT-J,
etc.). The black text is the templated input. The orange text is the input from the created causal
query triplets, where the first event text and second event text comes from the premise and
available set of choices. The mask-token prediction probabilities of ‘before’ and ‘after’ and next-
token prediction probabilities of the option IDs at the red text are used as the observed prediction
distribution for BERT-based and GPT-based open-weight models.

F Additional Results
F.1 Temporally Plausible Choices in Causal Triplets

Some of the initial studies [Do et al., 2011] highlight the difficulty in choosing between the causal-
effect events and temporal events (that occur in close proximity to the premise event), i.e., temporal
relationships are sometimes considered as a causal relationship by human annotators. We also create
another version of created causal triplets where the wrong choices are replaced by temporally near
nodes (nodes that are at a one-hop distance from the premise node). We call these ‘causally hard
triplets.’ Note the temporal nodes are obtained from the observational graphs Go . Table 7 shows
the performance comparison with causal triplets and causal-temporal triplets versions of the same
queries. We observe a significant performance drop on the causal-temporal triplets version for most
models, highlighting the increased confusion.

30
 CAUSAL REASONING ANALYSIS:
Context: For the activity activity name. During the activity, the following set of
sequences occurred in order:
[ ordered list of events present in a Trajectory till E1 ] # Trajectory (z = Ti )
Further, the event ‘event text for E1 ’ took place. # Intervention (do(E1 ))
Question: Given the above information, will the chances of the occurrence of the event
‘event text for E2 ’ increase or decrease?
A. Increase
B. Decrease
Answer: A # (p(E2 | do(E1 ), z = Ti )
CAUSAL REASONING ANALYSIS:
Context: For the activity activity name. During the activity, the following set of
sequences occurred in order:
[ ordered list of events present in a Trajectory till E1 ] # Trajectory (Ti )
Further, the event ‘event text for E1 ’ did NOT take place. # Intervention (do(¬E1 ))
Given the above information, will the chances of the occurrence of the event ‘event text
for E2 ’ increase or decrease?
A. Increase
B. Decrease
Answer: B # (p(E2 | do(¬E1 ), z = Ti )
Flipped options variant of the above Prompt Template
CAUSAL REASONING ANALYSIS:
Context: For the activity activity name. During the activity, the following set of
sequences occurred in order:
[ ordered list of events present in a Trajectory till E1 ] # Trajectory (z = Ti )
Further, the event ‘event text for E1 ’ took place. # Intervention (do(E1 ))
Given the above information, will the chances of the occurrence of the event ‘event text
for E2 ’ increase or decrease?
A. Decrease
B. Increase
Answer: B # (p(E2 | do(E1 ), z = Ti )
CAUSAL REASONING ANALYSIS:
Context: For the activity activity name. During the activity, the following set of
sequences occurred in order:
[ ordered list of events present in a Trajectory till E1 ] # Trajectory (Ti )
Further, the event ‘event text for E1 ’ did NOT take place. # Intervention (do(¬E1 ))
Given the above information, will the chances of the occurrence of the event ‘event text
for E2 ’ increase or decrease?
A. Decrease
B. Increase
Answer: A # (p(E2 | do(¬E1 ), z = Ti )

Figure 8: Input prompt formats for computing the causal estimand using the autoregressive open-
weight models (using backdoor criterion) (e.g., llama(-2), GPT-J, etc.). The black text is the
templated input. The orange text is the input from the created causal query triplets, where the
activity name denotes the description of the activity like baking a cake. The trajectory Ti is
obtained using the observational graph Go and contains the sequence of events before the event E1 .
The next-token prediction probabilities of the option IDs at the red text is used as the observed
prediction distribution. The flipped options variants of the prompts contain the same query with
flipped options (i.e., the option ‘Increase’ becomes ‘Decrease’ and vice versa). This is done to
make the causal estimand unbiased towards the predicted option token as highlighted in Algorithm 3.

31
Table 7: The table provides evaluation results of Language models over the causal and causal temporal
triplets.
Triplets Model Name cake shopping train tree bus
gpt-neo-125M 50.71 50.01 49.99 50.13 50.15
gpt-neo-1.3B 44.77 45.69 42.52 45.67 42.89
gemma-2b 53.76 52.19 60.57 60.71 53.64
causal triplets

gpt-neo-2.7B 50.00 50.01 50.00 50.01 50.00

phi-2 85.14 83.65 77.29 82.24 71.74
gpt-j-6B 49.59 50.02 50.29 49.92 49.93
Llama-2-7b-chat-hf 77.92 72.41 73.48 72.40 68.21
Mistral-7B-v0.1 77.64 69.38 68.46 72.43 69.37
gemma-7b 81.47 82.26 77.24 80.78 70.29
Meta-Llama-3-8B 80.79 76.46 76.08 78.21 67.39
gpt-neo-125M 50.60 49.80 49.90 50.00 50.20
gpt-neo-1.3B 49.50 51.20 48.80 47.50 48.00
causally hard triplets

gemma-2b 52.30 51.00 56.10 52.20 50.00

gpt-neo-2.7B 50.00 50.00 50.00 50.00 50.00
phi-2 80.00 74.70 67.90 87.50 66.50
gpt-j-6B 50.20 50.00 50.30 50.00 49.80
Llama-2-7b-chat-hf 71.60 66.60 68.30 77.00 65.40
Mistral-7B-v0.1 69.20 63.10 64.20 67.90 62.00
gemma-7b 76.30 76.40 69.70 89.70 63.70
Meta-Llama-3-8B 77.30 72.20 69.30 83.20 64.30

32
 choose-recipe

buy-ingredients

other-return-home

other-enter-kitchen other-read-instructions

get-ingredients

get-utensils

get-measuring-utensils

preheat

measure-ingredients

add-ingredients other-make-cake

prepare-ingredients

grease-cake-tin

pour-dough

put-cake-oven

set-time

other-clean

wait

other-return-oven

check

other-bake-extra-time turn-off-oven

take-out-oven

cool-down

take-out-cake-tin

decorate

eat

End Task

Figure 9: The figure shows the “observational graph" for the activity Baking a Cake.

33
 make-list

take-bags other-turn-car-on other-carry-list other-get-in-car

go-grocery-drive

other-park-car go-grocery-walk

go-grocery-bus get-list-from-car

enter

take-shop-cart

other-shop move-section

get-groceries

check-list

go-checkout other-checkout

wait

put-conveyor other-present-coupons

cashier-scan/weight

other-get-total

other-get-wallet

pay

pack-groceries

get-receipt other-put-bags-cart other-take-bags

bring-vehicle

other-place-cart-into-cart-corral

leave

set-items-in-kitchen

End Task

Figure 10: The figure shows the “observational graph" for the activity Going Grocery Shopping.

34
 decide-where-to-go

check-time-table

other-pack other-get-in-car

get-train-station

other-park-car

enter-station other-get-out-car

other-go-ticket-counter

other-wait-queue

get-tickets

other-find-platform

get-platform other-walk

wait

train-arrives other-wait-to-board

get-on

other-stow-luggage

find-place

conductor-checks

spend-time-train

arrive-destination

get-off

other-take-luggage

other-leave-station

End Task

Figure 11: The figure shows the “observational graph" for the activity Going on a Train.

35
 other-go-garden-center choose-tree

get-tree

take-home

find-place

get-tools

other-put-on-gloves

dug-hole

other-secure-tree place-fertilizers other-check-hole

unwrap-root other-score-roots other-get-soil

place-root

other-hold-tree refill-hole

other-check-stability tamp-dirt

other-put-mulch other-put-away-tools

water

tie-stakes-up

End Task

Figure 12: The figure shows the “observational graph" for the activity Planting a Tree.

36
 other-find-cost

check-time-table

find-bus other-decide

get-bus-stop

wait

other-signal-driver

bus-comes

other-driver-opens-door

board-bus

get-ticket

find-place

ride

spend-time-bus

press-stop-button

go-exit

other-reach

other-thank bus-stops

get-off

End Task

Figure 13: The figure shows the “observational graph" for the activity Riding on a Bus.

37
NeurIPS Paper Checklist
1. Claims
Question: Do the main claims made in the abstract and introduction accurately reflect the
paper’s contributions and scope?
Answer: [Yes]
Justification: We provide details about the main claims in the Abstract and Introduction
(Section 1).
Guidelines:
• The answer NA means that the abstract and introduction do not include the claims
made in the paper.
• The abstract and/or introduction should clearly state the claims made, including the
contributions made in the paper and important assumptions and limitations. A No or
NA answer to this question will not be perceived well by the reviewers.
• The claims made should match theoretical and experimental results, and reflect how
much the results can be expected to generalize to other settings.
• It is fine to include aspirational goals as motivation as long as it is clear that these goals
are not attained by the paper.
2. Limitations
Question: Does the paper discuss the limitations of the work performed by the authors?
Answer: [Yes]
Justification: We provide a separate section on Limitations (Sections 6)
Guidelines:
• The answer NA means that the paper has no limitation while the answer No means that
the paper has limitations, but those are not discussed in the paper.
• The authors are encouraged to create a separate "Limitations" section in their paper.
• The paper should point out any strong assumptions and how robust the results are to
violations of these assumptions (e.g., independence assumptions, noiseless settings,
model well-specification, asymptotic approximations only holding locally). The authors
should reflect on how these assumptions might be violated in practice and what the
implications would be.
• The authors should reflect on the scope of the claims made, e.g., if the approach was
only tested on a few datasets or with a few runs. In general, empirical results often
depend on implicit assumptions, which should be articulated.
• The authors should reflect on the factors that influence the performance of the approach.
For example, a facial recognition algorithm may perform poorly when image resolution
is low or images are taken in low lighting. Or a speech-to-text system might not be
used reliably to provide closed captions for online lectures because it fails to handle
technical jargon.
• The authors should discuss the computational efficiency of the proposed algorithms
and how they scale with dataset size.
• If applicable, the authors should discuss possible limitations of their approach to
address problems of privacy and fairness.
• While the authors might fear that complete honesty about limitations might be used by
reviewers as grounds for rejection, a worse outcome might be that reviewers discover
limitations that aren’t acknowledged in the paper. The authors should use their best
judgment and recognize that individual actions in favor of transparency play an impor-
tant role in developing norms that preserve the integrity of the community. Reviewers
will be specifically instructed to not penalize honesty concerning limitations.
3. Theory Assumptions and Proofs
Question: For each theoretical result, does the paper provide the full set of assumptions and
a complete (and correct) proof?
Answer: [NA]

38
 Justification: The paper does not have any theoretical results.
Guidelines:
• The answer NA means that the paper does not include theoretical results.
• All the theorems, formulas, and proofs in the paper should be numbered and cross-
referenced.
• All assumptions should be clearly stated or referenced in the statement of any theorems.
• The proofs can either appear in the main paper or the supplemental material, but if
they appear in the supplemental material, the authors are encouraged to provide a short
proof sketch to provide intuition.
• Inversely, any informal proof provided in the core of the paper should be complemented
by formal proofs provided in appendix or supplemental material.
• Theorems and Lemmas that the proof relies upon should be properly referenced.
4. Experimental Result Reproducibility
Question: Does the paper fully disclose all the information needed to reproduce the main ex-
perimental results of the paper to the extent that it affects the main claims and/or conclusions
of the paper (regardless of whether the code and data are provided or not)?
Answer: [Yes]
Justification: We provide details in Section 4.
Guidelines:
• The answer NA means that the paper does not include experiments.
• If the paper includes experiments, a No answer to this question will not be perceived
well by the reviewers: Making the paper reproducible is important, regardless of
whether the code and data are provided or not.
• If the contribution is a dataset and/or model, the authors should describe the steps taken
to make their results reproducible or verifiable.
• Depending on the contribution, reproducibility can be accomplished in various ways.
For example, if the contribution is a novel architecture, describing the architecture fully
might suffice, or if the contribution is a specific model and empirical evaluation, it may
be necessary to either make it possible for others to replicate the model with the same
dataset, or provide access to the model. In general. releasing code and data is often
one good way to accomplish this, but reproducibility can also be provided via detailed
instructions for how to replicate the results, access to a hosted model (e.g., in the case
of a large language model), releasing of a model checkpoint, or other means that are
appropriate to the research performed.
• While NeurIPS does not require releasing code, the conference does require all submis-
sions to provide some reasonable avenue for reproducibility, which may depend on the
nature of the contribution. For example
(a) If the contribution is primarily a new algorithm, the paper should make it clear how
to reproduce that algorithm.
(b) If the contribution is primarily a new model architecture, the paper should describe
the architecture clearly and fully.
(c) If the contribution is a new model (e.g., a large language model), then there should
either be a way to access this model for reproducing the results or a way to reproduce
the model (e.g., with an open-source dataset or instructions for how to construct
the dataset).
(d) We recognize that reproducibility may be tricky in some cases, in which case
authors are welcome to describe the particular way they provide for reproducibility.
In the case of closed-source models, it may be that access to the model is limited in
some way (e.g., to registered users), but it should be possible for other researchers
to have some path to reproducing or verifying the results.
5. Open access to data and code
Question: Does the paper provide open access to the data and code, with sufficient instruc-
tions to faithfully reproduce the main experimental results, as described in supplemental
material?

39
 Answer: [Yes]
Justification: We provide details in the Introduction.
Guidelines:
• The answer NA means that paper does not include experiments requiring code.
• Please see the NeurIPS code and data submission guidelines (https://nips.cc/
public/guides/CodeSubmissionPolicy) for more details.
• While we encourage the release of code and data, we understand that this might not be
possible, so “No” is an acceptable answer. Papers cannot be rejected simply for not
including code, unless this is central to the contribution (e.g., for a new open-source
benchmark).
• The instructions should contain the exact command and environment needed to run to
reproduce the results. See the NeurIPS code and data submission guidelines (https:
//nips.cc/public/guides/CodeSubmissionPolicy) for more details.
• The authors should provide instructions on data access and preparation, including how
to access the raw data, preprocessed data, intermediate data, and generated data, etc.
• The authors should provide scripts to reproduce all experimental results for the new
proposed method and baselines. If only a subset of experiments are reproducible, they
should state which ones are omitted from the script and why.
• At submission time, to preserve anonymity, the authors should release anonymized
versions (if applicable).
• Providing as much information as possible in supplemental material (appended to the
paper) is recommended, but including URLs to data and code is permitted.
6. Experimental Setting/Details
Question: Does the paper specify all the training and test details (e.g., data splits, hyper-
parameters, how they were chosen, type of optimizer, etc.) necessary to understand the
results?
Answer: [NA]
Justification: As discussed in Section 4, we only perform evaluation on pre-trained models
and do not train/fine-tune any new model.
Guidelines:
• The answer NA means that the paper does not include experiments.
• The experimental setting should be presented in the core of the paper to a level of detail
that is necessary to appreciate the results and make sense of them.
• The full details can be provided either with the code, in appendix, or as supplemental
material.
7. Experiment Statistical Significance
Question: Does the paper report error bars suitably and correctly defined or other appropriate
information about the statistical significance of the experiments?
Answer: [NA]
Justification: As discussed in Section 4, we only perform evaluation on pre-trained models
and do not train/fine-tune any new model.
Guidelines:
• The answer NA means that the paper does not include experiments.
• The authors should answer "Yes" if the results are accompanied by error bars, confi-
dence intervals, or statistical significance tests, at least for the experiments that support
the main claims of the paper.
• The factors of variability that the error bars are capturing should be clearly stated (for
example, train/test split, initialization, random drawing of some parameter, or overall
run with given experimental conditions).
• The method for calculating the error bars should be explained (closed form formula,
call to a library function, bootstrap, etc.)
• The assumptions made should be given (e.g., Normally distributed errors).

40
 • It should be clear whether the error bar is the standard deviation or the standard error
of the mean.
• It is OK to report 1-sigma error bars, but one should state it. The authors should
preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis
of Normality of errors is not verified.
• For asymmetric distributions, the authors should be careful not to show in tables or
figures symmetric error bars that would yield results that are out of range (e.g. negative
error rates).
• If error bars are reported in tables or plots, The authors should explain in the text how
they were calculated and reference the corresponding figures or tables in the text.
8. Experiments Compute Resources
Question: For each experiment, does the paper provide sufficient information on the com-
puter resources (type of compute workers, memory, time of execution) needed to reproduce
the experiments?
Answer: [Yes]
Justification: Appendix Section D.1 provides details.
Guidelines:
• The answer NA means that the paper does not include experiments.
• The paper should indicate the type of compute workers CPU or GPU, internal cluster,
or cloud provider, including relevant memory and storage.
• The paper should provide the amount of compute required for each of the individual
experimental runs as well as estimate the total compute.
• The paper should disclose whether the full research project required more compute
than the experiments reported in the paper (e.g., preliminary or failed experiments that
didn’t make it into the paper).
9. Code Of Ethics
Question: Does the research conducted in the paper conform, in every respect, with the
NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
Answer: [Yes]
Justification: Yes, we read the code of ethics and follow these.
Guidelines:
• The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.
• If the authors answer No, they should explain the special circumstances that require a
deviation from the Code of Ethics.
• The authors should make sure to preserve anonymity (e.g., if there is a special consid-
eration due to laws or regulations in their jurisdiction).
10. Broader Impacts
Question: Does the paper discuss both potential positive societal impacts and negative
societal impacts of the work performed?
Answer: [NA]
Justification: To the best of our knowledge the research proposed in the paper does not have
any negative social impact.
Guidelines:
• The answer NA means that there is no societal impact of the work performed.
• If the authors answer NA or No, they should explain why their work has no societal
impact or why the paper does not address societal impact.
• Examples of negative societal impacts include potential malicious or unintended uses
(e.g., disinformation, generating fake profiles, surveillance), fairness considerations
(e.g., deployment of technologies that could make decisions that unfairly impact specific
groups), privacy considerations, and security considerations.

41
 • The conference expects that many papers will be foundational research and not tied
to particular applications, let alone deployments. However, if there is a direct path to
any negative applications, the authors should point it out. For example, it is legitimate
to point out that an improvement in the quality of generative models could be used to
generate deepfakes for disinformation. On the other hand, it is not needed to point out
that a generic algorithm for optimizing neural networks could enable people to train
models that generate Deepfakes faster.
• The authors should consider possible harms that could arise when the technology is
being used as intended and functioning correctly, harms that could arise when the
technology is being used as intended but gives incorrect results, and harms following
from (intentional or unintentional) misuse of the technology.
• If there are negative societal impacts, the authors could also discuss possible mitigation
strategies (e.g., gated release of models, providing defenses in addition to attacks,
mechanisms for monitoring misuse, mechanisms to monitor how a system learns from
feedback over time, improving the efficiency and accessibility of ML).
11. Safeguards
Question: Does the paper describe safeguards that have been put in place for responsible
release of data or models that have a high risk for misuse (e.g., pretrained language models,
image generators, or scraped datasets)?
Answer: [NA]
Justification: Not applicable for our paper.
Guidelines:
• The answer NA means that the paper poses no such risks.
• Released models that have a high risk for misuse or dual-use should be released with
necessary safeguards to allow for controlled use of the model, for example by requiring
that users adhere to usage guidelines or restrictions to access the model or implementing
safety filters.
• Datasets that have been scraped from the Internet could pose safety risks. The authors
should describe how they avoided releasing unsafe images.
• We recognize that providing effective safeguards is challenging, and many papers do
not require this, but we encourage authors to take this into account and make a best
faith effort.
12. Licenses for existing assets
Question: Are the creators or original owners of assets (e.g., code, data, models), used in
the paper, properly credited and are the license and terms of use explicitly mentioned and
properly respected?
Answer: [Yes]
Justification: We have used only open-source resources and cited relevant owners of the
various resources, tools and models.
Guidelines:
• The answer NA means that the paper does not use existing assets.
• The authors should cite the original paper that produced the code package or dataset.
• The authors should state which version of the asset is used and, if possible, include a
URL.
• The name of the license (e.g., CC-BY 4.0) should be included for each asset.
• For scraped data from a particular source (e.g., website), the copyright and terms of
service of that source should be provided.
• If assets are released, the license, copyright information, and terms of use in the
package should be provided. For popular datasets, paperswithcode.com/datasets
has curated licenses for some datasets. Their licensing guide can help determine the
license of a dataset.
• For existing datasets that are re-packaged, both the original license and the license of
the derived asset (if it has changed) should be provided.

42
 • If this information is not available online, the authors are encouraged to reach out to
the asset’s creators.
13. New Assets
Question: Are new assets introduced in the paper well documented and is the documentation
provided alongside the assets?
Answer: [NA]
Justification: We do not create any new asset.
Guidelines:
• The answer NA means that the paper does not release new assets.
• Researchers should communicate the details of the dataset/code/model as part of their
submissions via structured templates. This includes details about training, license,
limitations, etc.
• The paper should discuss whether and how consent was obtained from people whose
asset is used.
• At submission time, remember to anonymize your assets (if applicable). You can either
create an anonymized URL or include an anonymized zip file.
14. Crowdsourcing and Research with Human Subjects
Question: For crowdsourcing experiments and research with human subjects, does the paper
include the full text of instructions given to participants and screenshots, if applicable, as
well as details about compensation (if any)?
Answer: [NA]
Justification: We do not perform any human experiments.
Guidelines:
• The answer NA means that the paper does not involve crowdsourcing nor research with
human subjects.
• Including this information in the supplemental material is fine, but if the main contribu-
tion of the paper involves human subjects, then as much detail as possible should be
included in the main paper.
• According to the NeurIPS Code of Ethics, workers involved in data collection, curation,
or other labor should be paid at least the minimum wage in the country of the data
collector.
15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human
Subjects
Question: Does the paper describe potential risks incurred by study participants, whether
such risks were disclosed to the subjects, and whether Institutional Review Board (IRB)
approvals (or an equivalent approval/review based on the requirements of your country or
institution) were obtained?
Answer: [NA]
Justification: Not applicable in our case.
Guidelines:
• The answer NA means that the paper does not involve crowdsourcing nor research with
human subjects.
• Depending on the country in which research is conducted, IRB approval (or equivalent)
may be required for any human subjects research. If you obtained IRB approval, you
should clearly state this in the paper.
• We recognize that the procedures for this may vary significantly between institutions
and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the
guidelines for their institution.
• For initial submissions, do not include any information that would break anonymity (if
applicable), such as the institution conducting the review.

43