Sequential Stroop Task Analysis

• Sequential Stroop Task is a cognitive paradigm that examines conflict adaptation by analyzing performance differences in incongruent trials following congruent versus incongruent trials.
• It holds the current trial constant while varying the previous trial’s congruency, enabling precise isolation of cross-trial conflict effects.
• Recent computational studies and VLM implementations support that improved performance in II sequences, as measured by token log probabilities and accuracy, signals dynamic cognitive control.

Searching arXiv for the cited papers and related work on the Sequential Stroop Task. The sequential Stroop task is a variant of the Stroop paradigm in which performance on a current trial is analyzed as a function of the congruency of the immediately preceding trial. In the underlying color-naming task, the target is the ink color and the distractor is the written color word; congruent trials align word and color, whereas incongruent trials mismatch them. The sequential version is used to examine conflict adaptation, conventionally operationalized as improved performance on an incongruent trial when it follows another incongruent trial rather than a congruent one. Recent work has also distinguished true cross-trial sequential effects from models that are only sequential within a trial, such as settling networks with discrete time steps but no trial-history dependence (Hu, 28 Oct 2025, Prabhakaran et al., 5 Feb 2025).

1. Paradigm, notation, and inferential target

The core logic of the sequential Stroop task is to hold the current trial type fixed while varying the congruency of the previous trial. In the standard notation used in recent model work, the four two-trial sequence types are as follows.

Sequence	First trial	Second trial
CC	congruent	congruent
CI	congruent	incongruent
IC	incongruent	congruent
II	incongruent	incongruent

The theoretically central contrast is between CI and II, because both sequences end with an incongruent current trial. If performance on the second trial is better in II than in CI, this is treated as evidence of conflict adaptation. One recent formulation states this directly: improved performance on incongruent trials when they follow other incongruent trials compared to congruent trials is taken as a signature of cognitive control (Hu, 28 Oct 2025).

A common source of terminological slippage is that “sequential” can refer either to cross-trial history effects or merely to temporal evolution within a single trial. Several computational papers relevant to Stroop interference implement the latter but not the former. In those cases, the system is sequential in the sense of repeated updates over discrete time steps, yet trials remain independent and there is no prior-trial congruency factor in the classical sense (Prabhakaran et al., 5 Feb 2025, Şengör et al., 2015).

2. Experimental and computational instantiations

A direct implementation of the sequential Stroop task for vision-LLMs presents an image containing two color words in colored fonts, arranged either left-to-right or top-to-bottom. The first word is treated as the previous trial and the second as the current trial. Each word has both semantic content and font color, each can be congruent or incongruent, and the two words do not share any colors in either modality. The color set comprises red, blue, green, yellow, pink, and brown. For each arrangement, the design yields 30 CC, 120 CI, 120 IC, and 360 II sequences. Models are instructed to report the ink colors in order, using exactly two words (Hu, 28 Oct 2025).

This VLM adaptation preserves the previous-trial congruency logic but changes the temporal ontology of the task. The model processes the entire image-prompt pair statically; the “previous trial” is embedded within the same input rather than supplied as a separate earlier event. The sequential dependence is therefore conceptual and stimulus-structural rather than a consequence of persistent model state across samples. This suggests a family resemblance to human sequential Stroop designs without identity of processing assumptions (Hu, 28 Oct 2025).

Other recent Stroop-related systems are important precisely because they are not classical sequential Stroop tasks. In the BiLex self-organizing-map model, each Stroop trial is simulated independently from a complete factorial design over 16 colors: 16 congruent cases, 16 no-input cases, and 240 incongruent cases. The target is always the ink color, and the lexical input can be congruent, incongruent, or absent, but there is no trial-to-trial carryover and no dynamic control adjustment across trials (Prabhakaran et al., 5 Feb 2025). In the cortex–basal ganglia–thalamus model of dopamine and action selection, the Stroop effect is produced by multi-step within-trial competition between word reading and color naming, again without explicit sequences of trials (Şengör et al., 2015). In the eye-tracking study based on a modified visual Stroop test, each condition is a separate static display with 16 areas of interest, the four condition-trials are randomized and balanced, and no cC/cI/iC/iI analysis is performed (Rossi et al., 2020).

3. Dependent measures and analytical contrasts

In classical sequential Stroop analysis, two constructs are fundamental. The first is the conflict cost, defined as the performance difference between incongruent and congruent trials. The second is conflict adaptation, defined as modulation of that cost by previous-trial congruency; in the sequential notation above, the critical test is II versus CI on the second trial (Hu, 28 Oct 2025).

Because reaction times are not meaningful for vision-LLMs in the same way as for human participants, the VLM formulation replaces human RT with token-level uncertainty measures. Each model is evaluated on all sequence types and both spatial arrangements using the log probabilities assigned to correct second color tokens, alongside second-token accuracy. Higher log probability and higher accuracy for II than CI constitute the relevant adaptation signature in that setting (Hu, 28 Oct 2025).

In time-resolved settling models, a different decision variable is used. In the BiLex implementation, time is discrete, activations evolve synchronously, and the lexical map’s uncertainty is quantified by entropy,

$E_t = -\sum_{c \in C} L(c,t)\log L(c,t).$

A response is emitted when lexical-map entropy falls below 1 bit, and the number of time steps to threshold is defined as response time. This is a sequential decision process in the sense of threshold crossing over time, but it is still a single-trial measure rather than a previous-trial congruency effect (Prabhakaran et al., 5 Feb 2025).

The eye-tracking Stroop variant operationalizes difficulty through aggregate gaze statistics rather than classical sequential contrasts. The extracted variables include number of fixations, average fixation length, maximum fixation length, horizontal regressions, vertical regressions, and multiple saccade summaries. The strongest reported fixation-level contrast is Naming With Interference versus Naming Without Interference, where number of fixations, horizontal regressions, vertical regressions, average fixation length, and maximum fixation length all differ significantly. These results characterize conflict and attentive load, but they do not instantiate conflict adaptation because the analysis is per condition rather than by trial history (Rossi et al., 2020).

4. Within-trial dynamics and mechanistic accounts of interference

The BiLex model implements the Stroop task with two 20×20 self-organizing maps: a semantic map trained on RGB color vectors and a lexical map trained on phonetic feature vectors for Spanish color words. Associative connections encode semantic-to-lexical translation, while explicit lateral connectivity within each map combines local Gaussian excitation with global uniform inhibition, yielding a Mexican-hat–like interaction profile. A Stroop trial begins with semantic color input and, depending on condition, congruent lexical input, incongruent lexical input, or no lexical input. Over discrete time steps, lexical activity reflects competition between the initial lexical seed and semantic-driven correction. The reported outcome is an overall accuracy of 84.2%, with 100% accuracy in congruent trials, 87.5% in no-input trials, and 83.0% in incongruent trials; incongruent trials are about 55 steps slower than congruent trials on average, and no-input trials about 36 steps slower than congruent trials (Prabhakaran et al., 5 Feb 2025).

This mechanism attributes Stroop interference to temporal competition within the lexical map rather than to an explicit conflict module. Congruent trials align semantic and lexical sources, so entropy drops rapidly. Incongruent trials initially seed the wrong lexical attractor, which is amplified by lateral dynamics and only later counteracted by semantic input. The model’s central claim is that this interference is a side effect of optimizing overall speed and accuracy through nonzero lexical routing strength. On that view, Stroop interference is not a separately optimized penalty but a cost of efficient average performance (Prabhakaran et al., 5 Feb 2025).

A distinct mechanistic account is provided by the dopamine-sensitive cortex–basal ganglia–thalamus model. Here, word reading is treated as habitual and color naming as task-relevant. Two competing C–BG–Th loops represent the two actions, and a slower cortico–cortico C1–C2–C3 loop performs error detection and attentional adjustment. Nigrostriatal dopamine is modeled as a parameter shifting the striatal activation function, thereby altering whether actions can be initiated and how effectively the system resolves competition. Under normal dopamine, the word-reading tendency initially rises faster, but control biases eventually allow color naming to win; dopamine excess prolongs competition, whereas dopamine depletion yields errors and delayed correction (Şengör et al., 2015).

Neither model is a true sequential Stroop model in the narrow experimental sense. Both are temporally explicit, both generate conflict resolution trajectories, and both supply candidate internal signals for future trial-history models, but neither implements CI versus II comparisons or trial-to-trial control adaptation (Prabhakaran et al., 5 Feb 2025, Şengör et al., 2015).

5. Conflict adaptation in vision-LLMs

The most direct recent computational study of the sequential Stroop task evaluates 13 open-source vision-LLMs from the Gemma, InternVL 3.5, Molmo, and Qwen2.5-VL families. Across left-right and top-bottom arrangements, 12 of 13 models show higher log probability for the correct second token under II than under CI, and condition accuracy beneath the bars indicates lower error rates in II than CI. Spatial arrangement does not meaningfully affect the observed pattern. The sole exception is Qwen2.5-VL 72B Instruct, which shows the reverse ordering, but the interpretation offered is a ceiling effect because both Qwen2.5-VL 32B and 72B achieve 100% accuracy across all conditions (Hu, 28 Oct 2025).

The same study adds a mechanistic interpretability layer using transcoders, a sparse-autoencoder variant, on InternVL 3.5 4B pretrained. Coactivation-based grouping is used to construct inter-layer feature networks and extract connected components termed supernodes. For the color red and the text RED, early- and late-layer supernodes are identified, with partial overlap between text and color features. In early layers, the text supernode contains more features than the color supernode; in late layers, the reverse holds. The paper interprets this asymmetry as mirroring the fact that reading is more automatic than color naming in humans even in the absence of interference (Hu, 28 Oct 2025).

Conflict-specific structure is then isolated with the summary tensor

$$A(c1 \neq t1, c2 \neq t2) - A(c1 == t1, c2 \neq t2),$$

which contrasts II-type against CI-type settings while holding the second trial incongruent. This analysis reveals a conflict-modulated supernode in layers 24–25 that is more active under II than CI. Causal ablation of that supernode increases Stroop errors 3.38-fold on CI trials, from 17.5% to 59.2%, and 8.33-fold on II trials, from 2.5% to 20.8%, while having minimal effect on congruent trials. After ablation, the model sometimes outputs “ink” instead of the second color, a failure mode absent before ablation (Hu, 28 Oct 2025).

The theoretical interpretation is deliberately cautious. The existence of II greater than CI effects and conflict-modulated supernodes suggests control-like internal organization, yet task comprehension remains a confound. A plausible implication is that part of the observed adaptation may reflect the model inferring the instruction “ignore the word; name the color” more clearly from conflict-laden inputs, rather than implementing a human-like scarce-resource control process (Hu, 28 Oct 2025).

6. Scope, misconceptions, and extension paths

The most important conceptual clarification is that not every Stroop paradigm with time dependence is a sequential Stroop task. A two-word image processed in a single forward pass, a settling network with entropy thresholding inside one trial, a dopamine model with repeated state updates before response selection, and a four-screen eye-tracking study with randomized condition order are all temporally structured. None, however, automatically instantiate the previous-trial congruency manipulations denoted by CC, CI, IC, and II in the classical sequential sense (Hu, 28 Oct 2025, Prabhakaran et al., 5 Feb 2025, Şengör et al., 2015, Rossi et al., 2020).

This distinction matters methodologically. In the BiLex model, routing is reduced to static parameters such as $r_{\mathrm{sem}}$, $r_{\mathrm{lex}}$, and $\gamma$, and there is no conflict monitoring or control module. In the dopamine model, C2 and C3 provide within-trial conflict detection and threshold adjustment, but control settings reset rather than persist across trials. In the eye-tracking study, each subject contributes only four long condition-trials, so no genuine cC/cI/iC/iI analysis is possible (Prabhakaran et al., 5 Feb 2025, Şengör et al., 2015, Rossi et al., 2020).

Several extension paths have been identified explicitly. For BiLex, future routing mechanisms based on the Conditional Routing Model and contrastive Hebbian learning are proposed so that routing becomes adaptive and task-dependent rather than fixed; the paper also suggests that within-trial variables such as time to threshold, entropy, or mismatch between semantic and lexical peaks could serve as conflict signals for later trials, though these are not yet implemented (Prabhakaran et al., 5 Feb 2025). For the dopamine model, plausible extensions include slow control states, tonic dopamine dynamics, residual activity across trials, and synaptic plasticity, all of which could convert intra-trial conflict correction into cross-trial conflict adaptation (Şengör et al., 2015). For eye-tracking paradigms, adding pupillary response, collecting larger datasets, and shifting from per-condition summaries to trial-by-trial modeling would make classical sequential analyses feasible (Rossi et al., 2020).

Taken together, the contemporary literature supports a layered understanding of the sequential Stroop task. At the narrowest level, it is a previous-trial congruency paradigm centered on CI versus II comparisons. At a broader mechanistic level, it serves as a template for studying how interference unfolds over time, how control-like variables might be represented internally, and how those variables could, in principle, be carried from one trial to the next. The current frontier lies in connecting these two levels without conflating within-trial settling dynamics with genuine sequential adaptation (Hu, 28 Oct 2025, Prabhakaran et al., 5 Feb 2025).