Since the early stages of research within the field of bilingualism, an important goal has been to shed light on the representation and processing of two languages within the bilingual mind. Debates continue to center around whether the two languages of a bilingual have a shared or a separate representation (Kroll and Tokowicz, 2009). Furthermore, there is ongoing discussion about whether each language is processed independently of the other when bilinguals are in a primarily “monolingual mode” as when they are actively communicating using only one of their two languages, with the other language essentially deactivated (Grosjean, 2010). Bilingual language processing is modulated by various factors such as the following: typological similarities and differences between the two languages, age at onset of exposure, contexts of acquisition (simultaneous versus sequential; immersion versus formal instructed learning), frequency of use, level of proficiency, and language dominance (Grosjean, 2010). A number of studies using different methodologies suggest that there are both shared and separate representations. Findings from the literature support a shared system at the semantic or conceptual level corresponding to word meaning (Francis, 2009) but separate representations at the lexical level corresponding to word form (Kroll and Stewart, 1994; Kroll et al., 2010; Costa, 2009). As for syntax, some studies have found evidence for a shared syntactic representation, particularly in relation to constructions which are functionally and/or structurally similar across the two languages (e.g., Hartsuiker et al., 2004; Loebell and Bock, 2003; Salamoura and Williams, 2006, 2007; Shin and Christianson, 2009). Other studies have provided evidence favoring separate representations, particularly for constructions that are structurally dissimilar as well as in contexts where the language mode is primarily a monolingual one (e.g., Ahn and Ferreira, 2024; Bernolet et al., 2007; Kim and McDonough, 2008).

A related issue concerns the role of crosslinguistic influence. Traditionally, influence of one language upon the other has been framed in terms of transfer (i.e., language transfer). Note that Uriel Weinreich (1953) in his seminal book on bilingualism viewed language transfer as potentially bidirectional, with L1 and L2 influencing one another. However, until very recently, most second language acquisition research focused on L1 transfer, i.e., unidirectional transfer where the L1 influences the developing L2 (Gass and Selinker, 2008). Recent research on simultaneous and/or sequential bilingual language development has shown that interaction among the languages is not solely unidirectional from L1 to L2 but bidirectional (Cook, 2003; Pavlenko, 2004; Pavlenko and Jarvis, 2002; Schmid and Köpke, 2017). In the present study, we prefer the term “crosslinguistic influence,” as it is a broader term that encompasses any form of interaction between the languages in a bilingual’s (or multilingual’s) repertoire. Crosslinguistic influence can manifest as either facilitation or interference. On the one hand, when facilitative it can support and enhance the development of the target language. On the other hand, it can result in interference, as when activation of properties from one language hinders the processing and acquisition of another. Furthermore, the term “language transfer,” which also stems from the behaviorist research paradigm, has recently been criticized (see Sharwood Smith, 2021) as a pernicious notion, misrepresenting how linguistic properties function in the mind. He suggests that, instead of transfer, languages are interconnected within a shared cognitive system, where structures coactivate without transferring properties. Furthermore, we concur with Berkes and Flynn (2016) that language learners do not build the new grammar by merely transferring the language-specific features of the L1 into the new language. Instead, they construct the new grammar via grammatical mapping as in the case of L1 acquisition (Lust, 2012). In other words, what learners do is “map from one primary structure to a more developed structure by dissociating modular grammatical components and integrating them in the ‘assembly’ of new language-specific grammars” (Flynn et al., 2005:2, cited in Berkes and Flynn, 2016; see also Flynn et al., 2004; Fernández-Berkes and Flynn, 2023; Lakshmanan, 2024; Lust et al., 2025). The primary structure is a minimalist Universal Grammar (UG), a fundamental and an invariant template guiding the mapping to a language-specific grammar. When acquiring a second or third language, the previously acquired language/s can cumulatively enhance the process of grammatical mapping to the language-specific grammar. Merge, a core UG operation at the heart of grammar building, provides the template for combining syntactic elements (e.g., words, phrases, clauses, sentences) into recursive hierarchical structures (e.g., coordination, adjunction, and embedding, etc.), enabling the unbounded expressive power of language.

With growing interest in language maintenance and attrition (in particular, L1 attrition), increasing attention has been given to how subsequent languages influence the L1. Existing research indicates that all linguistic subsystems, phonology, morphology, lexis, syntax, semantics and conceptual representation, as well as pragmatics, are susceptible to varying degrees of crosslinguistic influence, with certain aspects (e.g., lexis) being more vulnerable than other aspects, such as syntax (Gallo et al., 2021; Gürel, 2008; Schmid and Köpke, 2017; Lakshmanan, 2024; Teruya and Lakshmanan, 2019).

Historically, second language acquisition research has compared L2 learners to monolingual speakers of the target language in order to determine L2 learners’ success in converging upon L1 (native)-speaker norms for the target L2. This approach tends to emphasize the failure of L2 learners to achieve the same level of competence as L1 speakers.¹ As Grosjean (2010) has cautioned, a bilingual is not two monolingual speakers in the same mind/person. Furthermore, Cook’s (1991, 2002, 2012) multicompetence framework, which assumes that language transfer, i.e., crosslinguistic influence, is potentially bidirectional, holds that it would be erroneous to expect bilingual speakers to function exactly like monolingual speakers in each of their two languages. Furthermore, Falk and Bardel (2010) emphasize that an L2 can influence an L3 to a degree comparable to that of an L1, both at the lexical and syntactical levels. However, they also note that it is often unclear which specific background language—whether L1 or L2—plays a more decisive role in the crosslinguistic influence on an L3. This is further supported by scholars who have argued in favor of the cumulative enhancement model, according to which all languages known can potentially influence the development of subsequent learning (e.g., Flynn et al., 2004). A substantial body of research on bilingual code-switching has shown that bilinguals are able to code-switch between their two languages with ease when they are actively using both languages simultaneously, i.e., when they are in a bilingual mode. At the same time, even when only one language is in use, as when they are in a monolingual mode, the other language that is not in use is also “activated.” Such crosslanguage or parallel activation during a monolingual mode has been observed for different aspects of language from lexis to syntax. When the activation involves elements or aspects that are not shared across the two linguistic systems, inhibition would need to occur in order to ensure convergence upon the target language in use (Bialystok, 2001; Green, 1998; Grosjean, 2010; De Groot and Kroll, 1997).

Generative approaches to bilingualism and second language acquisition seek to determine in what way the innate biological endowment for language, that is the invariant and parameterized principles of Universal Grammar (UG), remains available across the two languages of a bilingual. It is expected that invariant principles of UG such as structure dependency will be accessible for acquisition and use of a subsequent language. This means that an invariant principle of UG would be instantiated in the same way across L1 and L2. However, even though we expect that both languages will adhere to an invariant UG constraint, it is possible that its operation could be impacted by other cognitive or linguistic properties or processes. Structure dependency is a fundamental property of human language (see for instance, Chomsky, 1971, 1980), according to which grammatical operations are based upon the hierarchical phrasal structure rather than upon a linear sequence of first, second, third, etc. It has been shown that acquisition relies on abstract structural principles rather than on a linear ordering of words (Crain and Nakayama, 1987; Berwick et al., 2011). However, in real time processing, parsing strategies (like those proposed by Fodor and Frazier, 1978 and De Vincenzi, 1991) can interact with structure dependency. Assuming that a UG principle is parameterized, as for example in relation to Head directionality (i.e., head-first versus head-last) or branching directionality (i.e., right-branching versus left-branching phrase structure),² the expectation is that acquisition and use of a new language will be easier in instances where the parametric setting or value is the same for both languages.³ In those instances where the parameter setting is different for each language, particularly in relation to sequential bilinguals, the L1 (or the dominant language), could impact the use of the L2. “Resetting” of the UG parameter (or more appropriately, activation of the L2 parametric value) would need to occur based on positive evidence, as well as in cases where the input is not sufficiently rich or precise (poverty of stimulus). The existing evidence in second language acquisition has been largely consistent with the strong continuity hypothesis, which holds that UG continues to remain fully available for subsequent language acquisition (Epstein et al., 1996, 1998; Flynn, 1987; Lakshmanan, 1994; Lardiere, 1998; White, 2003; Schwartz and Sprouse, 1996; among others). Continued access to UG, however, does not entail the rejection of L1 influence. Various proposals have been put forth in L2 research on the role of L1 transfer: No L1 transfer, Partial L1 transfer and Full L1 transfer (for review, see White, 2003; see also Pienemann et al., 2005 for discussion of psycholinguistic processing constraints on L1 transfer). As discussed previously, along with Berkes and Flynn (2016), we assume that second language learners do not construct the new grammar by mere transference of language-specific features from the L1 into the new language. Rather, as in the case of L1 acquisition, they build the new grammar via a process of grammatical mapping, which involves mapping from one primary structure (i.e., a minimalist UG-determined template) to a more developed structure. This involves a process of dissociation and integration of modular grammar components in constructing the new language-specific grammar/s, despite the wide variation both within and across languages (e.g., branching directionality/Head-directionality parameter, tense and agreement morphology, null arguments, associated semantics and pragmatics, etc.). As mentioned previously, in the case of second and third language acquisition, all previously acquired language specific grammars have the potential to cumulatively enhance the mapping from the initial UG template to the new language-specific grammar. In relation to real time processing of phrases and sentences, due to potential cross-language activation, in cases where the L1 and L2 differ, the processing of a new language for comprehension and production during a monolingual language mode could involve the inhibition of an L1 property or parametric value and activation of the appropriate L2 property or value.

In the current paper, we investigate the acquisition and use of recursive adjectives in an adult sequential bilingual context. Our focus here is on L1 Romanian-L2 English bilinguals who began learning English in a formal instructional setting in Romania as a foreign language during early childhood and were pursuing their undergraduate studies at a university in Romania. Importantly, our goal is to investigate how highly proficient English L2 users with Romanian as the L1 and their dominant language interpret complex recursive adjective phrases in both their languages. Our main focus is on their knowledge and use of these structures in both their L1 and L2. In this study, we test the universality of the Recursive Set Subset Ordering (RSSO) Constraint, a UG-based structure-dependent principle according to which set adjectives are merged to the head noun first, and subset adjectives are merged subsequently. Specifically, our study provides novel evidence from English, a language with left-branching adjective structures, and contrasts it with Romanian, a language with right-branching adjective structures. By comparing recursive modification across these languages, we aim to assess whether branching directionality, as well as other cognitive factors—such as Adjective Ordering Restrictions (AOR) in terms of conceptual properties (like size or color)—modulate adherence to the RSSO constraint. In so doing, the study contributes to our understanding of how universal grammatical constraints interface with language-specific variation and other cognitive factors in the L1 and L2 comprehension of bilinguals.

From a global perspective, and in terms of Principal Branching Directionality (Lust, 1983; Lust and Mangione, 1983), both English and Romanian are head initial, right branching languages. However, they differ in their branching directionality at the local phasal level (Lakshmanan et al., 2024). For example, adjectival phrases in English are left branching (prenominal) while in Romanian they are right branching (postnominal), as shown in (1) and (2) respectively. English is a natural gender language and Romanian is a language with grammatical gender, where adjectival modifiers agree both in number and gender with the head noun.

Both languages allow two different interpretations: (i) a coordinative interpretation, and (ii) a recursive set-subset interpretation. For instance, under a coordinative interpretation for (1a), the phrase big red flowers refers to flowers that are both big and red. Here, the two adjectives, which are not in a set-subset relationship, are combined together first before the resulting unit is merged with the head noun. In contrast, under a recursive set-subset interpretation, the same phrase refers to a subset of big flowers from a set of red flowers. The same is true of the corresponding Romanian adjectival phrase in (2a). As for the phrase red big flowers in (1b), under a coordinative interpretation, it refers to flowers that are both red and big. In contrast, under a recursive set-subset interpretation, the same phrase refers to a subset of red flowers from a set of big flowers. This is true of the corresponding Romanian adjectival phrase in (2b).⁴ Note that regardless of whether the adjectives occur prenominally (as in English) or postnominally (as in Romanian), the adjective that specifies the set must be closest to the head noun, and the adjective that specifies the subset is further away from the head noun. This universal constraint at the syntax-semantic interface is known as the Recursive Set Subset Ordering Constraint (Foucault, 2020; Foucault et al., 2021, 2022a; Bleotu and Roeper, 2021a, 2021b, 2022a, 2022b), as defined in (3), and it constitutes the focus of the current study:

In addition to the Recursive Set Subset Ordering Constraint (RSSO) on adjectives, there is also an Adjective Ordering Restriction (AOR) on the linear sequence of adjectives in language, associated with coordinative (non-set-subset) uses of adjectives, as in big red flowers understood as ‘big and red flowers’. Across many languages in the world, it has been argued that more objective properties, such as color, appear closer to the noun, whereas more subjective properties, such as size, appear further away from the noun (Dixon, 1982; Sproat and Shih, 1991; Scontras et al., 2017, 2019). This is illustrated in (4):

AORs have given rise to debates regarding their status (universal vs. language-specific property) and their nature (syntactic/semantic/cognitive). According to Cinque’s (1994, 2005, 2023) cartographic approach, AORs are syntactic universals subject to crosslinguistic variation. Apart from the fact that adjectives vary in how they are ordered with respect to the noun (prenominal vs. postnominal), adjectives also vary in how they are ordered with respect to one another. For instance, in a Germanic language like English, where adjectives are prenominal, the most natural order for a sequence of adjectives involving a Size adjective and a Color adjective is A_Size A_Color N. In contrast, in a Romance language like Romanian, where adjectives are postnominal, the most natural order for such a sequence would be N A_Color A_Size, which represents a mirror order of English. It is important to note that the AORs apply to coordinative contexts and not to recursive set-subset contexts.

For the purposes of the current paper, we assume that AORs involve a strict hierarchy across languages, in line with Cinque and others. However, it is relevant to mention that, for Romanian, some recent studies have suggested a less rigid AOR ordering (see e.g., Cornilescu and Cosma, 2019; Bleotu and Luciu, 2024; Trușcă and Bleotu, 2024, see also Leivada and Westergaard, 2019 for Greek). In coordinative contexts, native speakers of British English tend to observe AORs overall, e.g., the order A_Size A_Color N (e.g., big red flowers) is overwhelmingly preferred over A_Color. A_Size N (e.g., red big flowers). In contrast, in Romanian, besides the expected mirror order of English N A_Color A_Size (e.g., flowers red big), the order N A_Size A_Color (e.g., flowers big red) has also been reported. It is possible that the emerging flexibility recently observed in Romanian could stem from the influence of English, given that today in Romania, L1 Romanian speakers learn and use English regularly. Therefore, more studies are needed to conclusively shed light on the issue. In our study, we assume along with Cinque, that AORs are in principle observed in Romanian as well, with the understanding that there may be some flexibility, especially in the case of bilinguals (for discussion on the influence of English on Romanian, see, for instance, Avram, 1997; Stoichiţoiu-Ichim, 2001, 2002, 2003).

Moreover, and most importantly, recent research with adults and children shows that, in recursive set-subset contexts, the universal RSSO constraint overrides AORs in English, as well as in Romanian. One line of research, exemplified by Foucault et al. (2021, 2022a) and Bleotu and Roeper (2021a, 2021b), which build upon on Roeper and Snyder (2004) and Roeper (2011), looked at recursive adjectives such as small big flowers and big small big mushrooms, adjectives which involve gradable properties characterizing the same dimension (size). Such structures typically give rise to recursive set-subset interpretations because coordinative readings (e.g., small and big) involve contradictory properties and are therefore less plausible. In such recursive interpretations, the outer adjective modifies the property introduced by the inner adjective, reflecting a set-subset hierarchy. For instance, in a phrase like a large small shirt, large is interpreted relative to the category small shirt rather than as a simple conjunction of large and small. Another line of research involves adjectives which specify properties characterizing different dimensions (size, color): long green leaves or green long leaves. Such adjectives can give rise not only to recursive set-subset interpretations but also to conjunctive readings, without a set-subset hierarchical relationship. However, in contexts where a clear set-subset distinction is present—such as visually salient subsets—speakers tend to order adjectives to reflect this hierarchy. Bleotu and Roeper (2022a, 2022b) tested the validity of the Recursive Set-Subset Ordering Constraint (RSSO) as defined in (3) above, in order to see whether participants observe RSSO or AOR for such sequences in recursive set-subset contexts. In a forced choice task, participants had to choose the best answer to describe a picture. The authors found that, when describing a picture depicting a subset of long leaves within a set of green leaves (see Figure 1A), Romanian adults preferred to merge the Color adjective before the Size adjective (5a). Conversely, when referring to green leaves within a set of long leaves (see Figure 1B), they preferred merging the Size adjective first to the head noun (5b). The order frunze verzi lungi ‘leaves green long’ is not in line with the strict SizeA ColorA Noun AORs predicted for Romanian by Cinque’s approach. According to Cinque (1994, 2005), in coordinative contexts, the order for English should be SizeA ColorA N, while for Romanian, the mirror order Noun ColorA SizeA should hold. Interestingly, children and adults tend to choose N ColorA SizeA orders if the color adjective picks the set and the size adjective picks the subset, but they tend to choose N SizeA ColorA orders if the size adjective picks the set and the color adjective picks the subset. Their empirical findings confirmed that participants will order adjectives according to a set-subset hierarchy over AORs when the context requires it.

This study and others bring supporting evidence that monolingual Romanian and American English children and adults observe the by RSSO by choosing to associate the adjective merged closest to the head noun with a set interpretation and the adjective merged further away from the head noun with a subset interpretation, regardless of the adjective employed (size/color) (for Romanian data, see Bleotu and Roeper, 2021a, 2021b; for American English data, see Foucault et al., 2021, 2022a,b, 2024). Moreover, this seems to be the case for bilingual English-speaking children as well (see Lakshmanan et al., 2022, 2024). Crucially for our current study on adult Romanian-English sequential bilinguals, it is relevant to mention that Foucault et al. (2022a) found that, in set-subset contexts, American English L1 adult controls observe RSSO with very high accuracy, producing recursive set-subset adjective phrases involving gradable size adjectives (small big mushrooms) and interpreting these recursively.

The current study extends the aforementioned research on the syntactic-semantic interface of RSSO as a fundamental UG principle to a bilingual context. Given that the RSSO is a syntactic constraint, there are two possible analyses of recursive set-subset adjectives: an adjunction analysis and a cartographic analysis. According to the adjunction analysis (Kremers, 2003; Abels and Neeleman, 2010), languages differ based on branching directionality (Lakshmanan et al., 2024; Lust, 1983; Lust and Mangione, 1983). From a local perspective within this view, English places adjectives to the left of the noun, whereas Romanian positions them to the right. One could postulate that the RSSO is a further requirement within Adjunction, merging set adjectives to head noun first. In a cartographic approach (Cinque, 1994, 2005, 2010), the underlying UG default would correspond to the English order Subset Set Noun, and the Romanian Noun Set Subset surface order, which is a mirror order of English, would be derived via movement. In particular, it could be achieved through a complex set of operations known as Roll-Up-of-N, which manipulates the basic English adjectival sequence through successive NP movement through specifiers of functional projections. The analyses are illustrated for Romanian in (6).

Note that, in example (6a), the noun phrase (NP) moves from its original position to an outer specifier of FP_set (Step 1). Then, the newly formed FP, containing FP_set, moves again to the outer specifier of the projection hosting FP_subset (Step 2). Deciding between an Adjunction account of RSSO and a Roll UP Account of RSSO is not trivial, and, for this reason, we abstain from committing to either. However, what is important for the current study is that both English and Romanian observe RSSO in recursive set-subset contexts. This is supported by ample empirical evidence from both adults and children speaking these languages as L1.

Compared to studies on monolinguals, there are relatively few studies on bilingual children’s development of recursive structures. The few studies that have compared bilinguals and monolinguals suggest that they handle recursion similarly. For instance, Pérez-Leroux et al. (2017) and Pérez-Leroux et al. (2021) compared the recursive prepositional phrases (PPs) produced by 4- to 6-year-old Spanish-English bilinguals and monolingual English speakers. They found that bilinguals performed similarly to monolinguals with 2- and 3-level PP-recursion (e.g., the dog next to the tree next to the house). Avram et al. (2020) found that 5- to 7-year-old simultaneous Romanian-Hungarian bilinguals had the same acquisition path as their monolingual Romanian children for recursive PPs and relative clauses. Recent studies by Foucault et al. (2024) and Lakshmanan et al. (2022, 2024) found that both monolingual and bilingual American English-speaking children can comprehend and produce recursive possessives and recursive gradable adjectives (big small mushrooms, small big small mushrooms) from as early as age 4, with children between 4 to 6 experiencing more challenges than children between 7 to 12. Both groups found recursive adjectives more challenging than recursive possessives, but their overall mastery was quite high with both. Interestingly, Leandro and Amaral (2014) observed that bilingual Wapichana-English children outperformed monolinguals (up to the age of 7) in interpreting complex genitive constructions, supporting the idea of a bilingual advantage. These findings support the notion that recursion is an innate feature of Universal Grammar.

However, the few studies that have examined the acquisition of recursion by adult L2 speakers showed that they experienced some challenges in comparison to L1 monolingual speakers. For example, Limbach and Adone (2010) found that, although L1 German-L2 English speakers who had acquired English in adulthood were equally successful as L1 English-speaking children in their comprehension of recursive possessives, they were not as successful as L1 English-speaking adults. Nelson (2016) investigated the comprehension of recursive PPs in sequential bilingual adult speakers (L1 English-L2 Spanish, L1 Spanish-L2 English). The study found that while adult L2 speakers experienced more challenges in comprehension than monolinguals, their accuracy with recursion improved with increasing L2 proficiency.

Besides comparing monolingual and bilingual children’s ability to handle recursive phrases, Lakshmanan et al. (2024) investigated whether branching directionality (Left, Right, and Mixed) of nominals in the bilingual children’s non-English language influences their comprehension and production of left-branching recursive possessives and gradable adjectives in English, their stronger language. They assessed 1-level, 2-level, 3-level recursive adjectives, and found that the bilingual children in the Left-branching group (e.g., Mandarin Chinese, Hindi, Kannada) had the lowest accuracy scores, even though their other language matched the branching directionality of the target recursive phrases in English. The Mixed-branching group (e.g., German, Russian) had the highest accuracy scores on all except 3-level recursive adjectives (e.g., small big small mushrooms), for which the Right-branching group (e.g., Spanish, French) had slightly higher scores. However, the differences in accuracy scores between the three groups were not statistically significant, suggesting that further research is needed involving a larger sample size, as well as assessment of both languages of the bilinguals, and not just the stronger one.

Other relevant research on bilinguals focuses on how adjectives are linearly ordered with respect to the noun they modify. Some studies on L2 Spanish indicate that, regardless of L1-L2 differences in the linear ordering of the adjective and noun, advanced learners are able to successfully comprehend and produce the target L2 Spanish orders (Judy et al., 2008; Guijarro-Fuentes et al., 2009; Rothman et al., 2009, 2010). However, Nicoladis (2006) found that L1 French-L2 English preschool bilingual children experienced challenges with the English A N order, which differs from the predominant N A order in their French L1. Interestingly, in cases where the two languages are both Romance languages (e.g., French and Spanish) with predominantly postnominal adjectives, Androutsopoulou et al. (2008) found that French L1-Spanish L2 adult speakers experienced challenges with evaluative adjectives occurring prenominally in Spanish but postnominally in French (e.g., una fantástica película in Spanish / *un fantastique film in French “a fantastic film”).

As stated in the introduction, our study examines how adult Romanian L1-English L2 bilinguals with high level of proficiency in English L2 comprehend recursive set-subset adjectives in L1 (Romanian) and L2 (English). These two languages differ in branching directionality of adjectival modifiers of a head noun: English is prenominal (left-branching), while Romanian is postnominal (right-branching). Consequently, in set-subset contexts, Romanian is a mirror image of English: red small flowers is flori mici roşii ‘flowers small red’. Our study thus constitutes a direct test of RSSO against AOR, as two competing possible sources for adjective ordering and interpretation. While AOR captures ordering preferences of adjectives naming conceptual properties (e.g., size-before-color in English, color-before-size in Romanian), RSSO posits a universally determined, structure-dependent principle, merging set adjectives with the head noun first, and subset adjectives subsequently. By examining recursive adjective structures in both Romanian (right-branching) and English (left-branching), we take advantage of the fact that the two languages instantiate opposite surface linear orders, and we investigate whether they observe the same underlying structural constraint (RSSO). This crosslinguistic comparison allows us to test whether RSSO overrides AOR, even when the surface order is reversed due to branching differences. By observing whether bilinguals adhere to RSSO across languages despite different surface orders, we probe the universality and dominance of structure-dependent mechanisms in linearization, and their interaction with crosslinguistic influence and processing constraints.

According to the Recursive Set Subset Ordering Constraint, set adjectives are merged first to the head noun, and subset adjectives are merged subsequently. Given that RSSO is a UG principle, our prediction regarding its availability is as follows:

This means that when the context is a recursive set/subset context, participants will observe the RSSO constraint by providing the target recursive interpretation and not a coordinative interpretation, regardless of whether the linear order of adjectives is congruent or incongruent with AOR. For example, in a recursive context where the size adjective small represents the set and the color adjective red represents the subset, participants will interpret red small flowers in English and flori mici roşii ‘flowers small red’ in Romanian recursively (as ‘the subset of red flowers within the set of small flowers’) and not coordinatively (as ‘the red and small flowers’). It is important to mention that, because the coordinative interpretation may generally be more frequent, it is possible that the representation associated with the coordinative interpretation would be activated and would need inhibition in both L1 and L2.

At the same time, we acknowledge potential for crosslinguistic influence. Whether this influence is unidirectional or bidirectional depends on the proficiency level and relative dominance of the bilinguals’ L1 and L2. Recall that in relation to adjectival modifiers of the head noun, Romanian is right-branching and English is left-branching. Thus, we expect that crosslinguistic influence could be triggered in several ways. One possible trigger is the simultaneous activation of the branching directionality of both English L2 and Romanian L1 even when only one language is in use. Another possibility is that the crosslinguistic influence could be triggered by the linear ordering of recursive adjectives (from left to right), regardless of the noun: English has the order A_Subset A_Set Noun, whereas Romanian has the order Noun A_Set A_Subset. Given that in the linear ordering, the subset adjectival modifier precedes the set in English, but in Romanian the set precedes the subset, the reverse order may be activated and, if not inhibited, would lead to the non-target interpretation.

In our study, we focus on adult L1 Romanian-L2 English sequential bilinguals who are highly proficient in their L2. We formulate the following prediction in relation to crosslinguistic influence:

Specifically, even when only English is in use, the simultaneous crosslanguage activation of the left-branching directionality of English, as well as of the right-branching directionality of Romanian, could lead to errors with recursive structures in English because of failure to inhibit the more dominant L1. At the same time, given the high L2 proficiency of bilinguals, we cannot reject the potential for English to influence Romanian.

Importantly, another possibility also arises. In light of the competence-performance distinction, it may also be that recursion is well in place but handling it in real time may pose challenges because of performance related issues: for example, inhibiting L1 influence especially when there are time constraints.

The study received approval from the Research Ethics Committee at the University of Bucharest (59/17.02.2022), and participants provided informed consent prior to data collection. The participants were 41 Romanian L1 sequential bilinguals with C1 English⁵: 21 were tested in their L1 (L1 Romanian group), 20 in their L2 (English L2 group, age range: 18–19, 18 Female; 2 Male mean age: 18; 06).⁶ According to the European classification system, C1 learners are advanced users who can understand a wide range of long, complex texts. They can interact and express themselves fluently and spontaneously. All participants were undergraduate students at the University of Bucharest. They were exposed to English L2 from around age 4–5 and were subsequently exposed to it formally in an instructional setting at least two times a week for 2–6 h (from primary school to university). All the participants we tested studied Romanian as a major and English as a minor at the university. Importantly, we tested the two groups of participants independently: one group was assessed in their L1, and the other group was assessed in their L2. This was done to facilitate a monolingual mode for assessment purposes, as well as to avoid a possible priming effect or bias: if the same participant had been tested on both languages, they might have ended up producing more non-target answers simply given the co-activation of two grammars on the same structures. The groups were matched in terms of level of English, exposure to English and frequency of exposure, age range. Moreover, it is important to reiterate that previous research on recursion in bilinguals tended to assess only one language (typically L2), whereas in the current research, both L1 and L2 are assessed.

After providing informed consent, participants completed a story-based, forced choice comprehension task (Tables 1, 2). The task began with 2 practice trials to familiarize them with the task methodology. The experiment itself contained 8 test items⁷ that were interspersed with 4 fillers.⁸ The test items were designed to examine whether participants exhibited a preference for recursive adjectival orders in line with the RSSO (N SetA SubsetA in Romanian, SubsetA SetA N in English) or whether they have a preference for adjective orders that align with Adjective Ordering Restrictions (AOR) in terms of conceptual properties (like Color or Size), as predicted by Cinque (1994, 2010). Recall that AOR holds for both left- and right-branching structures. Thus, according to Cinque, AOR in Romanian (a postnominal adjective language) should mirror the order found in English (a prenominal adjective language). The 8 test items (see Table 1) involved 4 items containing recursive adjectival orders complying with AORs and 4 items containing recursive adjectival orders that are incongruent with AORs.

In the case of AOR-congruent items, adjectives followed the expected AORs. In English, this means SizeA ColorA N (e.g., the small red flowers), while in Romanian, this is the mirror order N ColorA SizeA (e.g., florile roșii mici, ‘flowers-the red small’).

In the case of AOR-incongruent test items, adjectives appeared in a reverse order, going against AOR. In English, this means ColorA SizeA N (e.g., the red small flowers), while in Romanian, this corresponds to N SizeA ColorA (e.g., florile mici roșii, ‘flowers-the small red’).

Importantly, prosody-wise, the items were presented to our participants with a neutral intonation throughout, with no contrastive stress or special emphasis on either of the adjectives. This was done to ensure that any interpretative preferences observed were not influenced by prosodic focus or stress.

Participants were introduced to two characters who each had a set of objects. Each set contained 3 subsets of 2 objects: the first character had objects of the same color but varying in size, while the second character had objects of the same size but varying in color. Participants then indicated who had the target subset/s (e.g., small red flowers). Table 2 exemplifies the procedure used for both Romanian L1 and English L2. Further consult Supplementary materials for examples of practice items and fillers.

This design specifically pits the structure-dependent RSSO against the AOR (e.g., English’s preference for the linear order size-before-color, Romanian’s preference for the linear order color-before-size). If in the set-subset context of our experiment bilingual participants consistently provide recursive set-subset interpretations in line with RSSO across both L1 and L2, regardless of language-specific AORs, this supports the hypothesis that the set-subset structure-dependent constraint guides interpretation more robustly than AORs. Conversely, deviations from this pattern—especially in L2—can provide evidence for processing difficulties in inhibiting language-specific features.

Overall, the results indicated that both the Romanian-L1 group and the English-L2 group performed at a high level of accuracy, interpreting the adjectives recursively in each language (i.e., N A_set A_Subset in Romanian and A_Subset A_set N in English). Figure 2 shows the proportion of the target (accurate) recursive responses for each congruent condition for the two language condition groups. As shown in Figure 2, in the Romanian L1 group the proportion of the target recursive responses was 0.988 (SD = 0.109) in each of the two AOR congruent condition types: Congruent with AOR and Incongruent with AOR. In the English-L2 group, the proportion of the target recursive responses was 0.873 (SD = 0.335 in the Incongruent condition), which is somewhat numerically higher than the target recursive responses in the Congruent condition (0.788 with a SD of 0.412).

We employed a mixed effects logistic regression model. The dependent variable was Response, which was coded binomially (1 = target-like and as 0 = non-target-like). The fixed effects were Language Condition (Romanian L1 vs. English L2), AOR Congruence condition (Congruent/ Incongruent with AOR), and their interaction. The Model also employed by-participant slopes per AOR Congruence Condition type and random effects per Group. Alpha level was set at the 0.05 level of significance. The model showed that the interaction between Language Condition and AOR congruence condition type was not significant [ß = 1.064, SE = 1.788, z = 0.595, 95% CI (−2.440, 4.568), p = 0.552]. Nor was there any significant main effect for the AOR congruence condition [ß = −1.940e-06, SE = 1.690, z = 0.000, 95% CI (−3.324, 3.324), p = 0.999]. Only the main effect for the Language Condition was found to be significant, indicating that, when overall performance was considered (i.e., regardless of AOR Congruence Condition type), performance in English L2 was somewhat less RSSO-compliant than in Romanian L1 [ß = −4.123, SE = 1.863, z = −2.212, 95% CI (−7.774, −0.472), p = 0.027*].

In terms of non-target-like responses, we would like to highlight that in the Romanian L1 group, there was a total of 84 responses to the experimental items, and, of these, 2 (2.38%) were non-target answers (1 in the Congruent with AOR condition and 1 in the Incongruent with AOR condition). In the English L2 group, there was a total of 80 responses in all for the items in these two conditions. Of these responses, 28 (i.e., 35%) constituted non-target answers, with 10 non-target answers in the Incongruent with AOR condition and 18 non-target answers in the Congruent with AOR condition.

To determine whether participants consistently adhered to the RSSO, we used the following criteria. A participant was categorized as adhering to the RSSO if they provided target-like responses on at least 3 out of 4 trials in each of the two congruence with AOR condition types. Based on this criterion, 13/20 participants adhered to the RSSO in English-L2 (18/20 participants in the Incongruent with AOR condition type and 15/20 in the Congruent with AOR type) and 21/21 bilingual adults adhered to the RSSO in Romanian-L1 overall and in both condition types. A Fisher’s exact probability test was conducted to examine whether there were significant differences between the distribution of participants who met and did not meet the criteria in adhering to the RSSO across Romanian-L1 and English-L2. The test revealed that the two language groups differed significantly in relation to the distribution of RSSO-compliant and RSSO-non-compliant participants (p = 0.003448), with a greater likelihood of occurrence of RSSO-compliant participants in the Romanian-L1 language condition. Further comparisons with Fisher’s exact probability test carried out separately for each AOR congruence condition type (Congruent with AOR and Incongruent with AOR) showed that the differences in the distribution of RSSO-compliant and RSSO-non-compliant participants across the two language condition groups were statistically significant only in the case of the recursive adjective orders in the Congruent with AOR condition (p = 0.02069) but not in the case of the Incongruent with AOR condition (p > 0.05).

As mentioned above, 13 participants adhered to the RSSO in English L2, while 7 (i.e., 35% of the 20 participants in the English L2 group) did not. The errors produced by the 7 participants who failed to meet the criteria set for adherence to the RSSO were analyzed. The patterning of the errors is shown in Table 3. Regardless of the congruence condition type, the responses of two of the participants adhered to an adjectival ordering where the color adjective was the set (i.e., closest to the head noun) and the size adjective was further away from the noun (e.g., A_Size A_Color N, such as small yellow leaves). As can be seen from Table 3, one participant who showed context sensitivity interpreted adjectival sequences in an inverse manner instead, consistent with the postnominal L1 Romanian order. Thus, this participant associated the congruent A_Size A_Color N order with contexts where color subsets were picked from size sets, and the incongruent A_Color A_Size N order with contexts where size subsets were picked from color sets. The errors of the 4 remaining participants who did not consistently adhere to the RSSO, revealed a mixed patterning: 2 of them accepted both orders for contexts involving color subsets among size sets, while 2 others accepted both orders for contexts involving size subsets among color sets.

According to our prediction in terms of UG , we expected that bilinguals, in particular highly proficient bilinguals, will adhere to RSSO in both their L1 and L2, given that the Recursive Set Subset Ordering Constraint (RSSO) is a UG principle. This prediction was confirmed by our findings. Bilinguals did indeed show sensitivity to whether an adjective denotes a set or subset in both their first (L1) and second (L2) languages. This suggests that RSSO is available not only for the L1, but also for subsequent language acquisition. Importantly, regardless of the differences in branching directionality between languages (adjectives in Romanian are typically postnominal, whereas adjectives in English are typically prenominal), participants interpreted the adjective closest to the noun (N) as the set and the one further away from the head noun as the subset (98.8% answers were recursive in L1, 83.05% were recursive in L2). The high-performance levels in the two within subject conditions (i.e., orders congruent with AOR, incongruent with AOR) of the L1 Romanian condition is not at all surprising when we consider the long-term and in-depth immersion of the adult bilingual participants in an L1 Romanian environment (in both informal and formal contexts). The fact that the RSSO was adhered to not only in the L1, but also in L2 is remarkable, given that their dominant language was Romanian and more frequently used than English. The onset of exposure to English for our participants occurred around age 3–4. Crucially, they were children and not adults when they began their acquisition of L2 and this could play an important role in their ability to adhere to the RSSO in their L2 as well. Recall that Romanian employs the linear order Noun Set Subset, while English employs the linear order Subset Set Noun. The bilingual participants’ adherence to the RSSO across their two languages, regardless of the differences in branching directionality and linear order, highlights their ability to inhibit the activation of the other language when only one is in use. These findings provide support for activation of the RSSO universal constraint at the syntax-semantics interface across the two different L1 and L2 syntactic systems.

It is important to note that, in recursive set-subset contexts, their adherence to the RSSO takes precedence over Adjective Ordering Restrictions (AORs). Thus, while AORs can explain the linear order and interpretation of coordinative adjectives (where small red flowers is understood as ‘small and red flowers’), they cannot explain the recursive set-subset ordering and interpretation of recursive adjectives (where red small flowers is understood as the subset of red flowers within the set of small flowers). Even if conceptually, it has been argued that speakers may prefer placing the color adjective closer to the noun than the size adjective, most bilinguals in fact choose to interpret the adjective closest to the noun as the set even if it refers to the size when the contexts involve recursive set subsets.

As the results indicated, in the English L2 condition the proportion of target recursive interpretations was numerically lower in the Congruent with AOR condition. In contrast, in the Romanian L1 condition the proportion of target recursive interpretations was numerically equally high across the two Congruence with AOR conditions. Furthermore, the results of the Fisher’s test indicated significant differences between the L1 and L2 conditions only for the Congruent with AOR condition. These findings demonstrate that the participants experienced more difficulties in inhibiting the coordinative interpretation in the English L2 but not in their L1. It further demonstrates that in their English L2, participants had more difficulty with inhibiting a coordinative interpretation in the Congruent with AOR condition but not in the Incongruent with AOR condition type. One explanation for this result for English L2 is that in the Congruent with AOR condition the linear order (size color N) is consistent with the English AOR and is also consistent with the RSSO. However, in each case the interpretation is different. In the case of AOR, it can only be a coordinative interpretation that is not sanctioned by the context. In the case of the RSSO, the interpretation is a recursive set-subset one that is sanctioned by the set/subset context. Having two interpretations of the same linear order sequence of adjectives poses difficulties in processing. In contrast, the Incongruent Condition does not pose the same level of difficulty, as the target adjectival phrase (color size N) is not consistent with the AOR of English and can only be interpreted recursively. Here the coordinative interpretation would not be triggered automatically given that the linear order violates the AOR. Since there is no competing interpretation, participants would be able to direct their attention to the set/subset context in order to interpret the target adjectival phrase.

As noted previously, some researchers have proposed (contra Cinque) that sequential adjectives in Romanian may have a less rigid linear order in terms of AOR (the size adjective can precede the color adjective, and the color adjective can also precede size adjective). Interestingly, in the case of recursive adjectival contexts, the participants interpreted the adjective closest to the head noun as the set, and the adjective further away as the subset. Thus, the fact that bilinguals adhere to RSSO in both languages suggests that their sensitivity to this constraint is rooted in Universal Grammar (UG). Crucially, our findings indicate that they are not simply transferring the language-specific features of their L1 (i.e., post-nominal adjectival modifiers, and a more flexible AOR) into their L2. Instead, as proposed by Flynn et al. (2005), learners construct the new grammar via grammatical mapping from UG. The fact that most of the participants adhered consistently to the RSSO suggests that they were able to keep the L1 and L2 systems separate. Additionally, as illustrated in examples (1) and (2), English and Romanian differ in relation to morphological gender. English is a natural gender language, and Romanian is a language with grammatical gender, where nouns are categorized as masculine/feminine/neuter, and adjectives agree with the gender of the noun they modify. Such overt morphological differences between the two languages could have facilitated their maintenance as two separate systems. Crucially, this additional structural difference between Romanian L1 and English L2 may have made it easier for them to inhibit activation of the postnominal adjectival order of their dominant language, i.e., Romanian, when using English L2. Furthermore, their adherence to RSSO may have been facilitated by their native mastery of Romanian (L1) and their high (C1) proficiency in English (L2). Notably, our findings align with recent research showing that higher second language proficiency leads to a more native-like organization of semantics and semantic memory (see Kogan et al., 2024).

As for our prediction in terms of crosslinguistic influence , given that the dominant language of Romanian L1-English L2 bilinguals is Romanian, we expected Romanian to influence their English to a greater extent. Overall, very few errors were attested in relation to recursive interpretation. The percentage of errors in the L2 English condition was 16.95%. In contrast, the percentage of errors in the L1 condition was negligible (1.2%), which suggests that, despite their high proficiency level in English, their L1 Romanian was impervious to influence from L2 English. Our findings suggest a slight influence of Romanian L1 on English L2 only in the case of the 7 participants who did not meet our criteria for adherence to the RSSO. A possible explanation is that for these 7 participants the adherence to RSSO may have been impacted by the inability to inhibit crosslanguage activation of language-specific features of their Romanian L1, leading to errors in the interpretation of recursive adjectives in L2.

The analysis of the 7 participants’ misinterpretations provides some insights into the mechanisms underlying the acquisition of recursive adjectival order in English L2. Recall that two participants consistently interpreted the prenominal color adjective as representing the set and the size adjective as representing the subset regardless of the condition. Under this view, these participants may be argued to have difficulty inhibiting a coordinate interpretation associated with the L2 AOR order. Such errors may have been further facilitated by prescriptive rules in formal instructional contexts English the correct ordering of adjectives is always ASize AColor Noun. Alternatively, another explanation for this error type is that it may be due to the existence of a universal semantic hierarchy of adjectives, where color properties, which are perceived as more objective than size properties (see Scontras et al., 2017, 2019), are somewhat more salient to participants, and tend to be favored as representing the set. As reported in the results section, one participant demonstrated sensitivity to the set-subset principle but interpreted adjectival sequences in the inverse order, in line with their L1 Romanian. Specifically, the size adjective was interpreted as the set in ASize AColor orders and the color adjective as the set in AColor ASize orders. This suggests failure to inhibit crosslanguage activation of L1, resulting in an interpretation based on a reversed direction of embedding. Recall that a mixed patterning was observed in the case of the errors made by four participants. Specifically, in at least one of the two congruence condition types, they provided the target recursive interpretation for 2 out of 4 trials and chose the non-target interpretation for the other 2 trials. We hypothesize that the mixed patterning of 2 non-target responses versus 2 target responses is not because of lack of L2 competence. On the contrary it may reflect difficulties in performance stemming from an inability to consistently inhibit L1 influence. An alternative explanation is that the four participants relied on a conjunctive default, treating the adjectives as coordinated (e.g., small yellow flowers interpreted as ‘small and yellow’). This strategy has been observed in previous research on child language acquisition in English and Romanian, where some children initially assign a conjunctive interpretation to recursive structures (e.g., flori mici mari ‘big small flowers’ interpreted as ‘big and small flowers’; see Bleotu and Roeper, 2021a, 2021b for Romanian; Foucault et al., 2021, 2022a, 2024 for English).

Recall that, when the distribution of the number of RSSO-compliant and RSSO-non-compliant participants was compared across the two language conditions using the Fisher’s test, significant differences were observed only in the Congruent with AOR condition, but not in the Incongruent with AOR type, with lower performance attested for English in comparison to Romanian. This can be explained in terms of failure on the part of a minority of participants to inhibit the L1 linear word order. Our overall findings, as well as the fine-grained analysis of the errors of the RSSO-non-compliant participants, fail to provide evidence in support of an L1 transfer based explanation, given that only one subject systemically showed difficulty in inhibiting the L1 pattern in L2.

Importantly, grammatical mapping from a UG-determined template (Lust, 2012; Berkes and Flynn, 2016; Fernández-Berkes and Flynn, 2023) helps account for our findings as to how Romanian L1-English L2 bilinguals are able to construct language-specific recursive set-subset structures. As discussed previously, Merge provides the universal minimalist template necessary for building hierarchical structure-dependent phrases. When interpreting recursive set-subset structures, participants take the adjective merged to the head noun first as the set and the adjective merged subsequently as the subset. They must, however, also integrate the branching directionality of the specific language in use (Romanian-right-branching, English-left-branching), ultimately assigning the interpretation Noun A_Set A_Subset in Romanian and A_Subset A_Set Noun in English. In our study, there were very few RSSO-non-compliant answers in English L2. Non-target response types involved: a coordinative interpretation, where the adjectives merge together first (Adj1^Adj2) before modifying the head noun, failure to inhibit the L1 recursive order or failure to inhibit the L2 AOR order. The very low frequency of occurrence of such non-target responses in our data, along with the overwhelming RSSO-compliant responses of most of the participants is in line with the cumulative enhancement model of acquisition (Flynn et al., 2004; Fernández-Berkes and Flynn, 2023).

One issue that needs to be discussed concerns the status of variability in our data, and whether it is sufficient to support our conclusions. As stated previously, one important goal of the current study was to compare how proficient Romanian L1-English L2 bilingual speakers, who were more dominant in their L1, observed the Recursive Set Subset Constraint across their two languages. Thus, we predicted that the group tested on L1 would perform at a very high level, while the group tested on L2 would experience some challenges in inhibiting the L1 influence. Importantly, the high-level performance of the group tested on L1 does not undermine the reliability or validity of the task. Rather, it serves to affirm the status of the Recursive Set Subset Ordering Constraint as a principle of Universal Grammar. In previous L1 research, there is overwhelming evidence in support of greater variability in children than in adult native speakers in relation to the RSSO constraint on adjectives (see Foucault et al., 2021, 2022a and Lakshmanan et al., 2024 for English children. See Bleotu and Roeper, 2021a, 2021b, 2022a, 2022b for Romanian children), as well as in relation to other recursive structures (Pérez-Leroux et al., 2017; Sevcenco and Avram, 2018; Giblin et al., 2019, and others). However, our focus here was on proficient Romanian L1-English L2 bilingual speakers. Additionally, the variability observed in the responses from the bilingual group tested on L2 is sufficient to shed light on the role of UG and cross-linguistic influence, given the existence of 7 participants who did not meet our criteria for target responses (representing 35% of the total number of 20 participants tested on L2). This ensures that the study captures individual variation and is not reliant on homogeneous groups. Thus, we can conclude that the RSSO-compliance of 13 participants (65%) in the English L2 condition compared to the overwhelming RSSO-compliance of all the participants in the Romanian L1 condition overwhelmingly asserts the validity of the universal status of the Recursive Set Subset Constraint. The fact that more non-target interpretations were attested in the L2 English condition than in the L1 Romanian condition points to a potential role for L1 in L2 processing.

Our data highlight a broader issue about how bilinguals navigate competing linguistic systems. At the outset of our paper, we mentioned that the term crosslinguistic influence rather than transfer better captures the interaction between two languages in the bilingual mind from a processing perspective. Acquisition of a new language is a creative process that does not simply entail transfer of language-specific constructions. Our overall findings clearly show that Romanian L1 - English L2 participants consistently observed the RSSO in English L2 regardless of language-specific differences in branching directionality between L2 and L1. In our English L2 data, of the 7 participants who did not consistently observe the RSSO, only one participant consistently exhibited difficulty in inhibiting L1 branching directionality when giving the target set-subset interpretation, while the evidence from the remainder of participants does not support L1 transfer. Instead, a few participants sometimes fail to inhibit L1 or competing coordinative L2 AOR orders, and this interferes with their interpretation. Overall, the crosslinguistic influence supports cumulative enhancement (Flynn et al., 2004; Fernández-Berkes and Flynn, 2023), as very few coordinative interpretations were observed. In other words, having acquired the RSSO in a previous language enables them to acquire their L2 more easily. Interestingly, if any form of transfer may be said to be at work, it is the transfer of a minimalist UG template, not language-specific structures.

Additionally, a few words are in order regarding the theoretical implications of our findings for syntactic accounts of RSSO in terms of Roll-Up and Adjunction. Interestingly, the fact that only 2 non-target responses were attested in Romanian L1, but more non-target responses (28, i.e., 35%) were attested in English L2 is compatible with both theoretical positions as long as we consider that English is the participants’ L2. If we assume, as in Roll-Up, that the default UG order is the prenominal adjectival order (as in English), and that the post-nominal adjectival order of Romanian is derived from this order via syntactic movement, then the higher rate of errors in English may seem unexpected given that bilinguals may be argued to have access to UG, and the English order is considered a default. However, it could nonetheless be explained through their greater familiarity with Romanian L1 and with the non-dominant status of L2. Thus, in spite of the higher syntactic complexity involved in deriving the L1 order, because participants have already established it as the order in their dominant language, they make fewer errors in L1. On the other hand, if we assume that both prenominal and postnominal orders are UG-consistent underlying orders (adjectives are adjoined to the right in Romanian but to the left of the head noun in English), the 7 participants’ difficulty with the prenominal order could again be explained through their challenges with L2, and their greater familiarity with L1. Thus, our results are in effect compatible with both theoretical positions (Roll-Up, Adjunction). Importantly, our paper makes a proposal regarding the universality of the RSSO without committing to a specific implementation (Roll-Up vs. Adjunction). What is crucial here is that the merging of the adjectives is hierarchically constrained by the set-subset adjective type, with set adjectives being merged to the head noun first and subset adjectives being merged subsequently, regardless of branching directionality.

Additionally, cross-linguistic variation in adjective ordering flexibility likely influences these findings. Romanian, often considered more flexible in adjective order regarding conceptual properties such as color or size, may allow the RSSO pattern to more easily override AOR patterns. Therefore, it would be important to conduct similar studies in languages which differ in branching directionality where both instantiate less flexible AORs. This would help determine whether participants face greater challenges with RSSO in languages with less flexible AORs.

Finally, it is worth noting that in our study, the target recursive phrases were presented using a neutral intonation pattern, without any contrastive stress or special emphasis placed on either of the two modifying adjectives. In future research, it would be interesting to examine whether the use of a marked prosodic intonation would impact the interpretation of recursive set-subset adjectives, what this would entail in a stress-timed language (e.g., English) or syllable-timed language (e.g., Romanian), and how it would interact with branching directionality (head-first vs. head-last).

Our study represents the first investigation of how adult sequential bilinguals handle recursive adjectives in contexts involving sets and subsets in their two languages (Romanian L1-English L2). Given participants’ high accuracy with recursion in both languages, our findings highlight the idea that the Recursive Set Subset Ordering Principle constrains bilingual acquisition and use. Our findings also provide evidence of crosslinguistic influence for some participants, which we attribute to language-specific differences in terms of branching directionality, linear order, and Adjective Ordering Restrictions.