Inferential Reading Skills in High School: A Study on Comprehension Profiles

Abstract

Reading comprehension of connected texts is a key skill in high school education, yet students exhibit varying proficiency levels, particularly in inferential reasoning. This study investigates reading behavior by means of finger-tracking technique and question answering among Italian 10th, 11th, and 12th year high school students, analyzing their performance on different types of questions: synonymy and reference vs. inference-based questions. Despite similar reading times and lexical effects across grades, students’ accuracy in answering inferential questions reveals significant variability. Subsequently, we identify three comprehension profiles—poor, medium, and good comprehenders—with the first two groups showing markedly lower performance on inference-based questions. These findings suggest that schooling alone may not be sufficient for all students to develop strong inferential skills, and some may benefit from targeted instructional support.

1. Introduction

Reading is a highly complex cognitive process that allows individuals to extract meaning from written symbols. It represents a critical skill for individuals’ personal growth, not only as the cornerstone for educational achievement but also as an essential tool for effectively navigating everyday challenges. Understanding written language requires visually processing printed words, to activate their phonological, orthographic, syntactic and semantic features, and to integrate these components accordingly, resulting in a coherent mental representation of the text. The intricate interplay of numerous simultaneous processes within a brief span of time demands efficient engagement of a highly interconnected and complex system.

It is not surprising that an exhaustive and well-established body of research has been conducted across the fields of psycholinguistics, education, and computer science. Over the last decades, scholars have made significant efforts to explore different aspects of reading from various perspectives, employing diverse methods and pursuing distinct objectives (Elleman & Oslund, 2019; Frost, 2012; Peng et al., 2022; Rayner & Reichle, 2010). Several theoretical models have been proposed to account for the cognitive processes involved in reading comprehension. One influential framework is the Simple View of Reading (Hoover & Gough, 1990), which posits that reading comprehension results from the interaction of decoding and linguistic comprehension. The Construction-Integration Model (Kintsch, 1988) distinguishes between the construction of meaning based on textual information and the integration of this information with prior knowledge. More recent models, such as the Landscape Model (Yeari & Van den Broek, 2011), further emphasize the dynamic and memory-based nature of comprehension, accounting for moment-to-moment shifts in attention and activation of inferences. Although these theoretical models emphasize various aspects of reading comprehension, they share the central notion that the cognitive processes at stake can be broadly categorized into two main types: lower-level processes (e.g., word recognition, syntactic parsing) and higher-level processes (e.g., inference generation, integration of information) (Kendeou et al., 2014; Perfetti & Stafura, 2014). These processes work together to enable the reader to extract meaning from written text and build a coherent mental representation of text content.

Lower-level processes involve efficient word recognition (decoding: Tighe & Schatschneider, 2016) and their translation into meaningful linguistic units (vocabulary knowledge: Oslund et al., 2018), as well as chunking portions of the text efficiently to read them out with the appropriate pace (reading rate: Brysbaert, 2019).

Higher-level processes play a crucial role in integrating these linguistic units to form a coherent and meaningful understanding of the text. Inference-making, in particular, represents the ability to connect various parts of the text and link it to prior knowledge to fill in gaps and build deeper understanding (Silagi et al., 2018). For inferences to be correctly drawn, previously encountered information must be kept in working memory while processing incoming linguistic material, allowing the reader to manage and structure content meaningfully (Pérez et al., 2014). Similarly, an inhibition mechanism enables suppression of irrelevant information, thus determining what to retain into the working memory buffer (Borella et al., 2010). These executive functions, together with attentional allocation and metacognitive monitoring (Liu et al., 2013), contribute significantly to reading comprehension and may help explain inter-individual variability beyond linguistic skills alone.

Furthermore, reading is a skill that develops with time and practice, which can be improved by expanding vocabulary, improving comprehension, and exercising critical skills in evaluating and analyzing information (Castles et al., 2018). Both lower-level and higher-level processes begin to develop before formal reading instructions are provided. These processes are also recognized as independent predictors of later reading comprehension ability. Lower-level processes undergo rapid development during early childhood and become more automatized, particularly during the initial stages of reading acquisition in the early grades of elementary school. Decoding skills improve significantly as children learn to recognize letters and sublexical elements such as ngrams or morphemes (García & Cain, 2014). Reading fluency emerges as children practice and gain confidence in recognizing printed words and phrases with higher speed and accuracy (Foorman et al., 2018). Vocabulary knowledge expands through exposure to both spoken and written language (Samuelson, 2021). As these processes become increasingly automatic, they unload cognitive resources, allowing readers to devote greater attention to higher-level comprehension tasks.

However, higher-level processes develop more gradually, requiring sustained cognitive effort and practice over an extended period. These processes evolve significantly from early childhood through adolescence and even into adulthood. Inference-making abilities grow as readers gain experience with varied texts and learn to apply their background knowledge more effectively (Daugaard et al., 2017). Executive functions, such as organizing and integrating information, strengthen as working memory capacity and cognitive control improve over time (Cartwright, 2012). Skills in attention allocation, such as focusing on central themes and monitoring comprehension, continue to mature with experience and exposure to increasingly complex texts (Arrington et al., 2014).

The interaction between lower-level and higher-level processes has been shown to be dynamic and interdependent (Verhoeven et al., 2011). Automatization of lower-level processes during early development allows readers to allocate more cognitive resources to the higher-level processes that are essential for deep comprehension. At the same time, higher-level processes continue to enhance and refine the reader’s ability to interpret and derive meaning from text throughout their lifetime. Vice versa, difficulties with lower-level processes can place additional demands on working memory, thereby hindering the formation of a coherent mental model.

Most research on reading acquisition has focused on beginning readers and simple sentence structures. However, reading skills continue to evolve throughout secondary school as students, who get older and more experienced though not yet mature readers, encounter increasingly complex linguistic material.

Today’s students face numerous academic and non-academic hurdles that require a fully-fledged reading skill to tackled. For a text to be understood, it is not enough to recognize words and understand their literal meaning. Readers must also engage in practices such as asking critical questions about terminology, concepts, and the purpose of the text itself; identifying elements of the author’s argument; evaluating the evidence and credibility of the content; and positioning themselves in relation to the text.

However, the literacy achievement of adolescents in Italy according to standardized national tests (INVALSI, https://www.invalsi.it, accessed on 13 May 2025; OECD-PISA, https://www.oecd.org/en/about/programmes/pisa.html, accessed on 13 May 2025) has been a concern for many years. In particular, results from the Invalsi Test conducted in 2024 showed a decreasing trend in the scholastic achievement for Italian classes: only 62% of students reached the basic level of competence, with a reduction of eight percentage points compared to the 2019 surveys. Such data may indicate that some students could struggle to reach proficient literacy levels, and that further training could be required for them to pursue academic success and personal growth. Although reading strategies and capability are essential for high school students to become effective and engaged readers, research on monitoring their linguistic competence appears to be lacking.

In reading research and education, the limited availability of Information and Communication Technology (henceforth ICT) tools for large-scale data collection has posed significant challenges (Petscher et al., 2020). It has hindered the replication of lab results on a broader scale, the evaluation of educational outcomes using pre-test and post-test designs, and the development of long-term educational practices tailored to specific groups of struggling readers—a concern frequently raised by major international organizations (ELINET, https://elinet.pro, accessed on 13 May 2025). In this context, ReadLet (https://www.readlet.it/apps/readlet/, accessed on 13 May 2025) represents an innovative ICT platform designed to deliver accurate, evidence-based assessments of reading efficiency through an ecological, non-invasive protocol for extensive data collection, storage, and analysis (Ferro et al., 2018; Taxitari et al., 2021). Developed as a web app on an ordinary tablet, it uses the tablet touchscreen to capture a subject’s reading behavior by tracking the movements of her/his dominant hand’s index finger as it slides along the text during reading. Recorded finger movements are continuous in both the space and time dimensions, as they extend over individual letters, words, punctuation marks and blanks evenly. This dynamic appears to be in contrast with the discrete nature of ocular movements captured by the eye-tracking technique, whereby the eyes alternate between fixation points via rapid, ballistic movements called saccades.

However, there is growing evidence that finger-tracking and eye-tracking yield remarkably similar patterns of reading behavior in proficient readers. Notably, it has been shown that finger- and eye-tracking measurements in young adult engaged in silent reading showed strong correlations with embedding levels of linguistic units, ranging from $\rho =0.66$ at the token level, $\rho =0.81$ at the chunk level, and $\rho =0.98$ at the sentence level (Crepaldi et al., 2022). Consistently, recent findings indicate that finger-voice span and eye-voice span similarly vary as a function of the syntactic and prosodic structure of the text to be read aloud (Nadalini et al., 2023). These converging findings underscore the validity of finger-tracking as a reliable proxy for traditional eye-tracking measures in capturing reading dynamics. Building on this foundation, the ReadLet protocol was developed to translate such research insights into scalable applications suitable for classroom and field-based settings.

In the Readlet protocol, a subject is asked to complete a comprehension questionnaire, at the end of each reading session. The collected data are then encrypted, pseudonymized, and securely transmitted to a central repository via a secure Internet protocol. Offline, the sequence of touch events is spatially aligned with the position of words on the page. Additionally, a suite of Natural Language Processing (NLP) tools annotates all words in the text along various linguistic levels, including word length, frequency, part of speech, morphological complexity, and syntactic dependency. These linguistically annotated texts can be further analyzed and classified by readability levels. By linking a subject’s reading performance to specific linguistic features in the text, ReadLet facilitates a deeper understanding of the factors shaping individual reading strategies. It has, in fact, been designed as an ecological and ubiquitous protocol for the collection of large data repositories. It can also be applied as a user-friendly application within classroom environments to identify students’ strengths and weaknesses. In this way, it opens the path toward evidence-based, targeted interventions for both typical and atypical readers, across different stages of reading development (Ferro et al., 2024; Marzi et al., 2022; Nadalini et al., 2023).

In light of this body of evidence, in this paper we propose a reading study with students from grade 10th to 12th attending a technical school located in the urban area of the city of Genova (Italy). We provide a fine-grained quantitative analysis of their reading profile through the ReadLet app, and analyze their behavior in relation with their proficiency of understanding as revealed by a specifically-designed comprehension questionnaire. In detail, the questionnaire was developed to take into account those skills providing readers with the necessary tools to efficiently interpret and understand written content, i.e., vocabulary knowledge, referential disambiguation and inference making. A rich vocabulary is essential for readers as it helps them grasp the meaning of words and terms, thus significantly enhancing overall comprehension. Identifying references, such as pronouns, is equally important for connecting referents and improving clarity, which facilitates understanding explicit details in the text. Lastly, making inferences involves drawing conclusions or insights that may not be directly stated in the text, relying on critical cognitive skills like reasoning and language comprehension. Inferential processing, therefore, refers to the ability to connect various parts of the text and integrate them with prior knowledge, filling in gaps to create a deeper understanding. While vocabulary and referential resolution contribute to the accurate processing of explicit textual information, inferential processing allows readers to go beyond the surface level by constructing meaning that is implied rather than overtly stated.

Building on these assumptions, the present study addresses the following research questions: (i) To what extent do students’ comprehension profiles, particularly regarding inferential processing, improve across grades? (ii) How do lexical features of the text (e.g., word frequency, word length) affect reading behavior, and are these effects modulated by comprehension skill level? We hypothesize that students will show increasing inference-making abilities with proficiency, regardless of grade level, and that more skilled readers will exhibit a reduced sensitivity to lexical features, reflecting a shift toward deeper, integrative processing.

2. Materials and Methods

2.1. Participants

63 participants took part in our study, with 18 tenth-grade (age range: 15–17 years; mean = 15.1 years), 26 eleventh-grade (age range: 16–18 years; mean = 16.1 years), and 19 twelfth-grade students (age range: 17–19 years; mean = 17.3 years), all attending the same technical school in Genova. Informed consent was personally signed by students aged 18 and above, or by parents in the case of underage students. All participants had normal or corrected-to-normal vision, with no reported history of neurological issues.

2.2. Materials

The materials used in the current study were sourced directly from the PRIN project ReMind (Albertin et al., 2024), which initially developed them as part of a broader effort to support diagnostic assessments for mild cognitive impairment (MCI). Reading texts were two short excerpts adapted and rearranged from original articles published in the Italian science communication magazine Focus. Given their relevance to the general public and for their linguistic characteristics, they can be considered as suitable for high school students—as confirmed by the readability ease calculated according to the Gulpease Index (Lucisano & Piemontese, 1988). To ensure transparency, an English translation of both texts is provided in the Appendix A. It is worth noting that participants’ prior knowledge of the topics addressed in the texts was not explicitly measured. Each text extended over two tablet pages. Lexical features are reported in

Table 1. Values of the lexical features of the two texts used in the current study.

	Text 1	Text 2
Text Length (words)	211	210
Mean Word Length (letters)	4.78	4.88
Mean Word Log Frequency	4.34	4.32
Type-Token Ratio	0.66	0.69
Lexical Density	0.52	0.58
Part-Of-Speech Type	11	10
Mean Sentence Length (words)	16.23	16.15
Readability index (Gulpease)	58.2	57.0

.

The questionnaire, shown to participants after each reading session, comprised seven questions addressing different aspects of reading comprehension. Two questions tackled vocabulary competence via synonym detection. Three questions were requiring the reader to disambiguate referential links such as pronouns or to recognize who performed a given action. Finally, two other questions tapped into the ability to draw inferences on the basis of information given in the text.

2.3. Procedure

The reading sessions were conducted in a quiet room inside the school, in the presence of at least one researcher. During each experimental session, up to eight participants were sitting in front of a school desk, in front of a tablet in portrait position. Finger-tracking data collection was administered via the ReadLet application installed on a 10.1 inches tablet, equipped with a 1.8 GHz Octa-Core processor, 3 GB RAM, 64 GB eMMC and Android10, at a sampling rate of 120 Hz rate. The tablet screen was 14.9 cm × 24.5 cm with a resolution of 1920 × 1200 pixels. Both texts were presented in Arial (21.25 pt).

Each student took part in two consecutive silent reading sessions. Before the actual experiment started, participants underwent a practice trial and were instructed to use the tip of the index finger of their dominant hand to finger-track a short passage. This practice trial required to cover at least 60% of the text to be completed. After each experimental reading session, participants were presented with seven reading comprehension questions with four randomly shuffled answers, of which only one was the correct answer.

2.4. Data Analyses

Text-to-finger alignment was performed automatically using a convolutional algorithm to find the best match between text lines and touch event sequences. For each uninterrupted sequence of touch events within a letter’s bounding box, tracking time was calculated as the difference between the final and initial timestamps of the sequence. The finger-tracking time for any other text unit was determined by summing the tracking times of the letters it spans. Data analysis was conducted with R (R Core Team, 2024). Descriptive statistics and data visualization were obtained via custom scripts using base R functions, while data modeling was implemented via generalized linear mixed-effect models (GLMM) using the lme4 package (Bates et al., 2015). This choice allowed us to avoid the need for transformation of skewed data or to select a theoretical distribution that matches the properties of the dependent variable (Lo & Andrews, 2015). Overall effects were tested using Type III sum of squares and ${\chi }^2$ Wald tests. For the analysis of reading times, we modeled finger-tracking data as a Gaussian distribution with an inverse link. Independent predictors included the categorical factor of grade level, in interaction with question accuracy (modeled as second order polynomial to account for nonlinearity), word’s length and frequency. Additionally, we included random intercepts for individual subjects and word tokens. For the analysis of the questionnaire, we modeled question accuracy as a binomial distribution with a logistic link. The model included categorical factors of grade level and question types (classified in synonym detection, referential link, and inference ability), as well as their interaction, and a random intercepts for individual subjects. Finally, we followed-up the aforementioned analyses by regrouping participants according to their comprehension proficiency as determined by their performance on the questionnaire, rather than by their grade level. In particular, a reader was labeled as a good comprehender if her/his accuracy was above the 3rd quartile of the overall distribution (85%, n = 16), a poor comprehender if her/his accuracy was below the 1st quartile (57%, n = 12), and a medium comprehender if her/his accuracy was in between (n = 29). GLMM analysis of the reading profiles checked whether there was any difference in the tracking times of such groups, while the GLMM analysis of the questionnaire assessed how different question types were managed by them.

3. Results

3.1. Descriptive Statistics

We collected data from 63 participants, for a total of 248 pages and 26,103 word tokens. Before running the analyses, we preprocessed the data to detect and remove outliers. First, we took out individual pages where the reader tracked less than 75% of the overall text, leading to the exclusion of 55 pages and 5897 word tokens. Next, we filtered out individual word tokens due to extremely low (<35 ms) or high (>1500 ms) tracking time, or because they were covered for less than half of their length (n = 3386). The final dataset thus comprised 57 readers, 193 pages and 16,820 word tokens. To get a first impression of readers’ behavior, we visualized the relationship between tracking times and question accuracy, averaged across participants, and assessed their correlation, for each grade level separately, as shown in Figure 1. Interestingly, we observed an overall trend according to which readers with faster tracking times were those achieving better accuracy in the questionnaire, independently of their grade level.

3.2. Reading Profiles (Tracking Times) by Grade Levels

For the tracking time analysis, the GLMM model (

Table 2. Summary of the tracking time GLMM by grade levels. The baseline level of the model corresponds to grade 10th. Due to the inverse link function, the sign of coefficients reflects an inverse relationship with reading time.

Tracking Time ∼ GradeLevel * (poly(QuestionAccuracy, 2) + Length + Frequency) + (1|Token) + (1|Subject)
	Estimate	Std. Error	t-Value	p-Value
Intercept	4.81	0.22	21.75	<0.001
gradeLevel_11th	0.59	0.22	2.67	0.008
gradeLevel_12th	0.03	0.26	0.14	0.893
poly(question accuracy,2)1	32.50	6.28	5.18	<0.001
poly(question accuracy,2)2	14.04	6.68	2.10	0.036
	Estimate	Std. Error	t-Value	p-Value
Length	−1.27	0.09	−14.25	<0.001
Frequency	0.25	0.10	2.48	0.013
gradeLevel_11th: poly(question accuracy,2)1	−1.86	8.81	−0.21	0.832
gradeLevel_12th: poly(question accuracy,2)1	−14.43	7.40	−1.95	0.050
gradeLevel_11th: poly(question accuracy,2)2	−43.26	7.46	−5.80	<0.001
gradeLevel_12th: poly(question accuracy,2)2	−18.79	8.17	−2.30	0.021
gradeLevel_11th:Length	−0.05	0.03	−1.93	0.053
gradeLevel_12th:Length	0.08	0.03	3.22	0.001
gradeLevel_11th:Frequency	−0.09	0.04	−2.21	0.027
gradeLevel_12th:Frequency	−0.07	0.03	−2.17	0.030
R2 = 0.989

, Figure 2) revealed a significant main effect of grade level (${\chi }^2$ = 11.26, p = 0.004): 11th graders were slightly—though significantly—faster than 10th and 12th graders, with no significant difference between the latter ones. The main effect of question accuracy was also significant (${\chi }^2$ = 30.70, p < 0.001), with faster readers achieving higher accuracy in the questionnaire. Benchmark effects of length (${\chi }^2$ = 202.99, p < 0.001) and (log) frequency (${\chi }^2$ = 6.13, p = 0.013) were also attested: longer and/or infrequent words resulted in higher tracking times than shorter and/or frequent ones. All interactions were also significant (all p < 0.05). As shown in Figure 2, the effects of word (log) frequency and length were more pronounced for 10th graders. However, contrary to what might be expected, no clear pattern of progressively decreasing dependence on word frequency and length emerges as grade level increases, as clearly shown in Figure 2 (bottom panels).

3.3. Reading Comprehension (Question Accuracy)

For the question accuracy data, the GLMM model (

Table 3. Summary of the question accuracy GLMM by grade levels. The baseline level of the model corresponds to grade 10th.

Question Accuracy ∼ GradeLevel * QuestionType + (1|Subject)
	Estimate	Std. Error	z-Value	p-Value
Intercept	0.20	0.36	0.54	0.59
gradeLevel_11th	−0.15	0.45	−0.33	0.74
gradeLevel_12th	−0.27	0.48	−0.56	0.57
questionType_ref	0.55	0.41	1.33	0.18
questionType_syn	1.37	0.50	2.77	0.01
gradeLevel_11th:questionType_ref	0.36	0.51	0.70	0.48
gradeLevel_12th:questionType_ref	0.60	0.55	1.10	0.27
gradeLevel_11th:questionType_syn	0.45	0.62	0.73	0.46
gradeLevel_12th:questionType_syn	0.21	0.64	0.33	0.74
R2 = 0.257

, Figure 3) revealed the significant main effect of question type (${\chi }^2$ = 46.67, p < 0.001): questions addressing lexical knowledge achieved the highest accuracy (syn: 83%), followed by those tapping referential disambiguation (ref: 68%) and inference making (inf: 55%). Interestingly, the main effect of grade level was not significant (${\chi }^2$ = 0.12, p = 0.94), as well as the interaction of grade and question types (${\chi }^2$ = 1.57, p = 0.81).

3.4. Results per Comprehension Proficiency

The GLMM analysis of reading profiles showed a main effect of comprehension group (${\chi }^2$ = 6.93, p = 0.031), whereby the tracking times of good comprehenders were faster than the tracking times of medium and poor ones, with no significant difference between these two. Remarkably, word length (${\chi }^2$ = 33.61, p < 0.001) and word frequency (${\chi }^2$ = 27.46, p < 0.001) were modulated by comprehension groups, as revealed by the significant interactions. Model results are reported in

Table 4. Summary of the reading profiles GLMM by comprehension groups. The baseline level of the model corresponds to the good comprehension group. Due to the inverse link function, the sign of coefficients reflects an inverse relationship with reading time.

Tracking Time ∼ Comprehension * (Length + Frequency) + (1|Token) + (1|Subject)
	Estimate	Std. Error	z-Value	p-Value
Intercept	5.31	0.17	30.50	<0.001
Comprehension_medium	−0.46	0.20	−2.34	0.019
Comprehension_poor	−0.50	0.25	−1.99	0.047
Length	−1.18	0.08	−14.93	<0.001
Frequency	0.14	0.09	1.51	0.131
Comprehension_medium:Length	0.12	0.03	4.36	<0.001
Comprehension_poorLength	0.19	0.03	5.77	<0.001
Comprehension_medium:Frequency	0.01	0.04	0.18	0.85
Comprehension_poorFrequency	0.17	0.04	4.03	<0.001
R2 = 0.986

and model estimates are plotted in Figure 4.

GLMM analysis of the question accuracy (

Table 5. Summary of the reading profiles GLMM by comprehension groups. The baseline level of the model corresponds to the good comprehension group.

Question Accuracy ∼ Comprehension * QuestionType + (1|Subject)
	Estimate	Std. Error	z-Value	p-Value
Intercept	1.51	0.31	4.94	<0.001
Comprehension_medium	−1.91	0.37	−5.12	<0.001
Comprehension_poor	−2.74	0.47	−5.79	<0.001
questionType_ref	1.01	0.48	2.12	0.034
questionType_syn	2.75	1.05	2.61	0.009
Comprehension_medium:questionType_ref	0.09	0.56	0.15	0.878
Comprehension_poorquestionType_ref	−0.35	0.65	−0.54	0.591
Comprehension_medium:questionType_syn	−0.37	1.12	−0.33	0.739
Comprehension_poorquestionType_syn	−1.99	1.15	−1.72	0.085
R2 = 0.428

, Figure 5) revealed the main effects of comprehension group (${\chi }^2$ = 120.76, p < 0.001) and question type (${\chi }^2$ = 43.84, p < 0.001), while their interaction only approached significance (${\chi }^2$ = 8.47, p = 0.076). Poor comprehenders performed below 50% accuracy for all the question types. Medium comprehenders showed an increase in their lexical knowledge, as evident by the higher accuracy in questions about vocabulary size (syn) and referential disambiguation (ref). However, they still lack the capability of making appropriate inferences (inf). Finally, good comprehenders seemed to master all the different types of knowledge that allowed for a deep understanding of the texts.

4. Discussion

Reading represents a complex cognitive process widely investigated by the scientific community. Two main lines of research can be seen in the related literature. Scholars from the field of psycho-linguistics have focused mostly on the factors that modulate word decoding and integration into a coherent text representation, as revealed by within-lab experiments monitoring on-line reading comprehension (e.g., via eye-tracking: Kuperman et al., 2024 or self-paced reading: Tremblay et al., 2011). Scholars from the field of education sciences have, instead, focused more on the different levels that allow the reader to comprehend a written text, such as lexical knowledge, executive functions and inference making. These components have been traditionally investigated via questionnaires and batteries of standardized tests conducted at a large scale (Clinton-Lisell, 2022). In the current paper, we attempted to bridge these lines of research, shedding new lights on the dynamic interaction between reading strategy and comprehension proficiency. In particular, we looked at finger-tracking times in the context of connected text reading and the corresponding level of understanding, in a group of junior high school students from grade 10th to 12th.

Our results did not reveal any striking difference across grades. By looking at the grade-level effect, we observed a similar performance in the question accuracy, while differences in tracking times did not follow a clear developmental pattern. This is likely due to the fact that the learning curve of reading competence is approaching its ceiling by the time teenagers start high school, although some capabilities continue to evolve throughout the whole adolescence. The absence of a clear developmental pattern in tracking times might reflect individual differences in processing strategies rather than a uniform progression across grades. In fact, those participants who show to have a good vocabulary knowledge also score high inferential reading comprehension percentage. Vice versa, those who have low vocabulary also have low inferential reading comprehension, in line with recent evidence that shows the significant relation between vocabulary knowledge and inferential reading comprehension (Royeras & Sumayo, 2024).

Interestingly, reading times of good and average comprehenders are not significantly affected by word frequency, as compared to poor ones. Typically, research on reading development has shown that early-stage readers (e.g., primary school children) are highly sensitive to lexical effects, such as word frequency and word length (Burani et al., 2008; De Luca et al., 2008). As reading proficiency increases, these effects tend to diminish, suggesting a shift from bottom-up decoding to more efficient, top-down comprehension processes. However, our results suggest that this transition is not merely a function of age or grade level and is rather more closely tied to comprehension abilities themselves. Rather than progressing uniformly across all students, the reduction of word frequency effects appears to be characteristic of those with sufficient and good inferential and integrative reading skills. This hypothesis seems to be strengthened by our finding that reading comprehension modulates tracking times so that readers who achieved a deeper level of understanding were those showing a faster reading pace, for all the grades involved in our study. The idea that processing speed represents a crucial cognitive skill for reading acquisition is not new (Christopher et al., 2012). Efficient text reading requires to smoothly match words to stored representations, and faster encoding of words and sentences allows for quicker integration and meaning construction before working memory resources decline. Additionally, according to bottleneck and processing speed theories (Christiansen & Chater, 2016), processing speed acts as a constraint on higher-level cognitive functions. Slower processing can limit the availability of information for cognitive processes like executive functions and inference making, potentially impairing performance in domain-specific skills such as reading.

Another important issue of the current study is represented by the analyses of the question accuracy. By regrouping students according to their performance in the questionnaire, we observed a clear-cut evidence of how comprehension of a connected text is based on different components, each of which provides its own contribution. In particular, we monitored (i) readers’ lexical knowledge via synonym detection; (ii) their ability to recognize entities in the text to disambiguate referents and assess who-did-what; and (iii) their ability to draw inferences to link the text content to their previous knowledge. At one extreme, we found that students with the lowest accuracy could not respond to any of the question types. At the opposite extreme, students who scored highest in accuracy were able to master all the different kinds of competence required for a deep understanding of written materials. In between, there were students who performed fairly well in questions monitoring lexical knowledge and entities recognition, yet lack the ability to make appropriate inferences.

Our results suggested that poor comprehenders lack low-level skills that lay behind efficient word recognition, which in turn makes it harder for them to select salient elements from the text while inhibiting irrelevant information. Without grasping the meaning of the text being read, it is not surprising their inability to perform inferential processing. The middle group instead performed fairly well in questions monitoring lexical knowledge (vocabulary size) and referential assignment (i.e., they could allocate enough attentional resources to keep track of who-did-what along the text). Apparently, they are able to grasp and handle local information that is explicitly stated in the text, by correctly recognizing the named entities and the action performed. However, they seemed to lack the ability to draw inferences. Inferences are mental representations constructed by the readers through a combination of their own knowledge and the explicit information in the text. This process enables the establishment of relationships and associations necessary for understanding implicit information. Effective inferential processing requires analyzing beyond the given content, drawing on world knowledge, assumptions, deductions, contextual factors, and textual cues to derive deeper meaning. This is in line with studies suggesting that inferential processing imposes a high cognitive demand (Barth et al., 2015; Singer & Leon, 2007).

Overall, the current study provides support to theoretical models on reading (e.g., Reading Systems Framework: Perfetti & Stafura, 2014; Simple View of Reading: Catts, 2018), whereby reading comprehension emerges from the interaction and integration of low-level (e.g., vocabulary size) and high-level (e.g., inference making) processes. This dynamic assumes that the extent to which cognitive abilities are involved in a reading task largely depends on the efficiency with which the task is performed. When lower-level skills are not well-developed or fully automatized, the engagement of the higher-level ones becomes too cognitively demanding. While the present study focused on linguistic aspects of reading proficiency—such as vocabulary knowledge, referential processing, and inference-making—we acknowledge that other non-linguistic cognitive factors also play a crucial role in shaping comprehension performance. In particular, working memory and inhibitory control are known to support the integration of textual information and the suppression of irrelevant cues during reading (Borella et al., 2010; Pérez et al., 2014). These constructs were discussed in the introduction section to emphasize their theoretical relevance; however, they were not directly assessed in our protocol. Future work should include dedicated cognitive assessments to evaluate how such executive functions contribute to different comprehension profiles, especially in readers showing weaker inferential skills. Including these measures would help clarify the interplay between linguistic and cognitive dimensions of reading and provide a more comprehensive account of individual differences.

5. Conclusions

Our results provide evidence that reading comprehension skills in high school students, and in particular inferential reasoning, do not develop automatically with grade progression. Despite comparable lexical effects across 10th, 11th, and 12th grade students, significant differences emerge in inferential comprehension. The classification into poor, medium and good comprehenders highlights that the first two groups struggle considerably more with inference-based questions, suggesting a specific gap in implicit information processing. In addition, our evidence shows that the more skilled a reader is and the less her/his reading strategy relies on lexical features. Taken together, our results align with cognitive models that conceptualize reading comprehension as relying on both automatic and higher-order processes. This interpretation is consistent with studies on reading expertise, which highlight how skilled readers gradually rely less on frequency-based recognition and more on discourse-level integration (Rayner, 1998; Staub & Rayner, 2007). The dissociation between comprehension proficiency and word-level processing underscores the importance of inferential reasoning as a key factor in reading development, beyond simple exposure and automatization.

This result is consistent with theoretical frameworks such as the Construction-Integration Model (Kintsch, 1988) and the Landscape Model (Yeari & Van den Broek, 2011), which propose that proficient readers use lexical features in a top-down, context-driven manner. Rather than engaging in bottom-up decoding of individual words, skilled readers are supposed to anticipate and integrate information across the text using predictive processing and higher-order comprehension strategies (Perfetti & Stafura, 2014). In such account, lexical features serve not as processing bottlenecks, but as cues that help guide the construction of meaning at the sentence and discourse level. Our findings support this perspective, showing that more proficient comprehenders are less sensitive to lexical variables such as word frequency and length, suggesting a shift from surface-level decoding to more global, meaning-oriented reading strategies. Accordingly, reduced sensitivity to lexical constraints among good comprehenders likely reflects a shift toward more efficient, predictive processing, where meaning is constructed proactively, rather than through sequential lexical decoding. In summary, this pattern reflects how good comprehenders are more likely to engage in more efficient semantic activation and context-based anticipation, thereby minimizing their sensitivity to lexical constraints such as word frequency (Kuperman & Van Dyke, 2011).

In contrast, poor comprehenders might rely more heavily on basic lexical processing, focusing on word-level recognition and surface features, which makes them more susceptible to lexical properties such as word frequency or familiarity. However, since our questionnaire focused primarily on inferential, referential and synonymy-based questions, we did not include items explicitly targeting literal comprehension.

Our findings have important educational implications. They underscore the need for targeted instructional strategies to foster inferential reasoning and processing of implicit textual information. Effective interventions could include guided questioning techniques to bridge explicit and implicit content, and exposure to diverse textual materials that require inference making. While our study did not directly assess students’ metacognitive strategies, our findings are consistent with the idea that enhancing students’ awareness of their own comprehension processes may support their ability to derive unstated meanings from connected text, in line with recent evidence on the role of metacognition in reading comprehension (Cartwright, 2023; Hall et al., 2020; Rice & Wijekumar, 2024).

These results underscore the relevance of targeted instructional approaches aimed at fostering inferential reasoning and processing of implicit textual information. Although our data do not provide direct evidence on the efficacy of such interventions, our findings highlight the need for further research in this direction, especially in educational settings. Accordingly, future research should explore the longitudinal effects of inference-focused training programs to determine their impact on students’ reading comprehension skills over time. Additionally, further investigations into cognitive factors such as working memory and inhibitory control could provide deeper insights into the mechanisms underlying inferential reasoning difficulties. Examining the role of metacognition in shaping comprehension profiles may also yield valuable perspectives for designing more effective educational interventions. Finally, examining the impact of targeted interventions aimed at strengthening inferential skills could help determine whether reducing lexical dependency is a trainable component of reading proficiency. To this aim, the adoption of the Readlet finger tracking technique in the present study provided evidence that it could be a valuable tool for identifying strengths and weaknesses in adolescent readers. A user-friendly application, easily integrable in ordinary classroom activities, would represent a step forward toward large-scale acquisition of naturalistic reading data.