Referential Adaptation in the Face of Changing Input Frequencies: The System Wants to Return to Baseline

Abstract

Evidence shows that adults adapt to patterns found in their language environment both at the syntactic level and referential level. However, little is known about how adaptation shifts in response to changes in the input, and existing theories make conflicting predictions. In two experiments, we examine referential adaptation with a between‐subjects blocked priming paradigm to test whether participants adapt flexibly if multiple referential patterns are present in the language input. The study exposed subjects to two blocks of exposure stimuli, each featuring a distinct referential pattern. Ambiguous critical sentences were interleaved throughout, in order to measure the effect of exposure to each reference remention pattern. Experiment 1 used a written paradigm, and Experiment 2 used an audio paradigm. Across experiments, we found that subjects were able to adapt flexibly to both referential patterns. However, this effect was limited: participants adapted to the second exposure condition only when they encountered a low‐frequency referential pattern first. These results suggest that participants are especially sensitive to the exposure they receive at the start of a new situation, and subsequent adjustments depend on whether the new pattern is consistent with lifetime processing biases. Our findings have implications for theories of implicit learning and adaptation.

1. Introduction

Language understanding requires comprehenders to make inferences, for example, judging which person is the more likely referent of “she” in “Ana saw Liz. She waved.” Even though “she” is ambiguous, adults tend to assume that it refers to Ana, the person mentioned as the subject of the prior sentence. This subject bias is one of several known discourse constraints that help people understand the speaker's meaning (e.g., Almor, ¹⁹⁹⁹; Arnold et al., ²⁰⁰⁰; Fukumura & Van Gompel, ²⁰¹⁵; Garvey & Caramazza, ¹⁹⁷⁴; Gernsbacher, Hargreaves, & Beeman, ¹⁹⁸⁹; Gordon, Grosz, & Gilliom, ¹⁹⁹³; Järvikivi, van Gompel, Hyönä, & Bertram, ²⁰⁰⁵; Kehler & Rohde, ^{2013, 2019}; Hartshorne, O'Donnell, & Tenenbaum, ²⁰¹⁵; Stevenson, Crawley, & Kleinman, ¹⁹⁹⁴).

To what extent can these discourse biases change on a situation‐by‐situation basis? For example, maybe I think that, in general, the subject is more likely to be rementioned (Arnold, 1998; Langlois, Zerkle, & Arnold, ²⁰²³), so this expectation guides me to assign pronouns to the subject when possible. But if George tends to frequently structure his stories so that he rementions the object frequently, would I shift my expectations when talking to George? Research shows that yes, people do adjust expectations based on referential patterns in recent experience (e.g., Contemori, ²⁰²¹; Fernandes, Luegi, Correa Soares, De La Fuente, & Hemforth, ²⁰¹⁸; Johnson & Arnold, ²⁰²³; Ye & Arnold, ²⁰²³). But an unanswered question is how specific this adaptation is: if I then talk to Martha, who instead tends to remention subjects, do I quickly shift back to my baseline biases?

Answering this question is critical to understanding the extent to which adults track the statistics of language and use these statistics to modulate processing biases. For example, we know that reading or hearing a new syntactic structure increases the chance that speakers will use the structure in production (Bock, 1986; Pickering & Branigan, ¹⁹⁹⁸) and that comprehenders become more likely to select that structure (Fine, Jaeger, Farmer, & Qian, 2013; Fraundorf & Jaeger, ²⁰¹⁶). Such syntactic adaptation can persist over many trials (Bock & Griffin, 2000) or even many days (Kaschak, Kutta, & Schatschneider, 2011; 2014). We also know that comprehenders can adapt to the most frequent referential patterns in the recent context (e.g., Johnson & Arnold, ²⁰²³). However, it is not clear how adaptation changes in the presence of conflicting input. Here, we examine this question with respect to referential adaptation.

Our focus is on questions about how people respond to different kinds of exposure over time, a phenomenon typically called adaptation (Chang, Dell, & Bock, 2006, 2012), cumulative priming (Kaschak et al., 2011), or statistical learning (Weiss, Gerfen, & Mitchel, 2009). Critically, existing theories either do not make explicit predictions about the impact of conflicting inputs, or they make different predictions. Given that the literature on referential adaptation is sparse, we consider predictions based on theories about structural priming, and related phenomena like adaptation and statistical learning. To understand how adaptation changes in response to new inputs, we consider the following four related questions and the predictions from existing theories.

How long does adaptation last ? Certain aspects of structural priming are known to be relatively short‐lived (Bock & Griffin, 2000). The classic example of a short‐lived effect is the lexical boost, in which priming is enhanced by overlapping wording choices in the prime and target sentences (Pickering & Ferreira, 2008; Kaan & Chun, ²⁰¹⁸). But other evidence shows that priming can also last for a long time. Research on sentence production shows that people produce structures they have recently encountered (Bock, 1986), and this type of priming persists after a block of 10 intervening sentences (Bock & Griffin, 2000; Bock, Dell, Chang, & Onishi, ²⁰⁰⁷), or even a week later (Kaschak et al., 2011; 2014). Contemori, Mossman, and Ramos (2022) found referential priming a week after initial exposure as well.

A significant challenge for theoretical accounts of structural priming is faithfully reproducing both short‐term and long‐term priming phenomena (Xu & Futrell, 2024). The residual activation mechanism accounts for short‐term structural priming using residual activation of the nodes representing a syntactic structure. Recently‐encountered structures are activated during the production or comprehension of a new sentence. This account predicts the short‐term persistence of recently‐encountered material (Bock & Griffin, 2000; Bock et al., ²⁰⁰⁷).

Another theoretical approach uses evidence for the longevity of structural adaptation to suggest that it is a form of implicit learning (Bock & Griffin, 2000; Chang et al., ²⁰⁰⁶; Chang, Janciauskas, & Fitz, ²⁰¹²; Kaan & Chun, ²⁰¹⁸). Error‐based implicit learning accounts specifically propose a mechanism based on prediction in which the language system is recalibrated in proportion to error signals, predicting more robust priming for surprising input (Chang et al., 2006). A related approach is hybrid models, which use two distinct mechanisms to account for short‐lived and long‐lasting effects. Implicit learning mechanisms model cumulative, long‐lasting priming effects, and spreading activation‐based mechanisms account for short‐lived effects. Bayesian models provide another account for long‐lasting priming effects. They separate short‐term and long‐term experience with linguistic patterns through the estimation of a prior, which corresponds to long‐term experience, and a continuously updated posterior distribution that allows the learner to rapidly shift their expectations in response to new input (Jaeger & Snider, 2013; Kleinschmidt, Fine, & Jaeger, ²⁰¹²).

How does local adaptation relate to prior biases? Work on syntactic priming has demonstrated that people respond more robustly to low‐frequency primes, a phenomenon termed the “inverse‐frequency effect.” For example, priming effects elicited by prepositional object datives (“The man gave the treat to the dog”) are more robust than the effects from double object dative structures (“The man gave the dog the treat”) for English speakers because the prepositional object structure is less frequent (Kaan & Chun, 2018). Similarly, Roy, Johnson, and Arnold (2025) found that referential adaptation also exhibited an inverse frequency effect. Participants were exposed to subject‐reference (the predicted condition), nonsubject‐reference (the unexpected condition), and a neutral condition with no anaphoric reference. Exposure affected pronoun comprehension only for the nonsubject‐exposure condition, while the subject‐ and neutral‐exposure conditions were not different.

One explanation for this pattern comes from error‐based learning models, which suggest that adaptation results from encountering surprising input. The system predicts the more frequent structure, and when a less‐frequent structure is encountered, it generates an error signal. The adjustment of expectations on new input is proportional to the strength of the error signal. Thus, the prediction of an inverse frequency effect is the hallmark of error‐based learning theories (Fine et al., 2013; Jaeger & Snider, ²⁰¹³). One type of error‐based learning can be found in Bayesian models, which attempt to optimize prediction and utilize prediction error. Bayesian accounts of implicit learning do not assume an error‐based mechanism. Prediction error is often represented using Bayesian surprise, the divergence between an individual's prior and posterior distributions at a given level of representation. These distributions are continuously updated as new information enters the system. A key distinction is that in Bayesian accounts, the changes to individuals’ expectations that are prompted by prediction error are weighted according to the reliability of prior information (Kuperberg & Jaeger, 2016). Reliability of linguistic cues can be estimated from lifetime language usage statistics, which are combined with the local language input to determine an efficient response (Kleinschmidt et al., 2012; Kleinschmidt & Jaeger, ²⁰¹⁵).

A different explanation for the inverse frequency effect comes from hybrid models through the base level learning function. Each retrieval from memory increases a structure's activation by a similar magnitude. For frequently used structures, each new retrieval has a proportionally small priming effect. But infrequently used structures are primed to a proportionally larger degree by each retrieval. As a consequence, hybrid models can account for inverse frequency effects (Reitter et al., 2011).

How does conflicting input guide adaptation? A critical question about adaptation is how people respond to conflicting input. Intuitively, repeated exposure to the same linguistic pattern should lead to stronger expectations for that pattern, while exposure to a mix of conflicting patterns should result in weaker adaptation. Initial evidence suggests that this pattern obtains. Jaeger and Snider (2013) found that the strength of priming depends on the relative proportion of PO versus DO structures. Likewise, evidence suggests that referential adaptation is weaker with mixed input within the same block (Arnold, in prep). These effects are consistent with error‐based accounts (including the Bayesian account), where prediction error changes over time based on new inputs. This implies that conflicting inputs should cancel each other out, weakening priming effects for mixed input.

However, it is not known what window of recent experience is considered for calculating either structural or referential frequency (the last few minutes, hours, months, years?) We also do not know how referential or structural adaptation changes over time if the frequency of input changes. Imagine that I adapt to George using lots of nonsubject‐reference patterns, and then switch to talking to Martha, who uses frequent subject‐reference patterns. Will I immediately recognize the new pattern and shift my expectations, or will I continue to follow George's pattern of speech?

Evidence on this question is indirect. Artificial language learning studies suggest a primacy effect: people prioritize the first of two artificial languages they encounter (Gebhart, Aslin, & Newport, 2009). But to our knowledge, no studies have explicitly tested how adaptation changes in a context where the frequency of input differs across blocks of input. Meta‐analyses of syntactic priming and adaptation (Mahowald, James, Futrell, & Gibson, 2016; Kumarage, Donnelly, & Kidd, ²⁰²⁴) suggest that most studies measure the effect of priming shortly following exposure to the prime. A less common approach to assessing syntactic priming is examining the cumulative priming effect that accrues after a participant is exposed to a syntactic structure multiple times over the course of an experiment. A subset of these cumulative priming studies present a block of stimuli containing a single type of prime and measure the effect of exposure to the block. Only a very small number of studies have examined complex, conflicting input using a blocked design (Fine et al., 2013; Jaeger & Snider, ²⁰¹³). Neither of these studies manipulated both priming condition and order of priming conditions across blocks. Thus, there is a gap in the literature regarding the effect of changing exposure patterns within a single experiment. Our study aims to close this gap.

Is learning generalized over all inputs or conditioned on specific contexts? The evidence indicating that adaptation is a long‐term phenomenon is comprehensive (Bock & Griffin, 2000; Contemori et al., ²⁰²²; Kaschak et al., ^{2011, 2014}; Kroczek & Gunter, ²⁰¹⁷). However, it seems implausible that encountering a few extra double‐object primes in an experiment will change one's expectations forever. In fact, evidence for the longevity of priming effects may instead reflect situation‐specific learning. Some research finds evidence of structural adaptation in language production a full week after initial exposure to the syntactic structure (Kaschak et al., 2011, 2014). Given the sheer quantity of language input a participant would encounter in the course of a week, the exposure effect from an experiment would be too minute to have any impact, unless the participant is learning something more specific. One possibility is that they are learning about the probability of syntactic structures in the context of the experiment or the specific task used in the experiment (Kaschak, Kutta, & Coyle, 2014).

The role of situation‐specific adaptation has not (to our knowledge) been directly addressed in the literature on syntactic or referential priming in comprehension. This is a theoretically important issue because evidence that implicit learning can be context sensitive is accumulating (Kamide, 2012; Kroczek & Gunter, ²⁰¹⁷; Nielsen & Christiansen, ²⁰²⁵), and models differ in the extent that they can capture this effect. Error‐based learning models do not offer an obvious mechanism to account for the role of context (speaker, location, etc.). By contrast, Bayesian accounts can explain speaker‐ and context‐specific adaptation effects as a consequence of reliable cooccurrence between these situational cues and linguistic structures (Jaeger & Snider, 2013). The point that contextual representations are an integral part of the dynamics of memory encoding and retrieval is well known to memory researchers (Tulving & Thomson, 1973) and is sometimes referred to as the “encoding specificity principle.” However, this principle has not been explicitly applied to linguistic phenomena like structural priming and adaptation until recently (Nielsen & Christiansen, 2025).

Evidence from other domains suggests that people do link calculations of linguistic frequency to specific situations or talkers. First, in pronoun production, exposure to pronouns increases pronoun production, but only if the exposure occurs at the beginning of the experiment (Arnold, 2025). This suggests a situation‐specific effect. Second, the statistical learning of artificial languages suggests that people learn the language they encounter first in a blocked design better than the one they encounter second (Gebhart et al., 2009), although speaker cues can facilitate learning multiple patterns at the same time (Weiss et al., 2009). Third, dialogue studies show that speakers and comprehenders entrain on terms, coming to expect that a partner will use the same label they previously used (e.g., penny loafer vs. shoe; Brennan & Clark, ¹⁹⁹⁶; Brown‐Schmidt, ²⁰⁰⁹; Horton & Gerrig, ²⁰⁰⁵). Similarly, comprehenders can learn different structural probabilities for specific talkers (Kamide, 2012; Kroczek & Gunter, ²⁰¹⁷).

However, there are no studies that explicitly test how adaptation in comprehension changes in response to changing input frequencies, either for syntactic or referential adaptation. Here, we focus on referential adaptation, using a blocked exposure paradigm to examine how adaptation changes over time.

1.1. Does referential adaptation respond flexibly to changes in the input?

Evidence shows that people adapt to the referential patterns they encounter in an experiment. Johnson and Arnold (2023) tested pronoun interpretation biases with stories like “Matt went skiing with Will. He wore the blue hat,” where they probed the interpretation of the pronoun with questions. Critically, participants also read numerous stories where either the subject was always rementioned with a pronoun (Ana played music with Matt. She played the piano.) or where the second character was always rementioned (… He played the piano.) People adapted to the pattern they were exposed to, so the pronoun was assigned to the subject more often in the subject‐exposure than nonsubject‐exposure conditions. This shows that people are sensitive to the frequency of referential patterns in the local context (see also Arnold, ²⁰²⁶; Contemori, ²⁰²¹; Contemori et al., ²⁰²²; Fernandes et al., ²⁰¹⁸; Kaiser, ²⁰⁰⁹; Kesan & Arnold, ²⁰²⁵; Roy & Arnold, ²⁰²⁵; Ye & Arnold, 2023a,b).

Evidence for referential adaptation is consistent with the idea that discourse biases stem from a sensitivity to the most frequent referential patterns in language. It is well established that by adulthood, there are population‐wide biases that are robust across studies, for example, the bias to assign pronouns to the subject. Arnold (1998, 2010; Arnold, Strangmann, Hwang, Zerkle, & Nappa, ²⁰¹⁸) suggests that people come to expect the most likely references based on remention frequencies; for example, subjects are more expected than other referents (Arnold, 1998; Arnold et al., ²⁰¹⁸). Another possibility is that people calculate the likelihood that a pronoun has a particular type of referent, for example, the frequency that pronouns refer to subjects. A third possibility is that comprehenders calculate the frequency with which pronouns are used in production for subjects, and integrate this with meaning‐based expectations (Kehler & Rohde, 2013; 2019). Regardless of which statistic is used, we can assume that people develop biases over a lifetime of learning. We can call these “lifetime statistics” about reference.

Thus, there are two ways to think of referential adaptation. One is a lifelong process whereby people continue to update their statistics about language. However, this cannot be the only explanation. If people were simply averaging all statistics together, an experiment with 40 exposure items (as in Johnson & Arnold, ²⁰²³) would represent a minute proportion of lifetime input, and would not shift interpretations noticeably (see Ye & Arnold, ²⁰²³, for this argument). A second possibility is that adaptation is a local adjustment to make interpretation more efficient, over and above any lifetime learning that may occur. People might have general biases like the subject bias, but be able to accommodate variations across speakers or across situations, making their expectations more situation‐specific and more precise.

Findings suggest that local adaptation occurs, but does not overturn prior biases. For example, Johnson and Arnold (2023, exp. 1a) observed that people assigned the pronoun to the subject 84% of the time in the subject‐exposure condition, but 69% of the time in the nonsubject‐exposure condition. If people only followed recent exposure (which was always either 100% subject‐ or nonsubject‐exposure), they should have exhibited ceiling and floor responses. Instead, they saw only a small shift, suggesting that people are integrating lifetime and recent statistics.

1.2. The current study

The current study seeks to understand the effect of referential exposure in contexts where people encounter repeated inputs with different referential structures. Using a paradigm like Johnson and Arnold's, we exposed people to either subject‐ or nonsubject‐antecedent reference structures, organized in blocks. Thus, if participants see subject‐exposure in the first block, they then move to a second block of nonsubject‐exposure. In another list, participants see nonsubject‐exposure in the first block, and subject exposure in a second block (see Fig. 1). We predict that block 1 will exhibit the standard adaptation effect. The critical question is what happens in block 2.

Fig. 1.

List and block design for Experiments 1 and 2.

Possibility 1: Recency. One possibility is that people respond strongly to the most recent input. At the start of block 2, we expose people to several instances of the new pattern, so by the time the test items occur, the new pattern is established. If recent input drives understanding, we should see a clear impact of the current exposure pattern (i.e., a reversal of the responses in block 1).

This pattern would be predicted by a mechanism like residual activation, which predicts that subjects should provide responses that align with the most recently perceived referential pattern. This activation should be short‐lived, and the remaining activation is likely to bias subjects’ interpretation of ambiguous critical items, making the reuse of the most recently encountered referential pattern more probable. This would influence which character is likely to be selected as the antecedent for “he” or “she.” However, the residual activation approach most strongly accounts for the increased priming associated with word repetition, and referential adaptation effects do not include repeated words. Thus, it is not clear whether this account would apply to referential adaptation.

A recency effect might also be predicted by a hybrid model, which predicts a recency effect along with an inverse frequency effect and other long‐term priming phenomena. A distinctive aspect of the base‐level learning mechanism is that it is not prediction‐oriented, it is based on memory retrieval. A structure that is retrieved repeatedly from memory should continue to become more active and, accordingly, is more likely to be reused. The hybrid model would also be associated with recency effects through its spreading activation mechanism that is used to model short‐term priming phenomena (Reitter et al., 2011).

Possibility 2a: Simple primacy effect. Another possibility is that we could see a bias to follow the pattern encountered at the start of the experiment. Once people adapt to the pattern in block 1, they may perseverate and continue to follow this pattern in block 2. This predicts a pattern of primarily subject pronoun resolutions across both blocks in the subject exposure first list and primarily nonsubject pronoun resolutions across both blocks in the nonsubject exposure first list. This pattern is not common in the implicit learning of syntactic structures. But in the statistical learning of artificial languages, a common outcome is for participants to learn the first pattern they are exposed to, but not a second pattern, unless the transition to the second pattern is cued or the exposure time for the second pattern is much longer than for the first pattern. It has been argued that this primacy effect occurs because in natural language use, rapid shifts between linguistic structures are typically accompanied by co‐occurring cues. Without these transitional cues, subjects could act conservatively and respond based on the first pattern they disambiguated (Gebhart et al., 2009).

Possibility 2b: Contingent primacy effect. Alternatively, a primacy effect could be contingent on whether the initial pattern matches baseline preferences from a lifetime of input. If the initial pattern is the same as the usual pattern in other contexts (i.e., the subject‐exposure condition), this provides a strong cue that the current situation is “normal” and participants can apply their usual expectations to the current situation, regardless of new input. But if the initial pattern is surprising (the nonsubject‐exposure condition), it may create a fragile shift in expectations that may be overturned as soon as evidence supports a return to the normal pattern. Under this view, subjects would respond to the pattern that is presented in the first block, but would only maintain this same pattern of responses into the second block if the first block presented subject‐antecedent exposure. This is the pattern that would be consistent with participants’ lifetime biases. This pattern is not explicitly predicted by any existing models, although it may be accounted for by a modified version of Bayesian accounts. We will return to this point in the General discussion.

Possibility 3: Averaging across blocks. Alternatively, participants may integrate the second block input with the first block input, potentially yielding an average response. For example, subject exposure in the first block would elicit a high rate of subject interpretations; when nonsubject exposure is encountered in the second block, participants might average their predictions and generate a moderate level of subject interpretations. The same would occur for participants starting with nonsubject exposure in reverse, such that both lists would elicit the same level of moderate subject interpretations in the second block. It is not clear that any existing models predict this outcome, but we consider it as a logical possibility.

Possibility 4: Surprisal—Larger difference between priming conditions in block 2. Another possibility is that exposure effects are strongest when the input is most surprising. Subjects form expectations about referential patterns based on their lifetime experience. In general, subject rementions are more probable and nonsubject rementions are less probable (Arnold, 1998; Johnson & Arnold, ²⁰²³). In theoretical frameworks like error‐based learning and Bayesian belief update, the learner predicts upcoming language. In a typical situation, the learner predicts that an upcoming reference refers to the subject of a prior sentence. If the learner encounters a nonsubject remention instead, this runs counter to their predictions and should be surprising. This registers as a prediction error, and the degree of unexpectedness associated with the nonsubject‐remention may be formalized using surprisal.

Error‐based learning models predict that the most robust adaptation effects should occur when this unexpectedness is strongest (Chang et al., 2006) or similarly for the most surprising referential patterns (Roy et al., 2025). Bayesian accounts make a similar prediction. An implication is that adaptation effects should be proportional to the size of the prediction error. This mechanism predicts the strongest adaptation effects in block 2 of the subject exposure first list. The exposure condition changes from the expected subject pattern (“Elaine and David were cleaning the car. Elaine soaped up the car with David and then she turned on the hose.”) to an unexpected nonsubject pattern (“Liz and Matt were sitting on the porch. Liz watched the sunset with Matt and then he put on some sunglasses.”). On this model, participants should experience prediction error, as they should still be anticipating subject reference referring to Liz (“…, and then she put on some sunglasses.”). This effect may even be greater than the effect of nonsubject exposure in the first block, because the expectation for subject remention is even stronger after an initial block of subject‐exposure.

To assess these predictions, we test the ability of participants to adapt to multiple referential patterns in two experiments. Experiment 1 used a written presentation of stories, while Experiment 2 used an auditory presentation. The auditory modality allowed us to use the speaker's voice as a signal that the second block constituted a new discourse situation, similar to talking to George versus Martha. It also allowed us to test whether the exposure effect extends to auditory experiments.

1.2.1. Sample size, preregistration, and data

Experiment 1 was not preregistered, but it followed the goal of 60 usable participants based on the grant proposal for this project. Experiments 2a and 2b used power analyses (described below) to determine the sample sizes of 80 (Experiment 2a) and 100 (Experiment 2b), which were preregistered (Experiment 2a https://osf.io/t4auv; Experiment 2b https://osf.io/re5w8). The data for all experiments are available at OSF (https://osf.io/a2kyz/overview?view_only). All studies were approved by the institutional review board. Informed consent was obtained from all participants in all of the reported studies before the procedure began.

2. Experiment 1

2.1. Methods

2.1.1. Participants

Participants in Experiment 1 were 74 native English speakers recruited using Amazon Mechanical Turk and were compensated for their participation in the study with $2.50. An additional 81 were excluded based on the inclusion criteria below.

Inclusion criteria. We used rigorous inclusion criteria to make sure that our crowd‐sourced participants were attentive humans. Participants had to be at least 18 years of age. The HIT was only available for participants located in the United States, the United Kingdom, or Australia. The survey automatically excluded individuals who reported their English proficiency as anything less than “Native” or reported beginning to learn English after the age of 6. The first 33 unambiguous questions were used to ensure people were reading the stories and questions; people with less than 75% accuracy were not allowed to continue the survey. Of the participants who completed the survey, we excluded from analysis those who had less than 75% accuracy on the unambiguous questions across the whole survey (n = 1) and those who reported having a language disorder (n = 3). In addition, we used the author recognition task (ART; Moore & Gordon, ²⁰¹⁵) as a measure that people were following directions, following the protocol developed in previous studies (e.g., Johnson & Arnold, ²⁰²³). Participants were asked to indicate which author names they were familiar with out of a list of real and fake author names, and they were instructed not to guess. Participants who were guessing on this task (signaled by providing 33% or more incorrect answers) were excluded from analysis (n = 54).¹ An additional 23 participants were excluded for not meeting more than one of the preceding criteria.

2.1.2. Materials and design

The experiment was created in Qualtrics and included a section for demographic questions, the ART, and then the main experimental task. The demographic questions asked participants about their age, education, language background, biological sex, gender identity, ethnicity, race, and whether they had ever been diagnosed with a language disorder.

The primary experimental task had participants read short stories featuring four characters (Ana, Liz, Matt, and Will). We manipulated whether participants read subject exposure stories (“Matt and Ana were visiting a pet store. Matt pet the cat with Ana and then he pet the dog.”) or nonsubject exposure stories (“Matt and Ana were visiting a pet store. Matt pet the cat with Ana and then she pet the dog.”). Underlining is used here to signal the conditions and was not present in the Qualtrics stimuli. Each set of exposure stories was presented in its own block, but from the participants’ perspective, the items were in a single survey with no signal of the second block. The order of the blocks was counterbalanced across the two lists. The effect of the subject and nonsubject exposures was measured using participants’ interpretation of ambiguous stories (“Ana and Liz were getting ready to go shopping. Ana brought the coupons to Liz, and then she wrote a shopping list.”). Participants’ interpretation of the story was probed using questions (“Did Liz write a shopping list?”) that asked participants about their final interpretation of the ambiguous pronoun, here “she.”

The main task began with two practice stories. Within the experiment, the stories were divided into two blocks. Each block began with a set of four training stories to establish the exposure manipulation. Then, eight ambiguous critical items appeared, interspersed with 10 more unambiguous exposure stories for the rest of the block; a total of 44 stories. The exposure items consisted of two sentences where the first sentence established a context for the two characters. Then, in the second sentence, the two characters were described as carrying out an action together; the first clause described an action in the “joint action” structure (“X did something with Y”), followed by “and then” and a clause with a pronoun. The story and questions were on separate pages.

Participants were asked two multiple‐choice questions about each story for a total of 88 questions: one about the story's content (comprehension question) and one that probed pronoun interpretation (reference question). The comprehension questions asked about what the characters were doing, who performed an action, where the characters were, or when an event from the story occurred. Each question collected responses using a multiple‐choice format with two response options.

The pronoun interpretation questions (“reference questions”) for the exposure items were unambiguous based on the gender of the pronoun and the gender of the characters in the story. The exposure reference questions asked about either the subject character or the nonsubject character, which determined whether the correct answer was “yes” or “no.” For the critical items, the two characters had the same gender, so the pronoun was ambiguous. All of the critical items used a source‐goal verb in the second sentence, rather than using a joint action structure. Critical items also only used one question type for the referential questions, asking “Did [name] do [x]?” The question always asked about the second person, so if people had assigned the pronoun to the subject, they would answer “no.” See Table 1 for examples of the items and various question types. Exposure items are shown in the subject‐exposure condition; the nonsubject‐exposure stories were identical except that the pronouns always referred to the second character. The key manipulation was in the ordering of the exposure conditions. In the subject‐first list, block 1 contained exposure to subject pronoun resolutions, and block 2 used nonsubject pronoun resolutions. In the nonsubject‐first list, the order of exposure conditions was reversed such that nonsubject exposure occurred in block 1 and subject exposure occurred in block 2 (see Fig. 1).

Table 1.

Item types used in Experiment 1

Item type	Question type	Example
Critical item	Reference (no): Did Liz write a shopping list? Content (What?): What did Ana bring to Liz?	Ana and Liz were getting ready to go shopping. Ana brought the coupons to Liz, and then she wrote a shopping list.
Exposure item	Reference (yes): Did Matt pet the dog? Content (Who?): Who did Matt pet the cat with?	Matt and Ana were visiting a pet store. Matt pet the cat with Ana and then he pet the dog.
Exposure item	Content (What?): What were Liz and Will doing? Reference (no): Did Will put the cake in the fridge?	Liz and Will were shopping for a birthday party. Liz bought a cake with Will and then she put it in the fridge.
Exposure item	Content (Who?): Who did Liz paint a picture with? Reference (yes): Did Liz frame the picture on the wall?	Liz and Matt were decorating the new apartment. Liz painted a beautiful picture with Matt and then she framed it.
Exposure item	Content (Where?): Where were Matt and Ana going? Reference (yes): Did Matt let the dog off the leash?	Matt and Ana were going to the park. Matt walked the dog with Ana and then he let the dog off of the leash.

2.1.3. Procedure

Participants on Amazon Mechanical Turk were redirected to the Qualtrics survey using a link. Before beginning the Qualtrics survey, the participant provided their consent to take part in the procedure. Participants were also provided with an explicit warning that they would not be compensated for their participation if they made too many mistakes. Participants responded to the demographic questions, the ART, and the main task that included both priming blocks. At the end of the procedure, the participant was given a random code to use to validate experiment completion and receive payment.

2.1.4. Analytical approach

The data were analyzed with mixed effects logistic regression models, using the lme4 package in R (Bates, Mächler, Bolker, & Walker, 2015). The outcome variable was whether the participant selected a subject or nonsubject interpretation for the pronoun (Subject = 1; Nonsubject = 0). We included a predictor variable called “Block” that represents whether the response occurred in block 1 or block 2 (Block 1 = 0.5, Block 2 = −0.5) and a variable that represented the list condition (List), which represents the order of exposure condition blocks (Subject first (subj‐non) = 0.5 and Nonsubject first (non‐subj) = −0.5), plus the interaction of block and list. Significant interactions were further probed by testing estimated marginal means in the emmeans package (Lenth et al., 2025). To begin, the model included the maximal random effect structure justified by the design. This included random slopes for block in the by‐subject random effects and random slopes for list in the by‐item random effects. The structure was simplified² by removing the random slopes that did not contribute significantly to model fit. This procedure was repeated until the model converged and was no longer overfitted. The statistical significance of the random effects structure was tested using the likelihood ratio test until only random effects that contributed significantly to model fit were retained (Matuschek, Kliegl, Vasishth, Baayen, & Bates, 2017). The final model included random intercepts for participants and items and random slopes for block by participants.

2.2. Results

The critical question in Experiment 1 was whether participants would respond flexibly to shifts in referential patterns within a single experimental session. We found that they did, but only in the nonsubject‐first list, when the procedure began with nonsubject exposure. As shown in Fig. 2, the first block revealed the expected exposure effect: more subject interpretations in the subject (Subj first: M = 59.5%, SE = 0.029) than the nonsubject exposure condition (Nonsubj first: M = 39%, SE = 0.028). But block 2 exhibited similar response in the subject‐ (Nonsubj first: M = 55%, SE = 0.029) and nonsubject‐exposure (Subj first: M = 57.1%, SE = 0.029) conditions. That is, people who started with a nonsubject block had a lower rate of subject interpretations in block 1. Then, they responded to the change in exposure condition in block 2 and began to produce a higher rate of subject interpretations. However, for people who started with the subject condition, there was no change in block 2, and they continued to use a high rate of subject interpretations.

Fig. 2.

Subject responses in Experiment 1. Note. subj (solid bars) = Subject exposure condition; NON (lattice bars) = Nonsubject exposure condition. The red column (left side) represents the first block, and the blue column (right side) represents the second block.

This pattern was supported by the results of the regression model (Tables 2–4, ³). The key finding was a significant interaction between List and Block; no additional main effects were observed. In order to probe the significant interaction between these variables, we used estimated marginal means generated based on the fitted regression model. Then, we used pairwise contrasts to test whether the difference between lists was significant, holding each block constant. We observed a significant difference in block 1 between the subject‐first and nonsubject‐first, but not in block 2 (Table 3). A second set of pairwise comparisons checked for differences between the first block and the second block in each list condition (Table 4). For the nonsubject‐first list, the difference between block 1 and block 2 was significant, but not for the subject‐first list.

Table 2.

Mixed effects logistic regression model results for Experiment 1

Fixed effects	Estimate	Standard error	z‐statistic	p‐value
(Intercept)	0.166	0.342	0.485	.6280
List (subj first vs. nonsubj first)	0.932	0.604	1.543	.1229
Block	−0.525	0.382	−1.375	.1690
List x Block	1.246	0.421	2.958	.0031 **

Signif. codes: 0 ≤ “***” < .001 < “**” < .01 < “*” < .05.

Table 4.

Pairwise contrast—Estimates for effect of block by list

Block pairwise	List	Estimate	SE	z‐statistic	p‐value
Block 2 – Block 1	List 1 (Subj first)	−0.098	0.438	−0.223	.823
Block 2 – Block 1	List 2 (Nonsubj first)	1.15	0.435	2.64	.008 **

Table 3.

Pairwise contrast—Estimates for effect of list by block

List pairwise	Block	Estimate	SE	z‐statistic	p‐value
List 2 (Nonsubj first)‐List 1 (Subj first)	1	−1.56	0.595	−2.615	.009 **
List 2 (Nonsubj first)‐List 1 (Subj first)	2	−0.309	0.682	−0.454	.65

2.3. Discussion

This outcome extends previous results showing that speakers adapt to the referential patterns encountered in an experimental session (Johnson & Arnold, 2023). Block 1 replicated this adaptation effect. Further, the significant contrast in blocks for the nonsubject‐first list showed that speakers are capable of tracking and responding to multiple different patterns of pronoun use within the same session. This shows that yes, participants can respond flexibly to changing referential patterns, at least under some circumstances.

However, a second key finding was that the order of exposure matters. In the subject‐first list, encountering subject exposure first seemed to anchor the participants’ expectations, making them less likely to respond to a change in the exposure condition in block 2. By contrast, when participants received unexpected input in the first block, they did not perseverate in the second block. One possibility is that the unexpected pattern was unstable, given its inconsistency with lifetime statistics. Another possibility is that it allowed the participants to maintain a sense of wariness for other possible shifts during the session.

These results are not consistent with a simple recency effect, simple primacy effect, or averaging across both blocks. They also do not exhibit a stronger effect of exposure in block 2. Instead, results reflect a contingent primacy effect—when the dominant subject‐exposure pattern comes first, people perseverate and follow this pattern in the second block, but when the infrequent nonsubject‐exposure pattern comes first, they switch to following the dominant pattern in the second block. However, this pattern is not directly predicted by any existing models.

Given this surprising finding, we aimed to replicate it in Experiment 2. An initial attempt to replicate this finding with the same paradigm produced a similar pattern of results, but the key interaction did not reach statistical significance (for the results of this study, see Supplementary Appendix A). This experiment provided tentative support for the observed pattern, but it raised questions about what experimental conditions would be most likely to reduce noise and provide a robust test of how responses change over blocks.

We hypothesized that flexible adaptation should be supported by signals that a new discourse segment has begun. The experiment itself is an example of a discourse situation, so it seems likely that participants may attend to the input at the start of the experiment in order to calculate the locally relevant statistics. Once learned, they do not need to attend to the input as much. We hypothesized that participants may be more likely to attend to the input if the second block appears to represent a new scenario—for example, having a conversation first with Martha and then moving to a new conversation with George. We, therefore, designed an experiment where the second block was signaled by changes in (a) who was speaking and (b) the characters being discussed.

3. Experiment 2

Experiment 2 tested whether people shift their adaptation when the input shifts in a second block, using a paradigm that signals that the second block represents a new discourse situation. Each new discourse situation should be associated with its own set of expectations about language style and the frequency of structural patterns. Discourse situations also tend to recur with particular individual speakers. This implies that one probabilistic cue to discourse context in natural language is a consistent interlocutor—that is, are you speaking with the same person or a new person? A second cue is the characters under discussion. The combination of both a new interlocutor and a new topic is particularly likely to signal a new situation. We, therefore, used a spoken version of our task, where blocks 1 and 2 had different speakers, and where the names of the characters discussed were different.

Before testing our primary question about blocked priming in Experiment 2b, we first confirmed that the basic referential adaptation effect would extend to a spoken paradigm in Experiment 2a. Experiment 2a is a conceptual replication of the subject‐ and nonsubject‐exposure conditions from Roy et al. (2025), using auditory stimuli.

3.1. Experiment 2a

3.1.1. Methods

3.1.1.1. Participants

The sample size for Experiment 2a was determined based on a power simulation using simr (Green & MacLeod, 2016), analyzing the data from the subject‐ and nonsubject‐exposure conditions from Roy et al. (2025, exp. 2). This analysis indicated 80% power with 70 subjects. Given that we did not know whether the audio‐based effect size would be the same as the effect size in our text‐based paradigm, we used a slightly larger sample with 80 subjects.

Participants were 80 English native speakers recruited via Prolific.⁴ A further eight participants were excluded from the final data set based on the inclusion criteria set out below. Participants were compensated for approximately 20–25 min of participation with $5.00.

Inclusion criteria. Using Prolific's prescreen options, we limited participants to those with (1) an IP address located in the United States, the United Kingdom, or Australia; (2) English as their first and primary language; (3) no language disorders; and (4) normal or corrected to normal vision and hearing.

Our preregistered standards were that individuals who correctly answered less than 75% of the “bot” (attention check) questions would be screened out immediately (n = 1). Subjects who correctly answered less than 75% of the unambiguous exposure questions (n = 5) and 50% of the literacy questions (n = 0) were excluded from the final data set following the procedure described below. Two individuals were excluded because they reported being “good” proficiency, rather than native proficiency, speakers of English.

3.1.2. Materials and design

Experimental stimuli were 50 stories from Roy et al. (2025), which were recorded by the first author. Participants responded to 68 questions about these stories. The structure for these stories was identical to the structure used in Experiment 1. Unambiguous sentences included a pronoun that was disambiguated based on the gender of the pronoun and the referent. Critical items included ambiguous pronouns such that the subject or nonsubject interpretation was plausible (“Liz and Ana were going to a concert. Liz offered the tickets to Ana, and then she waited in line.”). All the critical items used source‐goal transfer verbs. Each of these stories was accompanied by a question that required the participant to resolve the ambiguous pronoun. These questions had three variations (see Table 5): half of the critical items had “who” questions (Who waited in line?), 25% had Subject (“yes”) questions, and 25% had nonsubject (“no”) questions. Subject questions asked about the first‐mentioned character (Did Liz wait in line?), so the “yes” response indicates that they assigned the pronoun to the subject character. Nonsubject questions asked about the second character (Did Ana wait in line?), so “no” reflected a subject interpretation and “yes” reflected assignment of the pronoun to the nonsubject character. Half of the critical stories included a second reading comprehension question (e.g., “Where were Liz and Ana going?”).

Table 5.

Example items for Experiment 2a

Item type	Example
Exposure item (Subject)	Liz and Will were shopping for a birthday party. Liz bought a cake with Will and then she put it in the fridge.
Exposure item (Nonsubject)	Liz and Will were shopping for a birthday party. Liz bought a cake with Will and then he put it in the fridge.
Critical item	Liz and Ana were going to a concert. Liz offered the tickets to Ana, and then she waited in line.
Bot question	Ana and Will were practicing their dance moves.Ana twirled around with Will and then she collapsed in exhaustion.Who collapsed in exhaustion? Ignore the question, just select this if you're human Ana Will
Literacy question	Liz and Ana were having a conversation about their taxes. Liz said “Pardon my French, but…”What was Liz going to say next? Choose the best answer. A swear word Hello in French

Instead of the ART, we included a set of questions intended to detect nonhuman participants (bot questions) and a set intended to identify non‐native English speakers (literacy questions). Bot questions served as simple attention checks, for example, one answer might say “Ignore the question, just select this if you're human.” Literacy questions called on more difficult inferences or English language idioms (e.g., Liz retrieved the chips from Will and then he said he felt under the weather. How was Will feeling?).

Participants also responded to 28 unambiguous exposure items, which all used what we call “joint‐action” constructions (X did something with Y). Ten of the exposure stories had two questions (38 questions total). In eight of the exposure stories, the second question was a bot detection question. For the other two exposure stories with a second question, one second question was a comprehension question, and the other was a reference question. Participants were screened out immediately if they missed 3/8 bot questions and excluded from the final data set if they missed more than 3/6 literacy questions (see Table 5). There were also two practice items.

The recordings for each item were made using a Sennheiser headset microphone and Praat software. Recordings of the stimuli were saved as WAV files. The audio volume for all recordings was normalized to 70 dB in Praat. Critical stories, bot items, and literacy stories were identical across the two lists. Exposure items were spliced so that the story recordings across lists were identical except for the final clause. The two versions were initially recorded in their entirety, and then one was chosen to serve as the base. The final clause was cross‐spliced into the recording (“Liz and Matt were having lunch at McDonald's. Liz ate french fries with Matt {and then she spilled ketchup on the table.}” This involved cross‐splicing the final “and then” clause into the base recording for each of the exposure items {and then he spilled ketchup on the table.}. This controlled for potential variations in the delivery of the story that could have influenced participants. The selection of the base recording for the exposure items was rotated such that the number of intact and spliced recordings was balanced between the two lists.

Pauses within the recordings were standardized. The beginning and end of each recording included a pause of 300 ms. Pauses of 300 ms were also included between each sentence. A shorter pause of 180–220 ms was used between the end of the second sentence and the final “and then” clause. This variation in pause length was introduced to leave the recording as a whole sounding as natural as possible.

3.1.3. Procedure

Participants on Prolific were redirected to Qualtrics to take part in the survey. Participants consented to the procedure and provided their Prolific worker ID. Participants were informed that there were attention check questions in the experiment and that they would be dismissed automatically, without pay, if they made too many mistakes. Participants completed demographic questions next and proceeded through an audio check in which working audio playback was required to answer a multiple‐choice question. Then, they were instructed on how to go through the procedure using two practice items. Incorrect answers led to a repeat of the practice items. Participants went through the main survey next. At the end, they received a code for validating survey completion and were compensated with $5.00.

3.1.4. Trial structure

The trial structure was built in Qualtrics. HTML snippets were used to embed the audio files, and JavaScript was used to ensure that the survey advanced automatically after participants selected an answer for the multiple‐choice questions. Each trial began with a warning tone before the story audio began to play (see visualization in Fig. 3); the question audio followed automatically. The timers that switched between each page were fitted to the length of the recording. Stories ranged in length between 6 and 9 s. This was followed by a 1‐s pause and an additional warning tone that preceded the question audio. Next, the question was played; question recordings ranged from 2 to 4 s. After the question, participants had 10 s to select an answer.

Fig. 3.

Sequence of recordings in each trial for Experiment 2a.

Each question was multiple‐choice with two possible responses (three responses for bot questions). The placement of the answer choices was shuffled automatically. Once the participant selected an answer, the experiment automatically moved forward to the next page, and participants did not have the option of changing their answer except for the bot questions, where we allowed them to change their answer to ensure that they had the time to read all three of the answer choices. In bot questions, participants had the option of moving ahead after answering the question, once 3 s had elapsed. Individuals who failed to answer a question in 10 s received a warning (“Please respond more quickly next time. If you miss making a response on an excessive number of items you will be asked to return the experiment. You may not re‐start the experiment after that.”) We preregistered a plan to ask participants who failed to respond within 10 s three times to return the survey. However, 10 s is a relatively long period of time for answering a multiple‐choice question, and no participants were excluded based on this standard.

3.1.5. Analysis

The data analysis was performed with mixed effects logistic regression models using the lme4 package (Bates et al., 2015) in R. The outcome measure was whether the participant selected a subject interpretation or a nonsubject interpretation for the pronoun in the critical items (coded subject = 1; nonsubject = 0). The predictor variables were exposure type (Subject = 0.5/Nonsubject = −0.5), Subject question (Subject, 0.5 vs. Who/Nonsubject, −0.5); Nonsubject question (Nonsubject, 0.5 vs. Who/Subject, −0.5), Exposure × Subject question, and Exposure × Nonsubject question. Using the same procedure as Experiment 1, we began with a maximal random effect structure and simplified. The final model included random intercepts for items and participants.⁵

3.2. Result

The critical question in Experiment 2a was whether we would observe the expected priming effect with audio stimuli, and we did. Participants reported a higher number of subject interpretations in the subject‐exposure (M = 87.34%, SE = 0.013) than in the nonsubject‐exposure condition (M = 75.16%, SE = 0.017; see Fig. 4). This pattern was supported by our mixed effects logistic regression, which revealed significant effects of exposure and question type and no interaction between them. Subject interpretations were most common for the “who” questions (M = 85.94%, SE = 0.014), and subject questions (M = 83.44%, SE = 0.021), and least for the nonsubject questions (M = 69.69%, SE = 0.026). In the final model (see Table 6), there was a main effect of exposure type and a main effect of nonsubject (“no”) questions. The “yes” questions did not differ significantly from the “who” question reference group. The interaction between exposure type and question type was not statistically significant; this was the case for both the subject and nonsubject question type variables.

Fig. 4.

Subject responses in Norming study (Experiment 2a). Note. subj (solid bar) = Subject exposure condition; NON (lattice bar) = Nonsubject exposure condition.

Table 6.

Mixed effects logistic regression model results for Experiment 2a

Fixed effects	Estimate	Standard error	z‐statistic	p‐value
(Intercept)	1.69	0.278	6.09	.000 ***
List (Subject vs. Nonsubject exposure)	1.12	0.364	3.08	.002 **
Subject question	−0.208	0.417	−0.498	.618
Nonsubject question	−1.32	0.404	−3.26	.001 **
List x Subject question	0.39	0.431	0.903	.367
List x Nonsubject question	0.04	0.379	0.107	.915

Signif. codes: 0 ≤ “***” < .001 < “**” < .01 < “*” < .05.

3.3. Discussion

Experiment 2a confirmed that participants adapt to exposure even in an auditory modality. In other words, participants changed their interpretations in response to a shift in the local frequency of reference patterns. This result replicated the findings of all of the previous studies examining referential adaptation (Arnold, 2026; Johnson & Arnold, ²⁰²³; Ye & Arnold, ^2023a; Roy et al., ²⁰²⁵) and extended these findings to a new modality.

4. Experiment 2b

Using the auditory paradigm from Experiment 2a, we sought to test the effect of changing exposure between blocks of stimuli. This allowed us to incorporate cues to help participants to understand the exposure blocks as distinct discourse contexts by using different narrators to voice stimuli in each block. In both lists, stimuli for block 1 were voiced by a female speaker, and the stories and pictures involved one set of participants (David, Elaine, Nancy, Thomas). By contrast, block 2 was voiced by a male speaker and included stories featuring a second set of characters (Ana, Liz, Matt, Will). Varying the speaker across blocks imitated the real‐life situation of establishing a new conversational context with a new conversation partner. Similarly, changing the set of characters under discussion between blocks created a sense of shift in topic.

Research suggests that people do attend to the voice in spoken input, and they use it for various processing decisions. For example, research investigating human sensitivity to individual speaker characteristics suggests that speaker identity is used in predictive processing. Weiss et al. (2009) found that in a statistical learning task, people could learn different structures in speech streams when the gender of the talker was different, but not when it was the same. In addition, a larger N400 effect is observed when a speaker's identity is judged to be inconsistent with the content of their utterance (Van Berkum, Van Den Brink, Tesink, Kos, & Hagoort, 2008). Van Berkum et al. (2008) found that the phrase (“I like to drink wine.”) elicited a more negative N400 when the phrase was produced in a child's voice compared to the same utterance in an adult's voice. Comprehenders can also learn to associate the specific acoustic characteristics of an individual speaker's voice with their tendency to use high attachment versus low attachment structures (Kamide, 2012). The strength of structural priming effects can also vary according to individual speakers. Chia and Kashak (2022) measured the structural priming effects elicited by 43 experimenters over the phone and found substantial differences in the priming effects elicited by individual experimenters. These findings cumulatively suggest that adaptation and implicit linguistic processes can be tied to the traits of individual speakers.

For this reason, we expected that pairing different voices with each block would help participants to identify the shift in frequency patterns, despite the fact that Dempsey, Liu, and Christianson (2020) found that non‐voice cues had no effect on syntactic adaptation. In their study, they added probabilistic co‐occurrence cues to their stimuli, for example, nationality adjectives or red font on difficult trials. These cues had no effect, but these manipulations do not really take advantage of humans’ sophisticated ability to identify individual speaker characteristics and use these characteristics actively during language processing. While we did not directly test the effect of using different voices in a manipulation, we included these signals with the goal of eliciting the greatest possible response to the shift in exposure across blocks. In addition, we strengthened the shift to a new frequency pattern by lengthening the training set of exposure items in block 2 to include 10 stories.

4.1. Methods

4.1.1. Participants

We estimated the required sample size with two methods. First, we calculated the sample needed to observe a main effect of exposure in the individual blocks, using the data from Experiment 2a. A post‐hoc power simulation using simr (Green & MacLeod, 2016) suggested that we needed 80 subjects to reach 80% power. Second, we ran a post‐hoc power simulation on the data in Experiment 1 to estimate the sample size needed to observe the interaction between list condition and block. The interaction required 80 subjects for 80% power. Given that the effect size for the interaction in an auditory sample was unknown, we preregistered a larger sample size of 100 participants.

One hundred adult participants were recruited online via Prolific, limited to those with IP addresses in the United States, the United Kingdom, or Australia. A further seven participants were excluded from the final data set based on the inclusion criteria described for Experiment 2a: lower than 75% accuracy on the unambiguous questions (n = 4); 2) learning English after the age of 6 (n = 3). Subjects were compensated with $5.00 for roughly 15–20 min of participation.

4.1.2. Materials and design

4.1.2.1. Design

For Experiment 2b, we switched to PCIbex instead of Qualtrics for experiment display because the timing demands of the audio paradigm were time‐consuming to program in Qualtrics. The task included 48 stories, each with one question, plus two practice items. The stories were identical to those in Experiment 2a, with the exception of two exposure stories, which were added to increase the number of exposure items in the design. Four of the literacy items were modified in order to convey the exposure manipulation in addition to testing participants’ literacy. We also used only Who questions for the critical stories, given that all question types showed the priming effect in Experiment 2a.

In Experiment 2b, each list was divided into two blocks in the same manner as Experiment 1. The subject‐first list used subject exposure in block 1, followed by nonsubject exposure in block 2. The nonsubject‐first list reversed this order of exposures. The manipulation in each of the two blocks was associated with two covariate cues: each block was voiced by a different speaker, and each block discussed a separate set of characters (see Fig. 5). A number of different story types were included (see Table 7).

Fig. 5.

Characters used in each block in Experiment 2b. The drawings for block 1 were created by Chan, Hwang, and Arnold (2018). The drawings for block 2 were created by Klibanow & Arnold (2020) and Kakoki & Arnold (2020), and are used with permission from J. Arnold (https://arnoldlab.web.unc.edu/publications/supporting‐materials/art‐and‐psychology/.).

Table 7.

Experiment design for Experiment 2b with item types and block information

Design overview
Story type	Question type
Block 1 ( 22 stories ). Female voice
4 initial exposure items	2 What/1 Bot/1 Literacy
8 critical items	8 Who questions
10 exposure items	3 Who/3 What/3 Bot/ 1 Literacy
Block 2a (24 stories). Male voice
10 initial exposure items	3 Who/3 What/ 3 Bot/ 1 Literacy
8 critical items	8 Who questions
6 exposure items	2 Who/2 What/1 Bot/1 Literacy
Block 2b (2 stories). Male voice
2 Literacy items	2 Literacy

The design differed from Experiment 1 in three ways. We measured participant responses by presenting participants with images or text boxes to click on to indicate their answer. We also included three answer choices with each question. Finally, the block structure was slightly different in Experiment 2b (see Table 7). For both lists, the second block used a longer run of training items at the beginning of block 2. As in previous studies, block 1 used four training items at the beginning of the study, followed by an interspersed group of eight critical items and 10 exposure items. In block 2, the procedure started with 10 training items followed by an interspersed group of eight critical items and six exposure items.

Forty‐eight stories were included in Experiment 2b, with one question about each story. Four new stories were added to increase the number of available exposure items (see Table 7). We utilized the same question types as Experiment 2a, including reference, content, literacy, and bot questions (see Table 8). The 16 critical items and eight of the exposure items were the same as Experiment 2a and were paired with reference questions. All of the reference questions used a “who” phrasing (“Who waited in line?”). An additional 10 exposure items from 2a were included, and these items were paired with content questions that used a “what” phrasing (“What did Ana spill on the table?”). Participants responded to questions by clicking on pictures. The pictures always included the character's name to help participants remember which picture went with which name. Half of the “what” questions had participants respond by clicking on an image of the appropriate object from the story context. The other half had participants click on a text box. The bot questions were the same as Experiment 2a, except that the instruction to ignore the story was included in two different ways. Half of the questions had an instruction in the question audio to demonstrate that the participant was attentive and human (“Nancy washed the dishes with Thomas and then she dried the utensils. Who dried the utensils? If you are a human, ignore the question and click on the picture of Thomas”). The other half of the bot items included a similar prompt within the answer choices. This was a text box containing the prompt (“If you are a human, ignore the question and click here”). Literacy items were the same as 2a, with the exception that four of the items had the phrasing of their second sentence modified to include the priming manipulation. Two standalone literacy stories appeared at the end so they would not influence responses on the critical items; the literacy responses were only used for participant inclusion.

Table 8.

Example items with image answer choices from Experiment 2b

4.1.2.2. Audio and images

The audio for Experiment 2b was recorded by one female (block 1) and one male (block 2) adult native speaker of American English; both were trained research assistants. The recordings for each item were made using a Marantz solid‐state recorder, a Yamaha mixing console, and Praat software. The same process to record and edit the audio stimuli was used as in Experiment 2a. One exception was that the selection of the base recording for the exposure items was rotated such that the number of intact and spliced recordings was almost (but not fully) balanced between the two lists in order to prioritize the quality of the recordings. Due to PCIbex constraints, we used .mp3 instead of .WAV audio files. Images used in the experiment were .png files obtained from license‐free sources on the internet via web searches. The text boxes were created in Microsoft Word and saved as .png files.

4.1.3. Procedure

The procedure for Experiment 2b was the same as for 2a, except that it was run on PCIbex (not Qualtrics), and the procedure in 2b included a section before block 2 where the new characters were introduced with their names and pronouns. The survey for 2b typically took participants 10–20 min to complete, which was briefer than 2a.

4.1.4. Trial structure

The trial structure was the same as 2a, with two exceptions. Instead of a warning tone, a GIF of a cassette tape appeared when audio was presented and disappeared when the audio recording finished. Participants also responded by clicking on one of the three pictures/text boxes that constitute the answer choices, rather than indicating their selection via a text‐based multiple‐choice question.

4.1.5. Analysis

We used the same analysis procedure as for Experiment 1.⁶

4.2. Results

Our results in Experiment 2b echoed the pattern observed in Experiment 1 (see Fig. 6). Participants maintained a high rate of subject interpretations in the subject‐first list where they encountered subject antecedent exposure in block 1 (M = 83.9%, SE = 0.018), followed by nonsubject antecedent exposure in block 2 (M = 84.9%, SE = 0.018). In the nonsubject‐first list, participants were first exposed to nonsubject reference patterns, and they responded to this with a relatively low rate of subject interpretations (block 1: M = 68.2%, SE = 0.023). In block 2, when they received subject exposure, participants shifted and produced more subject interpretations (block 2: M = 81.8%, SE = 0.019). The rate of subject responses was higher in Experiment 2b, compared to Experiment 1. This is likely to be the result of a “yes‐bias” in Experiment 1, stemming from the fact that selecting the subject in critical items required participants to select a “no” response.

Fig. 6.

Subject responses in Experiment 2b. Note. subj (solid bars) = Subject exposure condition; NON (lattice bars) = Nonsubject exposure condition. The red column (left side) represents the first block, and the blue column (right side) represents the second block.

We followed the same process of model simplification as in previous experiments. The final model (see Table 9) included random slopes for block by participants and random intercepts for participants and items.⁷ Critically, we found a significant interaction between the block and the list condition, which replicated the interaction that was identified in Experiment 1. The main effect of block was not statistically significant, but the main effect of list was significant, reflecting a higher rate of subject interpretations in the subject‐first list than the nonsubject‐first list.

Table 9.

Mixed effects logistic regression model results for Experiment 2b

Fixed effects	Estimate	Standard error	z‐statistic	p‐value
(Intercept)	1.948	0.257	7.57	.000 ***
Block	−0.582	0.464	−1.26	.21
List	0.613	0.282	2.17	.03 *
Block × List	0.939	0.348	2.698	.007 **

Signif. codes: 0 ≤ “***” < .001 < “**” < .01 < “*” < .05.

Again, we probed the interaction with pairwise contrasts. Examining both lists while holding the block constant shows a significant effect in the first block (Table 10). This showed that the rate of subject responses was significantly different between the exposure conditions in each list for the first block. There was a second significant difference in subject responses between block 1 and 2 for the nonsubject‐first list (Table 11) but not the subject‐first list. Together, these comparisons suggested that participants changed their processing biases in block 2 only when they received the less common nonsubject remention pattern first, followed by the more common subject remention referential pattern.

Table 10.

Pairwise contrast—Estimates for the effect of list by block

List pairwise	Block	Est.	SE	z‐statistic	p value
List 2 (Nonsubj first) ‐List 1 (Subj first)	1	−1.082	0.331	−3.27	.001 ***
List 2 (Nonsubj first) ‐List 1 (Subj first)	2	−0.144	0.331	−0.433	.665

Table 11.

Pairwise contrast—Estimates for the effect of block by list

Block pairwise	List	Est.	SE	z‐statistic	p value
Block 2 – Block 1	List 1 (Subj first)	0.113	0.499	0.226	.821
Block 2 – Block 1	List 2 (Nonsubj first)	1.05	0.492	2.14	.0326 *

4.3. Discussion

In Experiment 2b, we replicated the essential findings of Experiment 1 but in the auditory modality. In the first block, we observed the expected difference between subject‐exposure and nonsubject‐exposure. But in block 2, when the exposure conditions reversed, only participants who got nonsubject‐exposure first showed any sensitivity to this change. This implied that local reference statistics can be tracked dynamically, as they change within a situation. However, this finding is limited to the case where the first block showed a less probable nonsubject‐exposure pattern. By contrast, when participants encountered the more probable subject‐exposure pattern first, this appeared to anchor their expectations for the remainder of the procedure, so they did not respond to the block 2 pattern.

We designed this study around the idea that different voices and characters across the two lists may signal a new discourse context in the second block and strengthen sensitivity to the changing reference patterns. However, given that a very similar pattern was observed to that of Experiment 1, it is unlikely that these cues substantially changed behavior. We do not have data about how the shift in topic and narrator was perceived by participants because we did not include a debrief at the end of the experiment. However, the new characters and their personal pronouns were introduced by the new narrator before the beginning of the second block, so it is likely that subjects were explicitly aware of both the change in narrator and the change in the named characters populating the stories.

5. General discussion

The primary finding from this study was that people can adapt flexibly to changes in referential patterns, but only when the first exposure block was a low‐frequency pattern. In a blocked design, we replicated the known referential adaptation effect in the first block. When the subject was always rementioned with a pronoun (Subject‐first List), people assigned pronouns more to the subject, and when the nonsubject was always rementioned, they assigned pronouns relatively more to the nonsubject. But our real question was what happened in the second block, when the pattern shifted. Here, we found a difference between the lists. If people started with the subject‐exposure block, the pattern more frequently found in natural language, they showed no sensitivity to nonsubject‐exposure in the second block. But if people started with the less‐common nonsubject exposure block, they quickly bounced back to the “typical” pattern of responses in the second block, where they encountered the more common subject‐exposure.

This study is the first to show that language users can adapt to multiple distinct referential patterns in a single experimental session, at least under some circumstances. Our results show that even after a participant has adapted to a first pattern, they are capable of forming generalizations about a second pattern and adapting to this second pattern as well. Our design mimics the real‐life situation of consecutive conversational partners. Correspondingly, the capacity to flexibly adapt to new reference patterns would have certain advantages, like being able to track referential frequencies by situation or by speaker. Multiple short‐term adaptations could be used to make processing easier and more efficient in a complex language environment. Longer‐term adaptations could contribute to a language user's long‐term learning about language and how it is used by specific speakers or in specific situations.

We considered four types of patterns for how people might adapt when resolving conflicting inputs: (1) a recency‐based pattern (where people shift to follow the most recent input, which here would be observable in the second block; (2) a primacy‐based pattern (where people follow the pattern encountered at the start of a situation, which here would be the first block—either as a simple effect or contingent on whether it matches lifetime biases; (3) a frequency‐based pattern (where lifetime and recent inputs are averaged); and (4) a stronger effect of exposure in the second block. Our results were inconsistent with the effects of recency, simple primacy, frequency‐averaging, and a stronger effect for block 2.

Instead, our results displayed a contingent primacy effect, suggesting that adaptation is sensitive to a combination of lifetime, situation, and recency calculations. Participants have a bias toward using more frequent structures, such as assuming the subject will be rementioned. But they also exhibit a situation‐specific primacy effect. If people first encounter subject‐reference pronouns in the first block, they set an expectation that the current situation is similar to other situations. The dominant pattern is so strong that people ignore the shift to nonsubject‐reference in the second block. But if people first encounter frequent nonsubject‐reference pronouns (Nonsubject‐first List), this atypical pattern triggers a local shift toward expecting nonsubject reference. However, this atypical pattern is unstable because it is inconsistent with prior expectations. Thus, as soon as participants get evidence that the dominant pattern has returned, which they do in the second block, they immediately pop back to their baseline pattern of pronoun interpretation. This pattern of results indicates that subjects privilege the material they encounter at the beginning of a discourse situation. But they only maintain the referential pattern they encounter first if it is consistent with their lifetime experience with reference.

The current set of findings makes a key contribution to our understanding of referential adaptation by identifying a new constraint, namely, the importance of calculating situation‐specific frequencies at the start of the experiment. We also found that this primacy effect is modulated by prior biases. We interpret our findings as evidence that people estimate whether the current input is typical or atypical, relative to their long‐term experience with language use. When a new situation presents a typical pattern (subject‐exposure), people stick with their prior biases for the entire situation. But when a new situation presents an atypical pattern (nonsubject‐exposure), their commitment to the new pattern is relatively weak, and can be overturned by new input. This interpretation is consistent with evidence that native speakers can usually tell very quickly whether they are conversing with another native speaker or a second language learner, a speaker that represents an atypical pattern of language use, and further, speakers are known to use this information (Kleinschmidt et al., 2012).

Situation‐specific adaptation effects represent a novel finding in the context of referential adaptation. But our findings are consistent with prior studies that find evidence that people learn about the language that is used in particular environments (e.g, specific registers; Ikeda, Kobayashi, & Itakura, ²⁰¹⁸) or by specific people (dialects or idiolects; Fraundorf & Jaeger, ²⁰¹⁶; Kroczek & Gunter, ²⁰¹⁷), and adaptation could be a part of how adults and children accomplish these types of learning. This is similar to the hypothesis that long‐term adaptations are fundamental to the formation of discourse processing biases (Johnson & Arnold, 2023; Ye & Arnold, ^{2023a, 2023b}). Rapid adaptation might be especially relevant to registers that are tied to physical locations or social situations, which can have a strong influence on how language is used. Consider the “frozen” register of a language that can be encountered in a religious institution, which is often very different from even the most formal variety of a language (Nordquist, 2025; Sun & Wang, ²⁰²³). Language also varies in formality across locations (e.g., bar vs. an office). Our findings raise questions about how people learn that certain language is appropriate in certain social situations. One way to approach these questions in future research is to see how subjects adapt if their physical context is changed between blocks or if the linguistic formality level changes between blocks.

However, it is not currently clear whether referential adaptation is triggered by a change in input frequency alone or whether other cues like speaker identity might also signal a shift. Prior research has shown that syntactic adaptation occurs for patterns that are associated with specific speakers (Kamide, 2012). For referential adaptation, we observed the same pattern in both experiments, even though Experiment 1 did not involve a shift in speaker or story characters, while Experiment 2 did. We hypothesized that hearing a new speaker might signal the shift and strengthen the effect in Experiment 2, but the effect observed was similar to that in Experiment 1. This suggests that people track the frequency of referential patterns for a situation even in the absence of signals about who is speaking.

Our results revealed that in the subject‐first lists, adaptation persisted from block 1 to block 2. This is inconsistent with a simple, additive view of priming effects. It is also inconsistent with an unweighted averaging of the referential statistics in the input. One possibility is that material presented in the first stage of the procedure is particularly salient, giving the initial pattern more prominence in memory and a greater influence on what subjects ultimately learn. A similar primacy effect was observed in an artificial language learning study. Gebhart et al. (2009) exposed subjects to two sets of artificial trisyllabic patterns and found that learning only occurred for the initial pattern, unless the transition between patterns was marked by an explicit cue or an exceptionally long period of exposure to stimuli. We did not observe a true primacy effect for both lists; our subjects only adapted to the less reliable nonsubject exposure pattern when it was the first pattern they encountered.

Our findings may also have some bearing on the question of how long discourse adaptation lasts. Contemori et al. (2022) found evidence of adaptation 6–10 days after their initial exposure paradigm, and the effect had not diminished in comparison to the post‐test administered immediately following the procedure. This appears to contrast with our findings. If we presented subjects with only our nonsubject exposure stimuli and used Contemori et al.’s multiday experiment structure, what would we expect? Given that subject reference is common in English in general (Arnold, 1998), we might reasonably imagine that subjects would encounter substantial subject exposure “in the wild.” If real‐life exposure functions like the second block of subject‐reference, we might expect people to return to typical levels of assigning pronouns to the subject, effectively eliminating the exposure effect. But this is not what Contemori et al. found. On the other hand, participants might consider the real‐life input as a different discourse context than the experiment. If so, we might expect the nonsubject‐exposure input to persist across several days. If so, this would suggest that in longitudinal adaptation studies like Contemori et al. (2022), subjects may think about the study as learning about one context and then returning to this same situation 6–10 days later.

Our findings point to a view of referential adaptation as a mechanism for optimizing comprehension in the current situation. This parallels other evidence that context is part of what subjects learn in a priming experiment. For example, Kroczek and Gunter (2017) exposed participants to SOV and OSV word orders that were associated with particular speakers. In a longitudinal design, these participants returned after 9 months. Subjects’ expectation for a particular syntactic structure was quickly reinstated after exposure to the speakers, and the priming effect was actually stronger in this delayed condition.

Most current theories of adaptation are not flexible enough to model this effect of context effectively. In models like the error‐based implicit learning model (Chang et al., 2006) and dual mechanism models that incorporate base‐level learning and spreading activation (Reitter et al., 2011), speaker and environment‐specific information is not maintained in the long term (Jaeger & Snider, 2013). This implies that these models would require additional mechanisms to explain our results or the results of prior studies that find evidence of context‐specific adaptation.

One potential avenue for modeling situation‐specific effects may be with Bayesian models, because they assume that language users will maintain situation‐specific information if the information is reliable and improves predictions for upcoming language (Jaeger & Snider, 2013; Kuperberg & Jaeger, ²⁰¹⁶, Kleinschmidt et al., 2012; Kleinschmidt & Jaeger, ²⁰¹⁵). This could be extended to account for a situation‐specific primacy effect if the link between context and language structure is both reliable and useful. In the present study, this would predict a stronger effect of initial input when it is the reliable pattern, that is, the one that matches pre‐experiment biases (subject‐exposure). However, this explanation would require a significant change to current Bayesian models, which predict that prior experience should consistently favor subject remention patterns, leading to surprise from the nonsubject exposure condition, whether it occurs in block 1 or 2. In order to account for our findings, we would need to add a mechanism for assessing situation‐based expectations based on input at the start of a new situation.

Our results also have implications for other current theories of implicit learning and adaptation. Theories of implicit language learning might be improved by fully taking into account the fact that memory encoding is inherently context sensitive. Memory researchers have been aware for decades that the informational content of memory is bound to the internal and external context in which a memory was formed (Eich, 1980; Tulving & Thomson, ¹⁹⁷³). The encoding specificity principle indicates that retrieval is easier when the context of encoding or original exposure matches the context of retrieval (Nielsen & Christiansen, 2025). This may help to explain context‐specific priming effects (Contemori, 2021; Kamide, ²⁰¹²; Kroczek & Gunter, ²⁰¹⁷). Subjects may find it easier to retrieve a particular syntactic or discourse structure if they are in a situation where they use the structure regularly or are with a familiar person who uses the structure.

Our results are also theoretically relevant because they provide a serious challenge to many error‐based models of adaptation. Our experimental design effectively provides a contrast between different levels of predicted bias. In block 1, everyone begins with a general prediction that referring expressions will remention the subject. In the subject‐first list, this prediction is upheld, so there is little to no predicted error signal. In the nonsubject‐first list, this prediction is violated, resulting in an adjustment to predictions such that the nonsubject referent is relatively more expected. Under this scenario, at the beginning of block 2, we would expect a very strong error signal for the subject‐first list, where suddenly people get the unexpected nonsubject exposure. This should lead to a substantial shift toward assigning ambiguous pronouns to the nonsubject. However, this is not what we found. Instead, there was no effect of the nonsubject exposure in block 2. By contrast, there was a strong shift in response to subject exposure in the second block of the nonsubject‐first list, even though the error signal generated by this block is predicted to be relatively weak. This provides further evidence that error‐based learning is unlikely to account for referential adaptation.

The fact that our exposure effect differs across blocks also challenges a hybrid model like the one proposed by Reitter and colleagues (2011). The base‐learning function in this model should be associated with an inverse frequency effect, where more common constructions are associated with weaker priming effects and uncommon structures are associated with relatively stronger priming effects. In addition, the spreading activation mechanism bolsters the current input, but this is a short‐term mechanism that favors recent input. In the subject‐first list, there is little behavioral effect of processing the primed structures from base learning (because the first block exhibits the frequent exposure pattern), but there is some activation of the subject‐remention pattern through spreading activation. But spreading activation is a short‐term mechanism that favors recent input, so it should start to support activation of nonsubject remention during the second block. In addition, the nonsubject pattern is less common, which suggests that the pattern's repeated retrieval from memory should further support a shift toward nonsubject interpretations. But this is not what we found.

In our study, the subjects were required to resolve a conflict between their preferences formed by long‐term experience and those derived from the recent linguistic input. As we have shown, this is a complex process. People do not simply respond according to their long‐term experience or discount valuable prior experience and shift with each new pattern. Our results point to a preference for the referential pattern that is more common in subjects’ lifetime experience. They both tend to continue using subject antecedent interpretations after the input has changed to a new pattern and tend to switch to subject antecedent interpretations when the input presents them with an opportunity to do so.

This can also be conceptualized as an effect related to reliability. Subject remention patterns signal that the environment is typical and reliably indicate that further subject rementions are likely. Nonsubject remention patterns, on the other hand, do not reliably signal more nonsubject remention patterns or more subject remention patterns based on lifetime experience with reference. Instead, frequent nonsubject remention patterns are likely to signal that the current situation is atypical, and eventually, participants will adapt to the high local frequency of nonsubject patterns.

Future research could test the prediction that the reliability of linguistic patterns influences adaptation by repeating the blocked priming procedure using a referential ambiguity where subjects’ lifetime experience is relatively weakly biased toward the higher frequency referential pattern. This should operate like having a weakly informative prior, which would change how subjects respond to new evidence. Subjects should be more sensitive to the evidence provided by the exposure conditions within the experiment, and thus be more willing to adapt their expectations in favor of the less frequent pattern in circumstances paralleling block 2 of our subject‐first condition (frequent pattern then less frequent pattern). Correspondingly, they should require more evidence before shifting in favor of the higher frequency pattern in circumstances resembling block 2 of our nonsubject‐first condition (less frequent pattern followed by more frequent pattern).

Our findings join other recent evidence against an error‐based mechanism for language structure learning. First, an analysis of structural priming found that the reading times for primes and targets were positively correlated (Tooley & Brehm, 2025). Second, a meta‐analysis by Fazekas et al. (2024) showed that the inverse frequency effect, a key prediction of error‐based models, was found only 42% of the time. Third, Ye and Arnold (2026, under review) found that referential adaptation occurred even in equi‐biased contexts without a strong prediction error.

Another major contribution of this study is that we established that the referential adaptation effect remains robust in auditory versions of the task. This replication is advantageous because the audio format may be a more natural task than reading. This avoids potential differences in reading experience across participants and offers the opportunity to test populations with reading difficulties, including children or individuals with cognitive disabilities. The auditory task also provides a method of testing the role of auditory cues on referential adaptation in future studies.

Conclusion

The present set of experiments suggests that language users are sensitive to the presence of multiple referential patterns in their language input. This implies that the statistical learning of referential patterns is not reckoned cumulatively, based on all of the language input found in the environment, or averaged across multiple speakers. Instead, learning from the local language environment appears to occur on a situation‐by‐situation basis. In addition, local learning is modulated by biases learned through lifetime exposure, such that adaptation to atypical patterns is weaker when it occurs after exposure to the typical pattern than when it is the first pattern in a new situation.

Conflicts of interest

We have no known conflicts of interest to disclose.

Supporting information

Supplementary Information