Think-Aloud Utterances (TAUs) in Cognitive Research

• Think-Aloud Utterances (TAUs) are sequential verbalizations produced during task performance that reveal immediate cognitive processes and strategy shifts.
• They are elicited using concurrent protocols and segmented using pause, semantic, or event-based methods to align speech with observable actions.
• TAUs bridge behavioral data with underlying reasoning, enhancing model development in domains such as usability testing, tutoring, and human–AI interaction.

Searching arXiv for recent and relevant papers on think-aloud utterances to ground the article. Think-Aloud Utterances (TAUs) are the time-ordered verbalizations produced while a person is performing a task and narrating ongoing cognition, rather than retrospectively explaining it after completion. Across usability research, educational measurement, intelligent tutoring systems, cognitive modeling, software engineering, and human–AI interaction, TAUs are treated as process data: they expose evolving interpretations, subgoals, uncertainty, strategy shifts, and evaluative judgments at a granularity finer than final answers or behavioral logs alone. The literature does not impose a single universal unitization rule. Some studies define TAUs as continuous speech segments tied to specific stimuli or actions, some segment by pause or semantic unit, some align utterances to task events, and some avoid utterance-level segmentation entirely and analyze full-trial transcripts or derived symbolic structures (Murali et al., 2023, Wurgaft et al., 29 May 2025). This heterogeneity reflects the fact that TAUs are simultaneously a methodological construct, a coding unit, and an increasingly important interface between human verbal protocols and automated analysis pipelines.

1. Definitions and conceptual scope

In concurrent think-aloud traditions, participants are asked to “say everything that comes to mind” while solving a task; the resulting TAUs are verbal traces of working-memory contents rather than post hoc reconstructions (Reinhart et al., 2019, Bujak et al., 2010). Within that family, specific domains instantiate the unit differently.

In usability testing on web stimuli, a TAU has been defined as each continuous segment of speech in which a participant refers to one or more page elements; these segments are time-stamped and linked to manually labeled Areas of Interest (AOIs) (Murali et al., 2023). In text-evaluation research, each discrete thought verbalized during concurrent think-aloud becomes a TAU, typically a sentence or short clause expressing an evaluative consideration about Coherence, Fluency, Consistency, or Relevance (Chu et al., 2024). In tutoring-system studies, TAUs are operationalized as concurrent utterances aligned to the next graded action, such as a hint request or attempt, so that each utterance becomes a predictor of immediate performance or self-regulated learning state (Borchers et al., 1 Feb 2026, Borchers et al., 2023). In software-refactoring research, TAUs are transcribed alongside code snapshots and action labels, and the utterance fragments are used to infer grounded-theory categories of reasoning about code quality (Oliveira et al., 2024). In dialog-personality modeling, TAUs are defined more narrowly as the verbalized thought immediately before an actual dialog turn, and are inserted before a speaker’s turn as augmentation for persona modeling (Ishikura et al., 10 Oct 2025).

Some work extends the notion beyond human participants. THiNK treats a model’s self-reflective step-by-step commentary during iterative revision as a TAU-like object, formalized as

$$u_i^{(r)} = \mathrm{LLM}_\theta\!\left(p_i^{(r-1)}, IS(p_i^{(r-1)})\right),$$

with the revised problem extracted from that sequence (Yu et al., 26 May 2025). InteractEval likewise treats both human and LLM concurrent verbalizations as TAUs that seed checklist construction (Chu et al., 2024). By contrast, work on large-scale automated verbal-protocol analysis for the Game of 24 explicitly does not define a TAU unit; instead, each trial transcript is treated as a single document and converted into a search graph (Wurgaft et al., 29 May 2025). This makes clear that TAUs are not synonymous with any single segmentation rule.

A consistent cross-domain property is that TAUs are valued because they preserve temporal order. Whether the target construct is attention to interface elements, model-based reasoning in electronics, self-regulation in tutoring, or cognitive model discovery, the central premise is that utterances proximal to actions, perceptions, or decisions can reveal mechanisms that outcome data alone underdetermine (Ríos et al., 2019, Xie et al., 6 May 2026).

2. Elicitation and recording protocols

TAU collection typically uses concurrent think-aloud with minimal prompting. In statistics education, best-practice guidance emphasizes that interviewers should remain silent except for neutral reminders such as “Please keep talking,” so that utterances remain as unprompted as possible (Reinhart et al., 2019). The physics-curriculum comparison study used a tic-tac-toe warm-up, then asked students solving Force Concept Inventory items to “say everything that comes to mind,” with the researcher prompting “What are you thinking now?” after roughly ten seconds of silence (Bujak et al., 2010). In usability sessions with Chinese non-native and native English speakers, three think-aloud protocols were contrasted: Classic Think-Aloud (CTA), Speech-Communication (SC), and Interactive Think-Aloud (ITA), with ITA adding active probing through five prompt types (Fan et al., 2022). CTA and SC produced similar verbalization distributions, whereas ITA increased Explanation utterances only (Fan et al., 2022).

Recording configurations depend on domain. In web-usability TAU–AOI linking, three scrollable screenshots from real municipal websites were shown on a 21.5″ display at $1920\times1080$; audio was captured by on-camera microphone at 48 kHz, gaze by a Tobii 4C Pro infrared eye tracker at 90 Hz, and mouse positions were logged at millisecond accuracy using common global timestamps (Murali et al., 2023). In risky decision-making for automated cognitive model discovery, audio was recorded by head-mounted microphone, with onset time-locked to option display and offset at button press during a 19-trial task (Xie et al., 6 May 2026). In chemistry tutoring studies, sessions were audio-recorded via built-in microphones or Zoom while tutor logs simultaneously captured step attempts and hint requests (Borchers et al., 2023). In PointAloud, live audio is streamed from a browser front end to Deepgram Nova-3, with downstream segmentation, labeling, summarization, and relevance scoring handled by GPT-4o and spatial reference inference by Gemini 2.5 Pro (Gmeiner et al., 10 Feb 2026).

Warm-up tasks are common where the literature reports them explicitly. Statistics-education guidance recommends a low-stakes warm-up such as describing how to make toast (Reinhart et al., 2019). The physics study used tic-tac-toe for acclimation (Bujak et al., 2010). The refactoring study included a 15-minute sample exercise and practice in thinking aloud before the recorded block (Oliveira et al., 2024). Other papers note no formal warm-up or do not report the exact instruction wording, as in the eye-tracking study of data evaluation with small and large datasets (Benz et al., 29 May 2026).

The choice between concurrent and retrospective protocols is theoretically consequential. Several sources state that concurrent think-aloud is preferred because it reduces retrospection bias and better captures immediate reasoning rather than reconstructed explanations (Reinhart et al., 2019, Chu et al., 2024). A plausible implication is that TAUs are most informative when the research goal concerns local process variables—attention, self-regulation, or mechanism discovery—rather than polished justifications.

3. Segmentation, alignment, and multimodal synchronization

Segmentation strategies differ sharply across studies, but the dominant patterns are pause/semantic segmentation, event-window alignment, and semantic chunking.

In the Chinese/non-native English usability study, two researchers segmented transcripts by pause or semantic unit before assigning one of five categories: Procedure, Reading, Observation, Explanation, or Others (Fan et al., 2022). In the data-evaluation eye-tracking study, the coding unit was any stretch of speech at least half a sentence long, with multiple codes allowed per segment (Benz et al., 29 May 2026). PointAloud formalizes segmentation with both silence and topic shift:

$$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$

so a new TalkNote is created when either a long pause or an LLM-inferred topic change is detected (Gmeiner et al., 10 Feb 2026).

In tutoring-system work, segmentation is driven by automatic speech-to-text and action logs. Whisper segments recordings into utterances with timestamps, and utterances are then grouped between successive ITS events so that each analysis unit corresponds to the interval preceding the next student action (Borchers et al., 2023). A related gaming-the-system study assigns each utterance to the most recent preceding action and groups utterances linked to eight consecutive actions into one clip; utterance length is then aggregated within clip $k$ as

$$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$

(Zhang et al., 8 Jan 2026).

The web-usability TAU–AOI study illustrates full multimodal synchronization. Each TAU $i$ is an interval $[t s_i, t e_i]$. Gaze points and mouse points are extracted over that interval:

$$G_i = \{(x_g(t),y_g(t)) \mid t \in [t s_i, t e_i]\},$$

$$M_i = \{(x_m(t),y_m(t)) \mid t \in [t s_i, t e_i]\}.$$

For each manually annotated AOI $A_j$, the system checks whether $1920\times1080$0 or $1920\times1080$1 intersects $1920\times1080$2, thereby linking spoken references to specific webpage elements (Murali et al., 2023). This yields a paired dataset of the form $1920\times1080$3 and supports downstream sentiment maps and expert filtering (Murali et al., 2023).

Not all research segments at the utterance level. “Scaling up the think-aloud method” processes each Game of 24 trial transcript as a whole and converts it into a search graph via an LLM “coder” and GraphBuilder API, with automatic checking and repair of generated code (Wurgaft et al., 29 May 2025). The absence of utterance segmentation in that work is methodologically important: it shows that some process-level insights can be recovered from aggregate verbal traces when the representational target is symbolic search structure rather than discourse function.

4. Coding schemes and representational frameworks

TAU analysis ranges from hand-built categorical codebooks to latent use within prompting contexts.

Several studies use explicit coding taxonomies. The Chinese/non-native English usability study employed five categories: Procedure, Reading, Observation, Explanation, and Others (Fan et al., 2022). The refactoring study used grounded theory to derive eight mutually exclusive reasoning categories: Unnecessary Code, Complex Code Structure, Long Method, Improve Readability, Improve Performance, Improve Maintainability, Code Comprehension, and Semantic Preservation (Oliveira et al., 2024). In the chemistry tutoring and gaming studies, utterances were labeled for four self-regulated learning categories derived from Winne and Hadwin: Processing Information, Planning, Enacting, and Realizing Errors (Borchers et al., 2023, Zhang et al., 8 Jan 2026). In data-evaluation research with eye tracking, a 14-code manual covered references to measurement data, experimental setup, force and position trends, single-point focus, and several forms of measurement uncertainty (Benz et al., 29 May 2026).

Other codebooks are domain-theoretic. The electronics-lab assessment study coded 30-second intervals against the five Modeling Framework subtasks: Make Measurements, Construct Models, Make Comparisons, Propose Causes, and Enact Revisions (Ríos et al., 2019). The physics-curriculum comparison study used question-specific code sets for two Force Concept Inventory items, including categories such as “Constant velocity ⇒ no acceleration,” explicit mention of momentum, “F=ma explicitly,” “air resistance,” and “Drew a diagram” (Bujak et al., 2010).

A distinct representational move appears in large-scale automated analyses. In the Game of 24 study, whole-trial transcripts are mapped into search graphs whose node types are game states and edge types are arithmetic operations or subgoal-setting moves (Wurgaft et al., 29 May 2025). Reliability is then assessed through normalized graph edit distance,

$1920\times1080$4

rather than utterance-level agreement (Wurgaft et al., 29 May 2025).

Still other systems treat TAUs as raw text context rather than codeable objects. In automated cognitive model discovery for risky choice, the full trial-by-trial transcript is supplied verbatim to LLaMA-3.1 alongside gamble parameters and choices, and no manual features enter the model (Xie et al., 6 May 2026). The effective evaluation objective remains behavioral BIC,

$1920\times1080$5

with no explicit $1920\times1080$6 term; TAUs instead act as prompt-level soft constraints on model generation (Xie et al., 6 May 2026). This suggests that TAUs may function either as direct annotations or as latent regularizers, depending on pipeline architecture.

5. Empirical findings across domains

Across domains, TAUs have been shown to reveal structure that is weakly recoverable from behavior alone.

In the web-usability study, 344 statements were recorded, of which 231 explicitly referenced AOIs. Hit rates were computed over verbally mentioned AOIs:

$1920\times1080$7

The denominator was 231 verbally mentioned AOIs. Gaze achieved $1920\times1080$8; mouse achieved $1920\times1080$9. A paired-sample $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$0-test yielded $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$1, with $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$2, indicating a significantly higher hit rate for gaze than mouse (Murali et al., 2023). The authors attribute this to passive mouse use for scrolling and tighter coupling between gaze and visual attention (Murali et al., 2023).

In automated cognitive model discovery, inclusion of think-aloud transcripts improved held-out BIC for 59.7% of participants, with a within-subject paired $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$3-test of BIC differences giving $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$4 and mean difference $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$5 BIC points, 95% CI $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$6, favoring the think-aloud condition (Xie et al., 6 May 2026). Structural consequences were also substantial: 69.4% of participants switched discovered model clusters between behavior-only and think-aloud conditions, with a predominant shift from Explicit comparator toward Integrated utility (Xie et al., 6 May 2026). This is one of the clearest demonstrations that TAUs can reshape mechanistic inference, not merely improve predictive fit.

In the Chinese/non-native English speaker study, native and non-native groups showed similar verbalization distributions and similar UX-problem patterns. The ANOVA found no main effect of Language, $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$7, but did find a Protocol effect, $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$8, and Category effect, $$\mathrm{Segment}(U_{t-1}, U_t)= \begin{cases} 1, & \text{if } \Delta t_{\mathrm{silence}} > 8\text{ s } \lor \mathrm{LLM}_{\mathrm{seg}}(U_{t-1},U_t)>\theta,\ 0, & \text{otherwise}, \end{cases}$$9 (Fan et al., 2022). No significant Language effect appeared for problem encounters or actual problems, and no significant Protocol effect appeared on actual problems (Fan et al., 2022). Observation utterances remained the strongest problem indicator across protocols (Fan et al., 2022).

In refactoring education, 203 reasoning occurrences were coded. Unnecessary Code accounted for 61 occurrences (30%), Complex Code Structure 18 (9%), Code Comprehension 11 (5%), Long Method 7 (3%), Improve Readability 7 (3%), Semantic Preservation 6 (3%), Improve Performance 5 (2%), and Improve Maintainability 3 (1%) (Oliveira et al., 2024). Novices concentrated more on issue presence categories 1–3, accounting for 84% of novice utterances versus 61% for experienced students, while quality-attribute reasoning categories 4–6 made up 2% of novice versus 11% of experienced utterances (Oliveira et al., 2024). The study also reports that students often overlooked long-method and suboptimal-loop issues (Oliveira et al., 2024).

In the physics-curriculum comparison, TAUs clarified why traditional students outperformed Matter and Interactions students on selected Force Concept Inventory items. On Q2, 60% of traditional versus 43% of M&I students chose the correct balancing-forces answer; 71% of M&I versus 50% of traditional students verbalized an incorrect net force in the direction of motion (Bujak et al., 2010). On Q8, traditional students were correct 90% of the time versus 57% for M&I; among M&I students, 5 of 14 explicitly invoked $k$0 and incorrectly inferred different fall times, whereas only 2 of 18 traditional students mentioned $k$1, and 14 of 18 correct traditional responses rested on “mass doesn’t matter” or “common knowledge” (Bujak et al., 2010). TAUs thus exposed not merely answer disparities but distinct error-generating reasoning chains.

In chemistry tutoring, TAUs have been used to relate self-regulation to immediate correctness. In one study, processing information predicted lower odds of a correct next step $k$2, realizing errors also predicted lower odds $k$3, while enacting predicted higher odds $k$4; adding SRL-cycle features significantly improved fit, $k$5 (Borchers et al., 2023). A later gaming-the-system study found that gaming clips had mean total words per clip of 105 versus 74 for non-gaming, and mean non-stop words 29 versus 20. A Poisson generalized linear mixed model yielded $k$6 for all words and $k$7 for non-stop words, both $k$8 (Zhang et al., 8 Jan 2026). Gaming clips were also more likely to contain Processing Information $k$9, less likely to contain Planning $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$0, and more likely to contain Realizing Errors $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$1 (Zhang et al., 8 Jan 2026).

In large-scale process analysis of the Game of 24, 640 participants generated 4,947 analyzed think-aloud trials after filtering. Division-only problems had 20% solve rate versus 59% for non-division problems, and trials with at least one subgoal had 65% success versus 53% without subgoals, both $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$2 (Wurgaft et al., 29 May 2025). Although that work does not segment TAUs, it demonstrates the continued value of verbal reports when scaled to thousands of trials.

6. Automation, LLM mediation, and hybrid human–AI pipelines

Recent work increasingly treats TAUs as machine-readable process traces rather than solely qualitative interview data.

Automation begins with transcription and diarization. In usability testing, OpenAI Whisper “medium” provides word-level timestamps and Pyannote performs speaker diarization so moderator and tester speech can be split automatically, with tester TAUs extracted without manual correction (Murali et al., 2023). In chemistry tutoring and gaming studies, Whisper-generated timestamps support alignment of utterances to tutor actions (Borchers et al., 2023, Zhang et al., 8 Jan 2026). In large-scale verbal protocol analysis, Whisper large-v3 transcribes audio, a Llama 3.3 classifier filters irrelevant transcripts, and a second LLM translates task-relevant transcripts into executable code for search-graph construction, with up to five repair iterations and temperature increases when errors persist (Wurgaft et al., 29 May 2025).

Hybrid human–LLM settings use TAUs to scaffold evaluation and documentation. InteractEval collects concurrent think-aloud outputs from four human experts and four LLMs, merges roughly 40–60 utterances per dimension, and runs a four-stage pipeline of component extraction, attribute clustering, question generation, and question validation to build evaluation checklists (Chu et al., 2024). Human TAUs were reported as stronger for internal quality dimensions such as Coherence and Fluency, while LLM TAUs performed better for external alignment dimensions such as Consistency and Relevance; combined TA peaked at roughly 85 utterances per dimension and produced the best evaluation outcomes (Chu et al., 2024). Diversity was quantified using self-ROUGE-L, Jaccard, and SBERT cosine similarity, with lower similarity interpreted as greater diversity (Chu et al., 2024).

THiNK uses a seven-agent loop aligned to Bloom’s Taxonomy, where agents $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$3 through $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$4 score a candidate problem and agent $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$5 provides improvement suggestions. Problem quality is defined as

$$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$6

with acceptance if $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$7 (Yu et al., 26 May 2025). TAUs are not directly scored, but the loop treats them as the revision mechanism through which feedback is internalized. Average performance improved in Remembering by $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$8, and Analyze, Evaluate, and Create each improved by more than 1 point after the TAU loop (Yu et al., 26 May 2025). The paper interprets this as evidence that structured feedback encourages richer reflective outputs in higher-order tasks (Yu et al., 26 May 2025).

PointAloud operationalizes TAUs as segmented, labeled “TalkNotes” spatially anchored to pointer traces and design-workspace context (Gmeiner et al., 10 Feb 2026). Each TalkNote may receive one or more labels from a six-category taxonomy: Design Intent, Process, ToDo, Important, Problem, and Question (Gmeiner et al., 10 Feb 2026). The system also generates cursor-adjacent TalkTips by periodically prompting GPT-4o and showing only those candidates whose relevance $$L_k = \sum_{i \in \text{utterances in clip }k} \#\text{words}_i$$9 exceeds threshold $i$0 (Gmeiner et al., 10 Feb 2026). In a within-subjects study with 12 professional designers, PointAloud improved suggestion relevance $i$1, process awareness $i$2, task support $i$3, and think-aloud support $i$4, with no change in distraction $i$5 (Gmeiner et al., 10 Feb 2026).

Other LLM-oriented work is more cautionary. When GPT-4.1 was prompted to continue student chemistry TAUs, the generated utterances were more coherent with context than human utterances, with local coherence differences $i$6 and $i$7, both $i$8 (Borchers et al., 1 Feb 2026). Model utterances were approximately twice as long on average, sentences were approximately 30% longer, and type–token ratio was approximately 20% lower than in human TAUs (Borchers et al., 1 Feb 2026). The authors interpret these outputs as overly coherent, verbose, and confident simulations of novice reasoning (Borchers et al., 1 Feb 2026). This raises a methodological caution: synthetic TAUs may be useful as engineering artifacts, but they are not automatically faithful proxies for human cognition.

7. Methodological issues, limitations, and implications

The literature identifies several recurring sources of error and several persistent misconceptions about TAUs.

A first misconception is that think-aloud always requires fine-grained utterance coding. Large-scale Game of 24 analysis shows that useful process representations can instead be built at the trial level through search graphs (Wurgaft et al., 29 May 2025). A second misconception is that mouse movement or surface interaction alone can substitute for attention-linked verbal context. In the TAU–AOI usability study, gaze substantially outperformed mouse in hit rate because the mouse was used primarily for scrolling rather than pointing (Murali et al., 2023). A third misconception is that LLM-generated think-alouds are automatically cognitively realistic. Evidence from chemistry tutoring indicates that they may be systematically more coherent and confident than authentic novice reasoning (Borchers et al., 1 Feb 2026).

Limitations often enter through segmentation, coding, and synchronization decisions. In the web-usability study, manual AOI definition may omit small elements, and speech–gaze latency means a verbal comment may precede or follow fixation by a few hundred milliseconds (Murali et al., 2023). In the data-evaluation study, the paper reports no formal eye-tracking-to-code regressions, only qualitative linkage between trend-based utterances and trend fixations (Benz et al., 29 May 2026). In several studies, prompts or instructions are only partially specified, which constrains reproducibility (Benz et al., 29 May 2026). The statistics-education guidance explicitly recommends reporting warm-up scripts, prompts, and coding rubrics to support transparency and replication (Reinhart et al., 2019).

Reliability practices vary. Some studies report substantial or near-perfect agreement, such as $i$9 initially and $[t s_i, t e_i]$0 after discussion in the data-evaluation coding (Benz et al., 29 May 2026), or Cohen’s $[t s_i, t e_i]$1 values of 0.78, 0.90, 0.77, and 1.00 across four SRL codes in tutoring data (Borchers et al., 2023). Others resolve disagreements by unanimous discussion without formal $[t s_i, t e_i]$2, as in the refactoring study (Oliveira et al., 2024). “Scaling up the think-aloud method” uses normalized graph edit distance rather than coder agreement on verbal categories (Wurgaft et al., 29 May 2025). This suggests that reliability standards are currently representation-dependent rather than unified.

The strongest general implication is that TAUs increasingly function as a bridge variable between observable behavior and latent process. In cognitive model discovery, they constrain mechanistic search beyond behavior-only trajectories (Xie et al., 6 May 2026). In intelligent tutoring, they expose the phase and sequencing of self-regulation cycles in ways that step logs alone do not (Borchers et al., 2023, Zhang et al., 8 Jan 2026). In usability, they can be grounded to visual context through gaze and pointer traces (Murali et al., 2023, Gmeiner et al., 10 Feb 2026). In educational assessment, they reveal why students arrive at correct or incorrect answers, rather than merely whether they do so (Bujak et al., 2010, Ríos et al., 2019). This suggests that TAUs are best understood not as ancillary commentary but as structured process evidence whose value depends on careful elicitation, alignment, and representation.