Sentence comprehension involves incremental parsing (Frazier, 1987; Frazier and Clifton, 1989; Kamide, 2008). Native speakers typically integrate linguistic cues in real time, using prior input to anticipate upcoming linguistic material. Rather than relying solely on bottom-up information (e.g., the linear input of words as they appear), they engage in predictive processing to guide interpretation. This predictive strategy supports more efficient real-time comprehension (Altmann and Kamide, 1999; Omaki et al., 2014). In second language (L2) processing, research has examined the extent to which L2 learners engage in predictive mechanisms similar to those used by native speakers of a particular target language. A central question in this earlier work was whether L2 learners anticipate upcoming input based on contextual and structural cues, or whether they rely more heavily on reactive, integration-based processing in which information is incorporated only after it is encountered.

Debates about the underlying mechanisms of L2 processing were strongly shaped by the Shallow Structure Hypothesis (SSH). SSH proposes that L2 learners tend to rely more heavily on surface-level cues such as word order and lexical associations than on abstract syntactic structures or hierarchical representations, constructing syntactic representations that are shallower or less detailed than those of native speakers (Clahsen and Felser, 2006; see also Clahsen and Felser, 2018; Felser, 2019 for updated reviews). Supporting this view, cross-modal priming and other studies have reported reduced sensitivity to syntactic dependencies, including filler-gap relations, suggesting that real-time syntactic integration is less robust in L2 processing than in L1 processing (Felser and Roberts, 2007; Marinis et al., 2005).

Although SSH does not explicitly address predictive processing in L2 comprehension, it has often been taken to imply that L2 learners may be limited in their ability to engage in structure-based anticipation because they underutilize abstract syntactic information. At the same time, because SSH concerns the depth of syntactic representations rather than the timing of processing, it remains compatible with the possibility that L2 learners engage in predictive processing, but that their predictions are guided more by surface level cues than by hierarchical syntactic structure. This interpretation highlights an important limitation of the predictive vs. non-predictive debate in L2 processing. Qualitative differences between native speakers and L2 learners may arise not from whether prediction occurs, but from the types of cues that guide anticipatory behavior in each group.

In the context of this shift in focus, recent work has moved beyond the earlier predictive vs. non-predictive framing and moved toward more fine-grained questions about what L2 learners predict, when do they predict it, and what cues support their anticipatory behavior (Grüter and Rohde, 2021; Kaan and Grüter, 2021). This emerging gradient refers to the view that predictive processing is not a categorial yes or no property but rather exists along a continuum, varying in strength and timing depending on how efficiently learners can recruit available linguistic cues in real time. Empirical findings illustrate this variability. Some studies report weaker or less consistent predictive behavior in L2 learners when prediction depends on verb semantics, thematic roles, or grammatical gender (Grüter et al., 2012; Hopp, 2013; Lew-Williams and Fernald, 2010; Grüter et al., 2017). Yet other work indicates that L2 learners can engage in robust predictive parsing when reliable cues are available. They have been shown to anticipate direct object positions based on filler-gap constraints (Omaki and Schulz, 2011), to use prosodic cues flexibly for referential prediction (Nakamura et al., 2021), and to exploit semantic and classifier cues during anticipatory looks (Grüter et al., 2020). Together, these findings suggest that L1 and L2 comprehenders both engage in predictive processing but might differ in the types of cues they rely on during anticipation.

A critical next step, therefore, is to examine whether L2 learners can recruit structure-based cues in contexts where native speakers reliably do so. Filler-gap dependencies offer a particularly revealing test case. Prior work on sentences such as (1), in which a wh-phrase can be linked either to the main clause (MC) verb (tell, i.e., Where did Lizzie tell someone?) or to the embedded clause (EC) verb (catch, i.e., Where was she going to catch butterflies?), has shown that native English speakers exhibit a robust preference for the MC interpretation (Frazier and Clifton, 1989; Omaki et al., 2014). This preference has been attributed to the Active-Gap-Filling Hypothesis and the principle of Minimal Dependency Formation (Phillips, 2006), which assume that comprehenders construct dependencies incrementally and prioritize structurally higher attachment sites. Related evidence from other languages shows similar structurally driven preferences in L1 processing (Nakamura et al., 2025a).

In L2 processing, prior studies have shown that learners are also capable of incremental dependency formation. For example, Omaki and Schulz (2011) demonstrated that L2 learners, like native speakers, process filler-gap dependencies incrementally and are sensitive to island constraints, indicating an ability to construct syntactic dependencies in real time. However, because their study did not examine structurally ambiguous configurations such as (1), this evidence does not address whether learners rely on when multiple interpretations are possible. Thus, while these findings suggest that L2 learners can build filler-gap dependencies incrementally, it remains unclear whether their predictive mechanisms are guided by the same structural cues that drive native speakers' active gap filling.

In addition to investigating the group level differences between native speakers and L2 learners, we also aim to investigate how different types of cues are used within the L2 learner population. Recent research indicates that the ability to compute syntactic dependencies does not always align with global proficiency, and that fine grained grammatical knowledge can be a better predictor of the cues learners rely on during comprehension (Hopp, 2013; Tanner et al., 2013; Cunnings, 2017). These findings suggest that individual differences in predictive behavior may reflect differences in the stability of specific structural representations rather than differences in overall proficiency, highlighting the need to examine variability with the L2 group rather than treating learners as a uniform population.

To investigate this possibility, we tested a relatively homogeneous group of upper intermediate to advanced L2 learners with comparable L2 learning background, allowing us to examine variation in sentence processing mechanisms independently of broad proficiency differences. Within this group, we assessed learners' accuracy on unambiguous filler-gap items such as (2) as an index of structural computation ability, that is, how reliably they construct syntactic dependencies in real time during comprehension. Unambiguous filler-gap sentences such as (2) require learners to identify the grammatically licensed gap position, for which there is only a single structurally valid analysis. Crucially, the correct interpretation in these cases cannot be derived from lexical associations or surface-level heuristics alone. As such, this measure provides a targeted index of learners' consistency in forming syntactic dependencies during online processing, capturing variability in underlying processing mechanisms that is not directly observable through standard proficiency tests.

By combining between-group (L1 vs. L2) and within-group comparison between high vs. low accuracy L2 learners on items such as (2), and by testing how this variation in structural computation ability shapes predictive eye movement patterns during the comprehension of globally ambiguous wh dependencies such as (1), our design provides a critical test of whether L2 sentence processing reflects parsing mechanisms that differ qualitatively from those used by native speakers, or whether L2 processing develops along a gradient path toward native like, structure-based prediction. To answer these questions, we employed a visual-world eye-tracking paradigm, which tracks participants' eye movements as they process spoken wh-questions and provides fine-grained temporal data about their moment-by-moment interpretive commitments. To determine when these commitments emerged, we conducted cluster-based permutation analyses, a data-driven statistical method that identifies time windows in which visual attention significantly diverges between competing interpretations.

We first report overall behavioral results based on key-press responses, followed by the eye-tracking results. Analyses are presented first for the overall comparison between native speakers and all L2 learners, and then for high- and low-accuracy L2 subgroups, defined by comprehension accuracy on unambiguous filler-gap questions.

The present study examined whether L2 learners engage in predictive parsing when processing structurally ambiguous filler-gap dependencies, and the extent to which their behavior resembles that of native speakers. When L2 learners were analyzed as a single group (hereafter, the combined L2 group), both key-press and eye-tracking measures for the ambiguous items showed that they, like native speakers, exhibited a robust preference for the MC interpretations. Crucially, this MC preference cannot be attributed to the selection of the “correct” interpretation, since both MC and EC interpretations were always grammatically available in our ambiguous materials. Instead, the consistent preference for the MC interpretation suggests that L2 learners, like native speakers, attempted to posit a gap at the earliest grammatically licit position, a hall mark of active gap filling in incremental sentence comprehension.

At the same time, the combined L2 analysis also revealed a clear difference from native speakers in the timing and cues that guided prediction. Whereas native speakers showed anticipatory looks toward the MC location only after hearing the matrix verb, the combined L2 group shifted their gaze toward the MC location even before any structural cue licensed an MC gap. Because no grammatical information supported an MC analysis at this point in the sentence, this early anticipatory shift cannot be attributed to a structure-based prediction. A plausible interpretation is that the aggregated L2 pattern reflects sensitivity to distributional regularities in the materials, particularly the frequent pairing of the fronted where with the matrix verb tell. Viewed at the group level, this early bias suggests that L2 learners initially relied on surface-level lexical contingencies rather than structural constraints. It is important to note that this interpretation reflects the averaged pattern obtained from the combined analysis and does not imply that all L2 learners contributed to this early bias in the same way; as the subgroup analyses indicate, this early effect appears to have been primarily driven by the low-accuracy learners, a point we that we elaborate on below.

Crucially, the combined L2 analysis also showed that the later portion of the fixation time course ultimately converged with the native speaker pattern. Both native speakers and the combined L2 group shifted their gaze toward the MC location once the matrix verb was encountered and maintained this preference throughout the EC region. Because this sustained preference cannot be explained by lexical associations alone, it suggests that once a structurally informative cue became available, L2 learners were able to use it to construct a grammatically licensed dependency. When viewed at the level of the combine L2 group, therefore, the pattern appears hybrid: early distributional, lexical cue-based anticipation followed by structurally guided active gap filling once the matrix verb became available.

However, the subgroup analyses revealed that the early and late effects observed in the combined L2 group were not expressed uniformly across all learners. Instead, they reflected two distinct processing profiles that differed in structural computation ability. High-accuracy L2 learners, whose performance on the unambiguous filler-gap items indicates more reliable structural computation ability, showed a fixation time course that closely resembled that of native speakers: they showed no early distributional effect and exhibited a single sustained MC preference once the matrix verb provided a structurally informative cue. Low-accuracy L2 learners, in contrast, showed an early distributional bias toward the MC location before any structural cue was available, and their later MC preference was delayed relative to native speakers and high-accuracy learners, and was fragmented across shorter clusters. These differences indicate that weaker structural computation ability is associated with greater reliance on surface-level lexical contingencies and delayed use of structural cues. Importantly, this suggests why the combined L2 analysis showed a hybrid pattern: the early lexical effect reflected the behavior of the low accuracy subgroup, while the later structural effect reflected the behavior of both groups, although more robustly for high accuracy learners.

Additional support for the interpretation that low-accuracy group rely more heavily on surface-level cues comes from the key press responses. While the high-accuracy group, like native speakers, showed a clear preference for the MC interpretation at the final decision stage, the low-accuracy group did not. Although eye-tracking data revealed MC commitment during online processing in the low-accuracy group, this commitment was not reflected in their key press choices. Instead, the low-accuracy group showed no reliable preference for the MC interpretation over the EC alternative. One possible explanation is that their structural representations are less robust, which makes their final decision highly sensitive to recency based lexical cues, that is, cues stemming from the most recently encountered portion of the sentence. Because the EC relevant material appears at the end of the question, this late occurring lexical information may have exerted a strong influence on the final choice. For example, after hearing “catch butterflies” in (4a), some low-accuracy learners selected the EC answer option, suggesting that the lexical recency cue may have overridden their earlier MC expectation. This seems to be the most plausible explanation of the observed pattern and is compatible with the idea that when structural representations are not firmly established, surface-level cues can pull comprehenders toward a competing interpretation, leading to a less stable overall commitment that differs between the real time fixation pattern and the final explicit choice.

To account for this L2 processing profile, we propose that L2 learners rely on a cue weighting mechanism in which the use of predictive cues is modulated by the learner's structural computation ability. Learners draw on a broad set of cues in real time, including cues that are not grammatically informative but non-etheless carry predictive value, such as distributional regularities and lexical co-occurrence patterns. L2 learners with weak or unstable structural representations place greater weight on these readily accessible surface-level cues, resulting in earlier non-structural predictions and a less robust structural commitment later in the sentence. In contrast, L2 learners with stronger structural computation ability weight structurally informative cues more heavily, and their resulting prediction pattern resembles that of native speakers. The idea that L2 learners with weaker structural computation ability would rely on information that native speakers do not use for prediction, and would in fact commit to a predictive interpretation even earlier than native speakers, might appear counterintuitive, particularly given proposals that L2 comprehension is constrained by working memory limitations (Cunnings and Felser, 2013; Cunnings, 2017; Kim and Christianson, 2017). However, a cue weighting perspective provides a rational explanation, since the early prediction observed in the low accuracy group is not structurally guided but instead arises from surface-level lexical associations. When structural computation is insufficiently stable, drawing on surface-level cues that are immediately available in the input allows learners to generate predictive commitments under real-time processing constraints (for frequency-based accounts of cue reliance, see Ellis, 2002).

This interpretation shares certain assumptions with the SSH (Clahsen and Felser, 2006), which holds that L2 learners rely less on abstract syntactic representations and more on lexical-semantic information during processing. However, our findings extend this hypothesis by showing that these differences emerge specifically in predictive processing. More importantly, our results reveal a pattern not explicitly predicted by SSH: the nature of predictive parsing in L2 learners depends on the stability or reliability of their structural representations. High-accuracy learners, whose computation of filler-gap dependencies was more reliable, displayed native-like structural prediction. Low-accuracy learners, in contrast, relied more heavily on surface-level contingencies such as lexical co-occurrence patterns, resulting in predictive processing that was less guided by structural information and more susceptible to moment-to-moment lexical cues.

This graded pattern suggests a developmental view in which L2 processing shifts from frequency-based prediction toward structure-based prediction as structural representations become more stable. This interpretation aligns with convergence-based accounts (Hopp, 2013; Cunnings, 2017), which hold that L2 learners can approximate native-like parsing mechanisms when their underlying grammatical representations are sufficiently robust. Rather than a categorical deficit in syntactic prediction, the present findings point to a continuum in which cue use changes as L2 grammatical knowledge becomes more established.

Importantly, our results also indicate that the L2 pattern observed here cannot be explained as a simple transfer of L1 processing strategies. In Japanese, the native language of our L2 participants, wh-phrases remain in situ and the interpretation of question depends on clause-final cues such as the question particle (Kuroda, 1965; Saito, 2017). Processing studies show that native Japanese speakers often postpone interpretation until they encounter this clause-final marker (Nakamura et al., 2025b; though see Aoshima et al., 2004; for evidence that Japanese speakers actively posit gap sites). In contrast, Japanese L2 learners of English in our study did not delay interpretation. Instead, they showed early commitment to the MC interpretation. This indicates that they adopted a more predictive parsing strategy characteristic of English wh-dependency resolution, rather than relying on a processing strategy typical of Japanese.

For future research, a central methodological challenge concerns how best to characterize individual differences in structural computation ability. While our high–low grouping used in the present study, based on accuracy for unambiguous filler–gap sentences, provides a theoretically motivated proxy for structural computation ability, future progress will require more fine-grained and independent measures of structural sensitivity that can better isolate the specific structural computational factors underlying individual variation in L2 predictive processing. Future research should also examine the generality of these cue-weighting patterns across syntactic structures and processing contexts. In particular, experimentally manipulating distributional environments will allow direct tests of whether early surface-driven predictions reflect the reliability of L2 learners' syntactic computation during line sentence processing. Finally, longitudinal or training-based studies may help illuminate how individual differences in structural representations develop over time and how these changes contribute to the transition from frequency-based to structure-based prediction in L2 comprehension.