Exam reading practice test 12 | English Test Prep

Reading Passage 1

When Scientific Models Meet Reality

AScientific models simplify the world so that researchers can study patterns that would otherwise be too complex to understand. A climate model, for example, cannot represent every cloud, tree, building and human decision individually. It uses selected variables and mathematical relationships to estimate how parts of the climate system may behave. Simplification is not a defect by itself; without it, many questions would be impossible to investigate. The challenge is deciding which details can safely be left out for the question being asked.

BA model is not the same as a prediction. Some models explore possible outcomes under different assumptions, while others are designed to forecast a specific event or trend. Confusion arises when a scenario is reported as if it were a guaranteed future. If a model estimates what could happen under a particular emissions pathway, its result depends partly on whether that pathway occurs. The model may be internally consistent and still produce an outcome that differs from reality because the assumptions change.

CValidation involves comparing a model's results with observations, ideally including data that were not used to build it. A model that reproduces the data on which it was developed may simply have learned the peculiarities of that dataset. Researchers therefore test performance in other periods or locations. Even successful validation does not prove that a model is universally correct; it provides evidence that the model is useful within a defined range.

DUncertainty is unavoidable, but it has different sources. Measurements may be imperfect, some processes may be poorly understood, and future human decisions may be unknowable. These uncertainties should not be collapsed into a single vague statement. Explaining which assumptions matter most can help decision-makers understand where further research would be valuable and which conclusions are relatively robust.

EModels are most useful when treated as tools for disciplined reasoning rather than machines that deliver unquestionable answers. Their value lies not in removing uncertainty but in making assumptions explicit, comparing alternatives and revealing where knowledge remains incomplete.

FThe choice of variables is one of the most important decisions in building a model. A researcher studying traffic may include road capacity, travel speed, and the number of vehicles, but may find it harder to represent how people change their plans when a new route opens. A model that leaves out such behaviour may still be useful for a narrow question, yet it could be misleading if used to predict the long-term effect of a major transport policy. The quality of a model therefore depends partly on how closely its design matches the question it is being asked to answer.

GModels also depend on the quality of the data used to build them. Historical records may contain gaps, measurements may have been collected using different methods, and some groups or locations may be under-represented. If the data reflect an unusual period, the model may treat that period as normal. Researchers can investigate these problems by comparing multiple data sources, examining where errors are concentrated, and explaining the limits of the available evidence. More data are not automatically better if the data are systematically incomplete.

HA model can be useful even when its estimates are not exact. For example, it may show that one policy is likely to create greater demand than another, or that a result changes sharply when a key assumption is altered. Such comparisons can help decision-makers identify which questions deserve further investigation. They should not, however, confuse a precise-looking number with certainty. A result reported to several decimal places may still rest on uncertain assumptions or imperfect measurements.

ICommunication is therefore part of modelling, not an afterthought. Researchers need to explain what a model includes, what it leaves out, and which conditions would make its results less reliable. A clear explanation allows other specialists and members of the public to question assumptions and consider alternative interpretations. Without that explanation, a model may appear more authoritative than the evidence justifies.

Reading Passage 2

The Problem of Overfitting

AA model can appear highly accurate when tested on the data used to create it yet perform poorly on new cases. This problem is known as overfitting. It occurs when a model captures random fluctuations or accidental features of its training data instead of the broader pattern researchers hope to understand. A complex model may fit the past exceptionally well but fail when conditions change.

BOne common response is to separate data into training and testing sets. The training data are used to develop the model, while the test data provide an independent check. Researchers may also use cross-validation, repeating the process with different portions of the data. These methods help estimate how well a model may generalise, although they cannot guarantee success in every future setting.

CModel complexity involves a trade-off. A simple model may miss important relationships, while a highly flexible one may follow noise too closely. The appropriate level of complexity depends on the problem, the quality of the data and the cost of different errors. In a medical screening system, for instance, failing to identify a serious condition may have different consequences from incorrectly flagging a healthy person. Evaluation must therefore consider the purpose of the model rather than rely on one score alone.

DResearchers should also document decisions made during development. If many versions are tested and only the best result is reported, the apparent performance may be overly optimistic. Transparent reporting allows others to understand how the model was built and what its evaluation can support. It also makes it easier to detect when a model is being used outside the conditions for which it was designed.

EOverfitting is particularly likely when a model has many adjustable features but is developed using a limited amount of data. The model can begin to represent accidental details that are unique to the examples it has seen. This may produce an impressive score during development while offering little help with cases that differ in small but important ways. The problem is not simply that a model is complex; complexity becomes risky when the evidence is insufficient to support it.

FThe separation between training and testing data must itself be handled carefully. If information from the test set influences decisions during development, the test set no longer provides a fully independent check. Researchers may unintentionally keep adjusting a model after repeatedly looking at its test performance. In such circumstances, the final score can become optimistic because the model has indirectly been shaped by the supposedly unseen data. A genuinely separate evaluation set, used only after the main development choices have been made, can help reduce this risk.

GCross-validation offers another way to estimate performance when data are limited. Researchers divide the available data into several portions and repeat the training and evaluation process, changing which portion is held out each time. The results show how sensitive performance is to the particular sample selected. However, cross-validation does not solve every problem. If all the samples share the same bias, or if the data do not represent the conditions in which the model will be used, the estimates may still be misleading.

HThe costs of overfitting depend on the application. In a low-risk recommendation system, a poor prediction may simply produce an irrelevant suggestion. In medicine, engineering, or public safety, the same kind of error could have more serious consequences. Researchers should therefore evaluate not only average performance but also the kinds of mistakes made, the groups most affected, and the consequences of a wrong result. A single score can hide important differences between errors.

IA further safeguard is to examine performance on data that reflect the environment where the model will actually operate. A system developed using records from one hospital, region, or time period may not transfer well to another. Researchers should investigate these differences before assuming that a successful result will generalise. When performance changes across settings, the difference can reveal a weakness in the model or a genuine difference in the populations being studied.

Reading Passage 3

Models in Public Decisions

AGovernments and organisations increasingly use models to support decisions about transport, health, energy and the environment. These tools can compare options and expose trade-offs, but they do not remove the need for judgement. A transport model might estimate that a new road reduces average journey time while saying little about the neighbourhoods divided by the route. A health model might identify areas of high demand without explaining why services are difficult to access.

BThe choice of what to measure is itself a value judgement. If an evaluation counts only average travel time, it may overlook effects on people who do not own cars. If an energy model focuses only on total cost, it may give insufficient attention to reliability or the distribution of benefits. Decision-makers should ask who gains, who bears the costs and which outcomes remain invisible in the analysis.

CPublic communication matters because model results can appear more precise than the evidence allows. A figure presented to several decimal places may create an impression of certainty even when the inputs are uncertain. Good communication explains the assumptions, the range of plausible outcomes and the conditions under which a conclusion might change. It should also distinguish technical questions, which experts can investigate, from social choices that require public discussion.

DModels can improve public decisions when they are open to scrutiny and used alongside local knowledge. Residents may identify constraints that are absent from a dataset, while technical analysis can reveal consequences that are difficult to see from individual experience alone. Neither source should automatically replace the other. The aim is not to make decisions without uncertainty, but to make better-informed choices while acknowledging what remains unknown.

EA model can make some consequences visible while leaving others outside the frame. When a transport plan is assessed mainly by average journey time, people living near a proposed road may experience noise, pollution, or reduced access to local shops that do not appear in the headline result. Similarly, a model that predicts total energy demand may not reveal which households will face higher bills or which communities will host new infrastructure. The choice of indicators therefore affects which concerns become part of the discussion.

FPublic decisions also involve values that cannot be settled by technical calculation alone. A model may estimate the likely cost of two options, but deciding how to balance cost against fairness, reliability, environmental impact, or cultural importance requires public judgement. Experts can clarify the likely consequences of different choices; they cannot determine everyone's priorities simply by producing a numerical result. Decision-makers should make those value choices explicit rather than presenting them as if they were dictated by the model.

GPublic trust depends partly on whether the modelling process can be questioned. Explaining the assumptions, publishing methods where possible, and inviting independent review can reveal weaknesses before a decision is finalised. In some cases, it may be appropriate to compare several models rather than rely on one. If different models reach similar conclusions despite using different assumptions, confidence may increase. If they disagree, the disagreement can identify which uncertainties matter most.

HModels should also be revisited after a policy is introduced. Actual outcomes can differ from forecasts because people respond to the policy, economic conditions change, or unexpected events occur. Comparing predictions with later observations helps researchers learn where the model was useful and where it needs revision. This process is not evidence that modelling has failed; it is part of using models responsibly in a changing world.

IThe public does not need to understand every equation to take part in decisions affected by models. It does need access to a clear explanation of what the model can and cannot say, whose interests were considered, and how uncertainty was handled. Good communication allows technical evidence to inform democratic debate without pretending that complex choices have one inevitable answer.