Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Sequential Sampling and Choice Dynamics

Abstract. Sequential sampling theories explain choice as a temporal process in which uncertain information changes a latent decision state until a stopping rule produces a response. This fundamental research monograph reconstructs the historical, mathematical, experimental, and neuroscientific development of sequential choice models. It compares the drift-diffusion process, racing and linear ballistic accumulators, leaky competition, decision field theory, multialternative selection, dynamic boundaries, and gaze-dependent value accumulation. The evidence review examines joint choice and response-time distributions, speed–accuracy control, neural population and electrophysiological signals, recognition and preference tasks, individual differences, hierarchical estimation, model recovery, and mechanisms of experimental discrimination. Particular attention is paid to the distinction between a good behavioral fit and independently established psychological or biological implementation. Recent comparative studies in 2025 and 2026 are integrated with classic primary research. The synthesis identifies sequential accumulation as a productive and testable family of explanations and finds that distinguishing its component mechanisms requires targeted interventions, temporally resolved measurements, and stringent generative validation. The work is a critical narrative synthesis of existing literature; it reports no new experiment, registered systematic review, or original clinical recommendations.

Keywords: sequential sampling; choice dynamics; evidence accumulation; drift-diffusion model; diffusion decision model; linear ballistic accumulator; leaky competing accumulator; decision field theory; response-time distributions; speed–accuracy tradeoff; first-passage time; model recovery; parameter recovery; multialternative decision making; perceptual decisions; value-based choice; neural decision signals; computational cognitive modeling; Artificial Era

Introduction: Sequential Evidence as a Theory of Choice

Sequential sampling models describe a decision as a temporally extended transformation from uncertain observations into a categorical commitment. Their central proposition is that information arriving over time changes a latent decision state until an action rule terminates deliberation. The scientific object is therefore a joint distribution of choices and their latencies, including correct and erroneous responses, rather than an isolated percentage of correct judgments. A model becomes informative when its accumulation law, starting conditions, stopping criterion, and nondecision components imply testable patterns across difficulty, incentives, prior probability, and time pressure. The diffusion tradition provides especially explicit connections between those behavioral observables and candidate processing operations Ratcliff & McKoon (2008) Forstmann et al. (2016).

The scope of this monograph is the general mathematical and psychological organization of evidence-driven selection: two-bound diffusion, continuous and discrete random walks, independent races, leaky and competitive accumulators, dynamic stopping policies, multialternative choice, value-guided attention, and estimation. Perceptual evidence is an important testing ground, yet recognition, value comparison, memory retrieval, and action selection present related questions about how uncertainty changes while a choice unfolds. The evidence is assessed across original theory papers, experiments, methodological tests, neurophysiology, and research reviews, with an emphasis on what specific task designs establish Usher & McClelland (2001) Brown & Heathcote (2008) Gold & Shadlen (2007).

The method is a critical narrative scientific synthesis of original and reviewed literature. It does not constitute a registered systematic review, exhaustive bibliometric census, new experimental study, or independent quantitative meta-analysis. The guiding questions are how evidence becomes a decision variable, which mechanisms generate observable time distributions, what task changes identify drift or caution, when rival models mimic one another, and which neural signals constrain formal explanation. Interpreting a fitted decision parameter as a unique psychological property requires model recovery, discriminating manipulations, and convergent measurement; predictive adequacy alone licenses a narrower claim Wilson & Collins (2019) O'Connell et al. (2018).

The Joint Evidence of Choice, Accuracy, and Response Time

Every response-time trial contains more information than a correct-versus-incorrect coding. The choice itself identifies the selected boundary or winning accumulator; latency indexes the duration from stimulus or decision cue to recorded response. Across repeated trials, the conditional latency distribution for each choice provides shape, spread, tail behavior, and differences between correct and error trajectories. A sequential model must account for these features simultaneously, under the actual stimulus contingencies and response mapping. A theory fitted only to average decision times may conceal discrepancies in distributional tails or error latencies that are decisive for distinguishing mechanisms Van Zandt (2000) Ratcliff & McKoon (2008).

Conditional accuracy functions and quantile-probability plots expose relationships that a single mean collapses. When sensory quality declines, decisions often become slower and less accurate, but the relative latency of errors changes with task structure and with variability between trials. A speed instruction can compress response distributions through threshold adjustments while a perceptual degradation changes the quality of sampled information. These manipulations predict different shapes rather than one universal latency shift. Experimental discrimination therefore depends on factorial designs, enough trials for both response categories, and analysis that retains the outcome attached to each time observation Voss et al. (2004) Ratcliff et al. (2016).

Several measurements require special treatment. Very fast guesses can have a different generating mechanism from the main decision process; lapses can produce slow responses disconnected from evidence; and missed deadlines select the portion of the latency distribution actually observed. Accuracy near ceiling leaves little information about incorrect decisions, whereas highly difficult tasks can render accuracy almost uninformative about a particular boundary asymmetry. The observation process, response device, and trial-exclusion rules belong in the likelihood or analysis plan. Parameter recovery under realistic contaminants is consequently part of scientific interpretation rather than a separate software detail Ratcliff & Tuerlinckx (2002) Lerche et al. (2017).

Historical Formation of Sequential Decision Models

The conceptual ancestry of evidence accumulation lies in sequential statistical testing, stochastic processes, psychophysics, and the measurement of choice reaction time. Sequential testing asks when observations are sufficiently informative to justify stopping rather than collecting another sample. Stone's 1960 analysis explicitly connected a Wald-style sequential decision procedure with the means and variances of choice latencies, error rates, and stimulus frequencies. This connection supplied a rigorous explanation of why caution and information quality interact in speeded selection. Historical continuity concerns the logic of sampling and stopping, even as psychological theories alter the definition of a sample Stone (1960) Bogacz et al. (2006).

Ratcliff's 1978 account of memory retrieval placed continuous stochastic evidence accumulation within an explicit psychological theory, linking recognition decisions to accuracy, response-time distributions, and competing retrieval outcomes. The model's historical significance includes its attention to the entire latency distribution and to response errors, which could otherwise be treated as incidental. Later diffusion applications generalized beyond the specific memory account, while modifications distinguished momentary noise, between-trial variability, starting bias, and nondecision time. A contemporary standard diffusion model should therefore be understood as a developed research family with empirically motivated assumptions, rather than a single immutable 1978 equation Ratcliff (1978) Ratcliff et al. (2016).

The 1990s and 2000s broadened the field through decision field theory, competitive accumulator architectures, and neurally motivated decision variables. Busemeyer and Townsend modeled preference dynamically under uncertain attributes, Usher and McClelland made leakage and competition central to perceptual choice, and Brown and Heathcote demonstrated the empirical reach of a deterministic-within-trial accumulation race. Neurophysiological work then permitted direct comparisons between behavioral accumulation predictions and temporally evolving brain activity. These strands produced a field in which several mathematically distinct mechanisms can satisfy similar benchmark behaviors, sharpening the requirement for discriminating tests Busemeyer & Townsend (1993) Usher & McClelland (2001) Brown & Heathcote (2008) Gold & Shadlen (2007).

Sampling, Likelihood, and Stopping Rules

A sequential decision algorithm specifies a stream of observations, a rule for translating each observation into evidence, and a stopping condition. In idealized binary inference, each increment can be a log-likelihood ratio favoring one hypothesis over another; the accumulated sum then records relative support. Under restrictive assumptions of stable evidence quality and a prescribed error constraint, a sequential likelihood-ratio procedure provides an efficient stopping architecture. Psychological applications typically replace exact likelihood knowledge with noisy, learned, or approximate evidence, creating a scientific question about how well the organism represents the task's statistical structure Stone (1960) Bogacz et al. (2006).

The relevant latent state is not an observed neural trace or a raw stimulus amplitude. It is a construct defined by a generative model: a scalar difference between alternatives, a vector of alternative-specific activations, or a set of fluctuating preferences. Decisions occur when a state hits a fixed or time-varying boundary, a race has a winner, or an external deadline interrupts the process. These variants imply distinct predictions for early errors, choice reversals, sensitivity to late samples, and the allocation of attention. Their formal differences must be linked to experimental conditions before being interpreted as psychological alternatives Usher & McClelland (2001) Roe et al. (2001) Hawkins et al. (2015).

Optimality has a precise conditional meaning. A policy may maximize accuracy for a given mean sampling time, minimize time for an error constraint, or maximize reward earned per total trial duration; these objectives yield different stopping policies once rewards, penalties, nondecision delays, and intertrial intervals change. Bogacz and colleagues analyzed formal relationships between several common two-choice models and the drift-diffusion architecture under particular assumptions. The result supports strong normative benchmarks, but establishes neither that all natural decisions are optimal nor that one behavioral fit identifies the brain's implementation Bogacz et al. (2006) Gold & Shadlen (2007).

The Two-Bound Drift-Diffusion Process

In its elementary form, the diffusion decision model treats a one-dimensional evidence state x as evolving according to a drift term plus stochastic fluctuations: dx = v dt + s dW. The state begins at z between lower and upper absorbing boundaries separated by a distance a; the first boundary reached determines the response. Drift v represents directional evidence quality in the model, s is the noise scale, and boundary separation controls the amount of accumulated evidence demanded before commitment. The initial state z represents a predecision asymmetry in support or response propensity. These quantities are model coordinates whose scale conventions must be declared for comparison Ratcliff & McKoon (2008) Ratcliff et al. (2016).

The predicted decision time is a first-passage time, not the total measured response time. An additional nondecision component commonly represents encoding, motor execution, and other operations outside the modeled accumulation interval. The observed latency is their sum under the model's assumptions. A high decision threshold tends to improve accuracy at the cost of longer deliberation, whereas stronger drift generally increases accuracy while shortening many decisions. Starting the process closer to one bound biases choice toward that alternative. The predictions are joint and distributional; they cannot be reduced to a universal one-parameter interpretation of fast or slow behavior Ratcliff & McKoon (2008) Voss et al. (2004).

Extensions add across-trial fluctuations in drift, starting position, and nondecision time, because ostensibly identical trials may begin with different internal states or evidence quality. Such variability helps account for empirical differences between the distribution of correct and incorrect response times. However, additional flexibility creates identifiability concerns, particularly when individual conditions supply few errors or limited latency spread. When the model's parameters are estimated, the report must state the noise scaling convention, boundary parameterization, allowed variability terms, contaminant process, and which parameters were shared or permitted to change across conditions Ratcliff & Tuerlinckx (2002) Lerche et al. (2017).

Interpretation of Drift, Boundaries, Bias, and Nondecision Time

An attractive feature of diffusion modeling is the correspondence between parameters and experimentally manipulable processing functions. Voss, Rothermund, and Voss tested this interpretation with discrimination tasks that varied accuracy motivation, sensory difficulty, response demands, and reward contingencies. The study found patterns consistent with higher boundaries under accuracy emphasis, reduced drift with difficult discriminations, longer motor-related components under a more demanding response, and biased selection with reward manipulation. Such parameter-selective experiments support construct interpretation, while the validity of any particular inference remains conditional on the fitted model and experimental design Voss et al. (2004).

Drift is frequently described as processing efficiency, yet this phrase can be misleading across tasks. In a perceptual judgment the drift may vary with sensory signal strength; in a recognition experiment it may encode diagnostic memory evidence; in a preference task it may depend on a relative value or on attention-weighted attributes. A drift estimate combines the task's evidence scale with the model's observation assumptions. A numerical comparison across experiments is meaningful only under defensible scale calibration and comparable decision rules. The label therefore identifies a generative role, while its psychological interpretation requires additional task theory Ratcliff & McKoon (2008) Krajbich et al. (2010).

Boundary separation is similarly linked to response caution, but caution is a task-dependent control variable rather than a timeless personality score. A speed–accuracy instruction can alter both preparation and decision policy, and participants may change strategy within blocks. Starting bias may arise from prior probabilities, payoff differences, learned expectations, or preferential encoding; the data alone seldom identify one source without targeted manipulation. Nondecision time aggregates processes that can change independently, and changes in motor dynamics may be misallocated to accumulation when the model is misspecified. Interpretational discipline consists in mapping parameters to independent experimental evidence Ratcliff et al. (2016) O'Connell et al. (2018).

Response-Time Densities and Diagnostic Distribution Shapes

First-passage models generate complete probability density functions for choices and decision latencies. The upper-bound and lower-bound distributions jointly determine accuracy, the relative frequency of each response, and the timing of errors. This is mathematically consequential: a parameter estimate supported by full distributions can differ from one derived by a few summary statistics because quantiles constrain tails and asymmetries. Response time is positive, often skewed, and influenced by nondecision operations; ordinary Gaussian residual assumptions on raw latencies need not respect these features. Model fitting should therefore use an appropriate likelihood or a carefully validated distributional approximation Van Zandt (2000) Ratcliff & Tuerlinckx (2002).

Quantile patterns can reveal differential mechanisms under experimental manipulations. An increase in evidence quality can disproportionately shift slow tails, whereas a change in boundary can alter both accuracy and the distribution of time spent deliberating. Between-trial drift variability may lead to erroneous decisions associated with relatively weak evidence and longer latencies; starting-point variability can yield rapid errors when the process begins near the incorrect boundary. These patterns are tendencies of particular parameterizations, not context-free laws about every model or every experiment. They become evidence when the proposed parameter restrictions predict held-out quantile structure Ratcliff & McKoon (2008) Ratcliff et al. (2016).

A complete data audit also looks beyond the central fitted quantiles. Very long tails can reflect attention lapses, fatigue, strategic waiting, or a genuine prolonged decision; extremely short responses may reflect anticipations, guesses, or recording errors. A misspecified contaminant model can make diffusion parameters appear unstable, and a flexible contaminant model can conceal poor fit of the substantive mechanism. Diagnostic plots should distinguish experimental conditions, selected responses, quantile bins, and trial order, and should examine whether simulations reproduce the observed frequency of both kinds of extreme event Ratcliff & Tuerlinckx (2002) Wilson & Collins (2019).

Leaky and Competing Accumulators

The leaky, competing accumulator framework replaces a single opponent-coded decision state with multiple evolving activations. Each response unit integrates evidence for its alternative, loses some accumulated activation through leakage, and inhibits competing units. Noise enters the activation dynamics, and a response is generated when a unit reaches a decision criterion. Leakage changes the effective memory of earlier evidence, while inhibition changes the degree to which support for one alternative suppresses its rivals. Usher and McClelland's 2001 account developed these principles in perceptual choice and tested them against standard reaction-time and controlled-timing paradigms Usher & McClelland (2001) Forstmann et al. (2016).

Leakage, inhibition, and input normalization can interact so that a model looks approximately diffusion-like in a restricted two-choice regime while producing different predictions for sequentially changing information or larger choice sets. A late contrary signal may be weighted differently depending on how much early evidence has decayed. Competition can create nonlinear changes in response activation as the relative strengths of alternatives shift. These mechanisms offer concrete ways to study temporal weighting rather than simply relabeling drift and threshold. The relevant evidence comes from experiments that vary the order and duration of information exposure and test the resulting time-sensitive choice patterns Usher & McClelland (2001) O'Connell et al. (2018).

Model distinctions should be evaluated after accounting for shared predictions. Both a leaky competitive process and a diffusion process can produce gradual accuracy improvements with increased deliberation and skewed reaction-time distributions. A successful fit to these basic regularities does not establish either leakage or inhibition as the operative mechanism. Strong comparisons manipulate the strength of late evidence, the number and similarity of alternatives, and response deadlines, then inspect distributional and trial-wise effects. Computational model recovery is essential when added mechanisms make one model flexible enough to mimic a simpler competitor Palminteri et al. (2017) Wilson & Collins (2019).

Independent Races and Linear Ballistic Accumulation

In race models, the alternatives are represented by distinct accumulator processes, and the first to reach its threshold determines the choice. Brown and Heathcote's linear ballistic accumulator, introduced in 2008, proposes a particularly tractable form: within each trial, the accumulator states advance linearly at fixed rates; variability is allocated to starting positions and rates across trials. Independent racers do not require lateral inhibition during the trial. Analytic expressions for choice and latency allow the same general framework to handle binary and multiple-alternative tasks, facilitating estimation and comparison Brown & Heathcote (2008) Forstmann et al. (2016).

The contrast between a ballistic race and continuously noisy diffusion is theoretically important. In the ballistic version, apparent trial-level stochasticity arises because the sampled starting state and rate differ across repetitions; in diffusion, the path itself fluctuates within a trial. Yet observing only an endpoint and a response time may offer insufficient resolution to distinguish the two fluctuation sources. A race can recover many canonical response-time effects, including skewness and changes under speed emphasis. The right experimental question is which patterns a particular race cannot reproduce under parameters constrained by another condition Brown & Heathcote (2008) Ratcliff et al. (2016).

Multiple-response tasks create an especially informative setting because the number of competitors can change without forcing an artificial binary coding of every response. The probability of each winner depends on its rate distribution, threshold, and starting variability relative to every other accumulator. Response-set size can affect both choice and latency even if sensory discriminability for an individual item remains constant. However, the assumption of independent races may be violated when response units inhibit one another, options share attributes, or attention alternates between candidates. These cases require comparison against competition-based models with explicit cross-option interactions Brown & Heathcote (2008) Roe et al. (2001).

Dynamic Preference and Decision Field Theory

Decision field theory treats preference as an evolving state influenced by uncertain attributes, fluctuating attention, and the evaluative consequences of each sampled comparison. Busemeyer and Townsend's 1993 theory linked sequential preference formation to phenomena including time-pressure changes, preference reversals, and violations of elementary stochastic choice constraints. In this account, the sampling process is not confined to external sensory noise; the sampled attribute and its subjective evaluation also vary. The mathematical dynamics therefore permit a choice option to gain and lose support during deliberation, even when the option descriptions remain unchanged Busemeyer & Townsend (1993) Roe et al. (2001).

The multialternative extension by Roe, Busemeyer, and Townsend articulated interactions among options, contextual attraction, and differential suppression of similar alternatives. When a third option enters a choice set, its effect can depend on its relationship to existing options, so a simple independent random-utility substitution rule need not describe the resulting frequencies. The theory provides a dynamic explanation of preference construction across time: attributes are sampled, relative valences change, and competition shapes the approach to a commitment boundary. The relevant predictions concern choice proportions, deliberation times, and context effects jointly Roe et al. (2001) Busemeyer & Townsend (1993).

Preference models are particularly vulnerable to flexible post hoc explanation because the attribute sampling sequence and attentional weights may be partially unobserved. Strong tests require experimental control or independent measurement of attention, prespecified option similarities, and held-out choice sets. One must distinguish the discovery that context affects preference from evidence for a specific dynamical equation. In addition, a theory validated for uncertain multiattribute selection may not immediately transfer to a fixed sensory discrimination without an account of how each domain generates evidence increments. Comparative modeling preserves this difference in psychological substance Roe et al. (2001) Palminteri et al. (2017).

Two Choices, Multiple Choices, and Winner Selection

The two-alternative task affords a special simplification: evidence favoring one option can sometimes be represented as evidence against the other, producing one relative decision coordinate. With three or more alternatives, a single scalar is rarely sufficient to represent all relevant comparisons. The modeler must choose between a set of independent accumulators, a coupled vector process, a sequence of pairwise comparisons, or another explicit choice rule. These alternatives make distinct assumptions about what happens to the unselected candidates while attention or sampling favors one option Churchland et al. (2008) Brown & Heathcote (2008).

Churchland, Kiani, and Shadlen studied monkeys making two- and four-choice motion decisions and observed patterns of neural activity in the lateral intraparietal area consistent with evidence accumulation across alternative counts. Their study is important because it connects task complexity, neural dynamics, and decision commitment within a shared experimental setting. However, a neural signal approaching a terminal level is compatible with multiple mechanistic interpretations, and the activity of one recorded population does not fully reveal network-wide competition. Neural support is strongest when quantitative model predictions survive changes in the option set and response mapping Churchland et al. (2008) Gold & Shadlen (2007).

Krajbich and Rangel extended a value-based drift-diffusion approach to decisions among three food alternatives while tracking gaze. Their model related gaze allocation to choice and response time, and the experiment supported continuity of certain computational principles between binary and trinary value decisions. This illustrates a productive bridge between independent race and coupled preference theories: measured fixation sequences constrain the information sampling process, reducing the freedom to explain any final choice after the fact. The number of alternatives becomes both an experimental manipulation and a stress test for assumptions about parallel versus serial comparison Krajbich & Rangel (2011) Krajbich et al. (2010).

Speed–Accuracy Policies and Reward Optimization

The speed–accuracy tradeoff is a central signature of sequential stopping. Raising a decision threshold generally increases the evidence demanded before commitment, which can improve accuracy while adding latency; lowering the threshold does the reverse. A more cautious policy is not always normatively preferable, because additional deliberation carries opportunity costs. Formal optimality depends on feedback, the reward for each outcome, time penalties, and the delay between trials. Bogacz and colleagues demonstrated that specific diffusion-related decision algorithms can implement optimal or approximately optimal policies under explicit task assumptions Bogacz et al. (2006) Ratcliff & McKoon (2008).

A model that estimates wider boundaries after an accuracy instruction provides a plausible process explanation, but the manipulation itself may also affect sensory engagement, anticipation, attention, and action preparation. This is why parameter-specific validation experiments matter. Voss and colleagues found experimentally induced selectivity consistent with diffusion parameter interpretations, while later studies of speed–accuracy manipulation have also emphasized that parameter estimates should be evaluated together with the task procedure. A single contrast in mean reaction time does not identify boundary adjustment as the only behavioral change Voss et al. (2004) Ratcliff et al. (2016).

Reward-rate optimization illustrates how the theoretically optimal response policy can depend on experimental timing. If one trial's extra second reduces the number of future rewarded trials, relatively low thresholds may be optimal; if errors are costly, the balance shifts. Even within a condition, participants can depart from a stationary optimum because they learn reward contingencies, anticipate block changes, or assign subjective values that differ from experimenter-defined payoffs. Distinguishing objective optimality from fitted behavior requires reporting the reward function, time budget, observed policy, and assumptions about trial-to-trial stationarity Bogacz et al. (2006) Busemeyer & Townsend (1993).

Collapsing Bounds, Urgency Signals, and Time Pressure

Fixed stopping boundaries are a convenient and empirically successful baseline, but some theories propose that less evidence is required as time elapses. A collapsing boundary can make late decisions easier to terminate, and an urgency signal may transform decision activity through a time-dependent gain. These proposals arise naturally when an approaching deadline or increasing opportunity cost makes prolonged deliberation less valuable. Their psychological interpretation depends on what actually changes over time: the criterion, input gain, motor readiness, or a combination of processes Hawkins et al. (2015) O'Connell et al. (2018).

Hawkins and colleagues compared fixed versus collapsing boundaries across human and nonhuman-primate datasets and multiple experimental paradigms. Their reported analysis generally favored fixed boundaries for the evaluated tasks while acknowledging procedure-dependent differences and difficulties of model discrimination. The result is a reminder that a plausible normative advantage of collapse is not direct empirical confirmation of collapsing boundaries in every task. Model selection is sensitive to the available distributional information, included alternatives, fitting constraints, and the particular late-decision region in which the hypotheses diverge Hawkins et al. (2015) Ratcliff et al. (2016).

Experiments should distinguish externally imposed deadlines from spontaneous urgency by varying deadline timing, payoff schedules, and the temporal profile of evidence. Neural signals that rise with elapsed time may mix evidence, motor preparation, and an urgency-related variable. Behavioral evidence is strongest if an urgency model predicts conditional choice accuracy at long latencies and response distributions under changed deadlines better than a well-specified fixed-bound alternative. The scientific question is a discriminable control mechanism, not the mere observation that many decisions occur before a deadline Hawkins et al. (2015) O'Connell et al. (2018).

Time-Varying Evidence, Change of Mind, and Temporal Weighting

Many standard diffusion derivations assume a stable mean drift within a trial. Natural observations change, and carefully designed experiments can manipulate the direction, strength, or reliability of evidence at different moments. A sequence of temporally controlled pulses reveals whether early and late information contribute symmetrically to commitment, or whether leakage, adaptive weighting, or an impending threshold makes their influence unequal. Such designs probe the processing trajectory rather than only its endpoint. Usher and McClelland's work explicitly considered changes in the arrival of information, demonstrating the value of controlled evidence timing Usher & McClelland (2001) O'Connell et al. (2018).

Choice reversals and changes of mind introduce a further temporal distinction: the internal commitment criterion may be reached before the executed response is irreversible. Post-threshold processing, continued sampling, and motor delays can produce corrections after an incipient decision. A strict one-boundary-hit-and-stop model may be inadequate for such behavior without an additional stage. Consequently, a model of initial commitment and a model of final recorded response need not assign identical temporal intervals to evidence accumulation. Experiments using continuous response traces or instructed delayed reports can make that difference observable Gold & Shadlen (2007) O'Connell et al. (2018).

Temporally structured stimuli expose important boundary conditions for applying simple stationary models. If evidence reverses midway through a trial, fitting a constant drift may summarize average performance without representing its actual computation. One can instead specify a drift function driven by observed stimulus history or construct a rival model with leak and adaptive gain. Comparison should then use trial-level prediction conditional on the actual evidence stream, not just aggregate correct percentages. Generalizing sequential sampling beyond static laboratory displays depends on representing information history as part of the generative process Ratcliff et al. (2016) Wilson & Collins (2019).

Neural Evidence Accumulation and the Limits of Localization

The neural evidence for accumulation emerged prominently from motion-discrimination paradigms in which stimulus strength could be controlled while monkeys reported a direction judgment with an eye movement. Shadlen and Newsome described activity in lateral intraparietal neurons that varied with decision formation, linking the temporal course of neural firing to evolving evidence about the task's alternatives. Their findings provided a striking biological connection to sequential mathematical models because behavioral difficulty, choice, and response time could be analyzed alongside neural dynamics. The relationship remains an inference from population activity and task coupling rather than a direct reading of a unique mathematical decision variable Shadlen & Newsome (2001) Gold & Shadlen (2007).

Subsequent experiments extended these approaches to changes in the number of response options and to human electrophysiology. Churchland, Kiani, and Shadlen compared two- and four-alternative perceptual decisions, while O'Connell, Dockree, and Kelly identified a human electrophysiological signal in a gradual target-detection task that shared accumulation-to-bound features and could be dissociated from some sensory and motor signals. These findings support the physiological plausibility of temporally integrated decision signals. They do not by themselves determine whether integration is perfect, leaky, recurrently competitive, or implemented in multiple coupled populations Churchland et al. (2008) O'Connell et al. (2012).

Gold and Shadlen's synthesis and O'Connell and colleagues' later review illustrate the shift from fitting final decisions to constraining a mechanistic account with several measurement levels. A useful neural test should specify whether the measured activity predicts choice, elapsed time, evidence reliability, and trial-wise deviations beyond confounds related to motor preparation and attention. Neural signatures may covary with a fitted decision parameter without identifying it uniquely, especially when recording and filtering procedures create temporal integration-like shapes. Convergent perturbation, simultaneous behavior, and explicit observation models are required for increasingly specific claims Gold & Shadlen (2007) O'Connell et al. (2018).

Attention, Eye Movements, and Value-Guided Sampling

Value-based choice introduces an important form of endogenous evidence selection. When deciding between consumer goods or food items, observers repeatedly look between alternatives, and their gaze changes which information is available for comparison. Krajbich, Armel, and Rangel modeled binary food choices with an attentional drift-diffusion account in which fixation affects the evidence used in value comparison. Their experiment demonstrated quantitative relations among measured fixations, choices, and response times. This connects attention to the temporal mechanics of selection while retaining the need to distinguish attention's causal effect from value-driven gaze allocation Krajbich et al. (2010) Milosavljevic et al. (2010).

Visual fixation is observable, but the model's attentional weight remains a theoretical parameter. A person may inspect an item because it is preferred, because it is uncertain, or because an experimental instruction directs attention; each mechanism produces different correlations. To establish a contribution of gaze to accumulation, designs can manipulate display duration, fixation sequences, or informational accessibility while holding other value-relevant properties stable. The resulting choice and response-time changes should be predicted from independently estimated weights. Without such constraints, the gaze term can absorb unmodeled preference or presentation effects Krajbich et al. (2010) Krajbich & Rangel (2011).

The transition from binary to three-alternative value choice creates additional tests because attention now moves among more candidates while preference competition unfolds. Krajbich and Rangel reported that a multialternative drift-diffusion approach captured relationships among fixations, choices, and reaction times in a trinary task. The study does not establish a universal attentional decision mechanism across settings; it demonstrates a productive empirical integration of sampling trajectories and final behavioral outcomes in the tested domain. Models of choice dynamics should explicitly state whether attention selects observations, changes their gain, changes their reliability, or alters the stopping threshold Krajbich & Rangel (2011) Roe et al. (2001).

Memory Retrieval and Recognition Decisions

The modern diffusion tradition has a historical connection to memory recognition, where subjects decide whether a probe belongs to a studied set. Retrieval evidence is often noisy, and a response may reflect accumulating match or mismatch information until a criterion is crossed. Ratcliff's 1978 theory combined continuous random-walk comparisons with predictions for both recognition accuracy and latency distributions. The framework made it possible to test whether changes in memory quality and decision caution could be differentiated rather than explaining every slow response through an undifferentiated notion of memory strength Ratcliff (1978) Ratcliff & McKoon (2008).

Recognition experiments present difficulties that matter for psychological interpretation. Stimulus familiarity, studied-list length, encoding conditions, response bias, and item heterogeneity may alter the information process or the distribution of trial-level drifts. A diffusion fit can describe retrieval decisions accurately under a given observation model, but a fitted evidence-quality parameter is not itself a complete theory of how a memory trace is constructed or accessed. It must be connected to independent manipulations of learning and retrieval, and its relation to signal-detection sensitivity must be defined for the experimental protocol Ratcliff (1978) Ratcliff et al. (2016).

Response-signal paradigms strengthen the temporal analysis by constraining when a response must occur, making an accuracy-versus-processing-time function observable. Ratcliff's work on jointly modeling response-signal and standard response-time data shows how multiple task formats can constrain the same candidate process. When a model fits self-paced choices but fails under externally controlled response timing, the failure may identify incorrect assumptions about accumulation or nondecision stages. A cross-paradigm test therefore can be more informative than an especially accurate fit to one unconstrained reaction-time distribution Ratcliff (2006) Usher & McClelland (2001).

Individual Differences, Aging, and Practice

Sequential sampling parameters are often used to study individual differences in processing because they partition speeded performance into candidate components of evidence quality, caution, bias, and nondecision operations. Such partitioning can clarify why two individuals obtain similar accuracy with different latencies, or why slower responding need not imply lower sensory sensitivity. Yet reliable individual interpretation requires adequate within-person observations, stable measurement, and evidence that the same model and parameter scale apply across the groups being compared. Group differences in fitted parameters are conditional explanations, not direct measurements of immutable traits Ratcliff et al. (2016) Lerche et al. (2017).

Ratcliff, Thapar, and McKoon examined age and practice in masked letter and brightness discriminations. Their diffusion-model analysis related older adults' initial performance and improvement across practice sessions to more than one processing component, including changes in evidence quality and response criteria. The study illustrates how longitudinal manipulation may distinguish learning-related changes from stable speed differences within its sampled age groups and specific tasks. A scientific synthesis should retain those sampling boundaries and avoid converting one set of perceptual practice findings into a universal aging law Ratcliff et al. (2006) Ratcliff et al. (2016).

Individual differences research must separate within-person fluctuation from between-person variation. A hierarchical model that shrinks unstable estimates toward group distributions may improve precision, while potentially concealing meaningful subgroups or systematic model violations. Practice changes sensory encoding, strategy, and motor execution over time, undermining assumptions that all trials within a block are interchangeable. Appropriate designs track sessions, order, fatigue, and measurement invariance; recovery studies should reproduce realistic numbers of trials per participant. Findings about change gain interpretability when they survive alternative parameter constraints and replicated protocols Wiecki et al. (2013) Lerche et al. (2017).

Likelihood Estimation, Contaminants, and Numerical Fitting

Diffusion likelihoods assign probability to each observed response and decision time under a specified parameterization. Estimation may use maximum likelihood, chi-square comparisons of response-time bins, quantile methods, or approximate algorithms when the likelihood is computationally difficult. Ratcliff and Tuerlinckx examined parameter estimation under simulation with known generating values, comparing fitting methods and the effects of contaminant responses and parameter variability. Their work shows why numerical success should be judged by bias, spread, and recovery properties, rather than only the optimizer's reported convergence Ratcliff & Tuerlinckx (2002) Van Zandt (2000).

Finite-sample behavior depends strongly on task difficulty, balance between response categories, and the number of experimental conditions. Lerche, Voss, and Nagler used simulation to evaluate parameter recovery across a wide range of trial counts and fitting criteria. The scientific lesson is a design-dependent one: investigators should not select a universal minimum number of trials without specifying which parameters and individual-level effects must be recovered. An experiment may adequately estimate a mean drift difference while leaving drift variability or starting bias weakly identified. Planned sample size and trial count should follow simulation calibrated to the target inference Lerche et al. (2017) Ratcliff & Tuerlinckx (2002).

Good practice includes multiple optimizer starts, explicit convergence criteria, checks for implausible boundary solutions, and sensitivity to alternative treatment of fast guesses and lapses. A contaminated trial may be generated by a separate state of attention or motor readiness; simply deleting it according to a post hoc cutoff can distort both tails and the estimated nondecision distribution. Numerical fitting must record data exclusions, latency units, time origin, censoring, stimulus coding, and response mapping. These details determine the likelihood and make computational reproduction possible. A satisfactory loss value is the beginning of inference rather than its endpoint Van Zandt (2000) Wilson & Collins (2019).

Hierarchical Bayesian Models and Parameter Uncertainty

Hierarchical estimation represents participant parameters as draws from population distributions while fitting the joint pattern of observed choices and latencies. This can stabilize inference when individual datasets are sparse by pooling information across persons. Wiecki, Sofer, and Frank introduced HDDM as a computational approach to hierarchical Bayesian estimation of drift-diffusion parameters, facilitating uncertainty assessment and regression-style analysis. The methodological advantage is conditional: partial pooling changes estimates according to the group structure and prior assumptions, so the researcher must justify how participants and conditions are modeled Wiecki et al. (2013) Forstmann et al. (2016).

Posterior distributions make uncertainty visible, but they do not automatically repair incorrect generative assumptions. A narrow posterior may result from a restrictive prior, excessive pooling, or an insufficiently flexible noise model. Convergence diagnostics, effective sample information, posterior predictive simulations, and sensitivity analyses are therefore necessary to establish computational trustworthiness. The model should specify whether trial-level covariates affect drift, threshold, nondecision time, or multiple parameters, and whether those relations are estimated at the individual or population level. Interpretability depends on the stability of such choices under scientifically plausible alternatives Wiecki et al. (2013) Wilson & Collins (2019).

A practical challenge is distinguishing group effects from parameter tradeoffs. A condition difference that can be explained by drift or by response caution may have a broad posterior ridge even when fitted choice predictions look excellent. Hierarchical constraints sometimes obscure this nonidentifiability by lending apparent precision to weak individual information. Study design, informative stimulus variation, and simulation-based recovery are better solutions than unexamined computational sophistication. Posterior predictive checks should display outcome-conditioned quantiles and errors for every relevant task condition, rather than only pooled mean reaction times Lerche et al. (2017) Ratcliff & Tuerlinckx (2002).

Parameter Recovery, Model Recovery, and Mimicry

Parameter recovery tests the inferential chain under known generating conditions. A researcher samples plausible parameter vectors, simulates decisions and latencies from the proposed model, fits the estimation procedure, and compares recovered estimates with the generating quantities. Recovery should assess bias, variability, correlations, and systematic misidentification over the range relevant to the actual experiment. Repeating the procedure at the actual trial count reveals whether a psychological claim about drift, bias, or caution could have been distinguished from its alternatives. Successful recovery is a property of model, design, and estimator together Ratcliff & Tuerlinckx (2002) Lerche et al. (2017).

Model recovery addresses a different question: whether the proposed experiment and comparison procedure can distinguish among rival generative models. Simulate data under each candidate, fit every candidate to those data, and construct the resulting confusion pattern. A complicated model may fit datasets generated by simpler rivals because its parameter space contains nearly equivalent behavior. When recovery is poor, a model-selection score obtained on the observed data should not be interpreted as decisive mechanistic proof. The literature on computational falsification emphasizes testing what each model cannot reproduce, including qualitative effects defined before estimation Palminteri et al. (2017) O'Connell et al. (2018).

Cross-model mimicry is especially acute when diffusion, linear ballistic races, and leaky competitive dynamics are constrained by the same aggregate choice and reaction-time distributions. More discriminating experiments record stimulus history, gaze, intermediate actions, electrophysiology, or changes of policy across task conditions. These observations add dimensions against which rival processes must make simultaneous predictions. Model recovery also needs robustness checks for misspecified noise, omitted lapse processes, and trial dependencies. The strongest conclusion is sometimes a set of compatible accumulation mechanisms rather than a unique neural algorithm Usher & McClelland (2001) Brown & Heathcote (2008) O'Connell et al. (2018).

Recent Tests of Relative and Absolute Evidence

The family resemblance among accumulators has become an empirical research question in its own right. Kvam formulated diffusion and separate-accumulator accounts within a broader geometry of evidence and identified ways to distinguish relative from absolute stopping rules. His orientation-discrimination analyses considered the sampled information, manipulations separating stimulus–option match from interoption discriminability, classification of simulated datasets, and inference about the geometry of option representations. In those analyses, the four approaches converged in favor of accumulator-based representations. This result is evidence for the tested orientation task and the compared models, and it provides testable methods for additional domains rather than a universal replacement of diffusion modeling Kvam (2025).

The study's importance lies in its shift from comparing goodness-of-fit scores under familiar tasks to engineering observations with different predictions. In a relative-evidence representation, evidence for one alternative enters the same decision coordinate as evidence against another; in an absolute-evidence representation, options can maintain separately evolving support. Varying total evidence independently of the relative difference, where the stimulus permits it, places a discriminating constraint on these representations. Such experiments clarify what a model assumes about information, even when baseline accuracy and mean latency are nearly indistinguishable Kvam (2025) Brown & Heathcote (2008).

A recent source of methodological refinement is the explicit taxonomy of sequential-sampling processes by stopping rule, temporal granularity, and evidence representation. Smith and Ratcliff's 2025 treatment places continuous diffusion alongside discrete random walks and counter processes, making clear that the label sequential sampling names a broader computational family. This classification also helps prevent false generalization: two models can share a first-passage metaphor while differing in the sampled variable, dependence structure, and interpretation of a threshold. Progress depends on experiments that isolate those differences Smith & Ratcliff (2025) Forstmann et al. (2016).

Causal Tests, Perturbation, and Process Tracing

Correlational fit becomes stronger mechanistic evidence when the research design intervenes on a hypothesized component. If drift is claimed to reflect evidence strength, selectively degrading stimulus information should alter the decision distribution in the predicted way; if threshold represents caution, changing incentives or speed instructions should produce quantitatively separable effects. Voss and colleagues used such manipulations to validate diffusion-parameter interpretations, demonstrating the general logic of testing parameter meaning through selective interventions. A convincing intervention additionally examines whether the manipulated factor changes other subprocesses, because psychological tasks rarely isolate one mechanism perfectly Voss et al. (2004) Ratcliff et al. (2016).

Time-resolved process tracing creates a second line of evidence. Eye movements can establish which alternatives received attention; continuous motor trajectories can indicate early response commitments; neural recordings may reveal the timing of evolving task-related signals. These measurements can be built into the observation model instead of being used only as post hoc validation. When a candidate accumulation model predicts both latency distributions and a trial's fixation-dependent choice probability, it faces more exacting evidence than if its parameters are adjusted separately for each measure. However, recording a correlated signal establishes neither exclusive causation nor the unique neural locus of a decision Krajbich et al. (2010) O'Connell et al. (2012).

Intervention and tracing have complementary limitations. Changing a stimulus can alter its salience and the participant's response strategy; measuring attention can change the timing of responses; neural stimulation may affect several circuits. The design should therefore include control conditions, quantitative competing-model predictions, and tests of whether effect directions hold across laboratories or measurement techniques. Convergent evidence arises when controlled changes, measured process histories, and out-of-sample behavior agree under a shared generative account. Where they diverge, the discrepancy identifies a target for revising the mechanism Gold & Shadlen (2007) O'Connell et al. (2018).

Construct Validity, Measurement Reliability, and Transport

Fitted parameters are latent model quantities, and their scientific meaning depends on whether they retain a stable relation to the construct under investigation. A participant's estimated drift can change with the stimulus set, scoring convention, response mapping, or permissible decision strategy. Threshold estimates can change with incentives, block length, and feedback. A claim about individual differences therefore requires repeatability over the intended observation window and invariance across relevant task variants. Without those checks, a parameter is a useful description of performance within a fitted generative system but an uncertain general psychological measurement Lerche et al. (2017) Ratcliff et al. (2016).

Reliability should be assessed at the level of the actual inference: within-session recovery, between-session stability, condition differences, or rank-order consistency. These are distinct quantities. A model may accurately predict the group-average response-time distribution yet provide poor participant-level rank reliability for starting bias because each individual contributed few biased choices. Hierarchical estimation can reduce variance but may introduce dependence on its population assumptions. Empirically relevant simulations should preserve the group structure, trial counts, error rates, and parameter covariance of the target study Wiecki et al. (2013) Lerche et al. (2017).

Transport across domains deserves explicit investigation. A diffusion interpretation of threshold in a two-choice motion task need not map directly onto a threshold for choosing between foods, where value construction and gaze alter the sampled information. An apparent common parameter can support an abstract computational comparison while failing measurement invariance across populations or tasks. The proper claim should specify which mathematical roles are shared and which cognitive operations provide the inputs. Cross-domain analogies generate fruitful hypotheses when they preserve the distinction between model structure, construct validity, and biological realization Forstmann et al. (2016) Krajbich et al. (2010).

Designing Experiments That Discriminate Decision Models

An experiment tailored to competing sequential accounts begins with predictions, not a preference for a familiar estimator. The researcher identifies a behavioral pattern that the models generate differently, chooses stimulus and payoff manipulations that maximize the divergence, and simulates their likely observations under realistic sampling variability. Candidate discriminators include late evidence pulses, relative versus total support for options, choice-set composition, speed–accuracy instructions, and response deadlines. A preregistered analysis can then test these predictions before viewing the observed condition differences. The comparative program outlined by Kvam illustrates the value of selecting tests that target representational assumptions Kvam (2025) Palminteri et al. (2017).

Design adequacy should be judged through parameter recovery and model recovery at the planned number of participants and trials. Simulated datasets can reveal whether a true accumulator process is routinely misclassified as diffusion under the intended design, or whether estimates of drift variability trade off with contaminants. A model-comparison procedure should also be tested against generating processes outside its nominated candidate set, since all compared accounts can be misspecified. Held-out predictions are particularly valuable when the held-out conditions include altered evidence timing, reward structure, or response options Lerche et al. (2017) Ratcliff & Tuerlinckx (2002).

Reproducibility requires preserving stimulus-generation scripts, trial sequence, instructions, response devices, preprocessing decisions, simulator settings, priors, optimizer configuration, code versions, and model-selection procedures. Parameter values without definitions of the noise scale or starting-state conventions are difficult to compare across laboratories. Posterior predictive or simulation checks should report conditional outcome frequencies, latency quantiles, and diagnostic tails, along with failures to reproduce relevant qualitative patterns. A complete record enables independent investigators to distinguish an inference supported by the design from a convenient fit enabled by flexible implementation choices Wilson & Collins (2019) Van Zandt (2000).

Applications, Generalization, and Explanatory Boundaries

Sequential sampling provides a common language for several laboratory phenomena: discriminating perceptual signals, recognizing studied items, comparing values, selecting responses under time constraints, and revising preference with additional information. Their common feature is temporally extended commitment under uncertainty. The evidence entering the decision process, however, differs substantially among domains. The sampled quantity may arise from externally controlled sensory measurements, retrieved memory evidence, fixated attributes, or learned value estimates. A single abstract boundary-crossing architecture therefore contains several candidate psychological theories, each requiring independent domain-specific validation Ratcliff (1978) Gold & Shadlen (2007) Krajbich & Rangel (2011).

The framework also has explanatory boundaries. Some choices unfold through well-practiced action routines, symbolic deliberation, or collective negotiation in which evidence and stopping policies are continually reorganized. A simple diffusion fit may describe their response distributions without reproducing their internal computations. Preference change during choice provides a particularly concrete test. Lee and Pezzulo's 2026 modeling analysis examined choice-induced shifts in value ratings and found that variants of sequential sampling with dynamically constructed values could generate several previously reported relations among choice consistency, latency, and post-choice preference. These results extend the explanatory repertoire of value-based accumulation, while the paper's model demonstrations require independent empirical discrimination from rival mechanisms Lee & Pezzulo (2026) Busemeyer & Townsend (1993).

Applications are strongest when models predict outcomes under new conditions. An account fitted under stable information should predict changed sensory reliability, a novel reward schedule, or a new option set after parameters are constrained by independent observations. The resulting success warrants a statement about generalization for those manipulations, while a persistent failure helps identify which assumptions require revision. This approach keeps the architecture connected to experimentally measurable processes and prevents the language of drift, caution, or evidence from becoming an unfalsifiable paraphrase of any observed selection Palminteri et al. (2017) Wilson & Collins (2019).

Artificial Decision Systems and Cross-System Analysis

Sequential choice mathematics can also describe artificial decision systems that collect uncertain information, update scores, and stop when a threshold or computational budget is reached. In this context the observable process can often be instrumented directly: system logs may expose evidence increments, policy scores, branch selection, and explicit termination rules. The resulting measurements offer opportunities to evaluate formal relations between sampling time, informational quality, and chosen actions. Such applications belong to the Artificial Era's comparative research horizon, where human and artificial processes are evaluated with a shared vocabulary of testable operations Bogacz et al. (2006) Wilson & Collins (2019).

Cross-system use requires a precise translation between mathematical and computational levels. A human diffusion-model drift is estimated from behavioral distributions under psychological assumptions; a software system's decision score may be available directly from its implementation. A numerical resemblance between these variables establishes no equivalence of memory, subjective valuation, experience, or decision architecture. Investigators should define the sampled information, its stochastic source, the stopping condition, and the observable response separately for each system. Comparable experiments can then test functional correspondences without assuming the same implementation or the same psychological ontology Forstmann et al. (2016) O'Connell et al. (2018).

Artificial decision processes may be especially useful for methodological calibration because the generating algorithm can be known by construction. Researchers can design benchmark systems with distinct absolute and relative stopping rules, simulate choices under controlled perturbations, and quantify whether a diagnostic procedure recovers the generating mechanism. Such computational tests validate a method within the tested family but do not establish that human decisions are generated by the same code. Human–Artificial comparative research advances when it distinguishes the methodological role of engineered ground truth from empirical claims about biological cognition Kvam (2025) Brown & Heathcote (2008).

Open Scientific Questions and Research Priorities

A major unresolved problem is mechanistic identifiability among models that produce similar choice–latency distributions. Accumulators may differ in their representation of absolute versus relative support, continuous versus trial-wise noise, leakage, competition, and time-dependent boundaries; ordinary speeded tasks often insufficiently constrain these distinctions. Research priorities include factorial manipulations of evidence geometry, informative late-sample conditions, multialternative settings, and joint behavioral–physiological prediction. The discriminating tests proposed by Kvam provide a concrete example of how model geometry can be turned into observable competing predictions Kvam (2025) Usher & McClelland (2001).

A second problem concerns the stationarity of decision processes. Trial history, learned reward values, fatigue, adaptation, gaze, and response preparation can change during a block or even within a single decision. Most tractable standard models summarize such changes through fixed parameters or limited variability terms. New investigations need designs that separately measure changes in the evidence stream and changes in the decision policy, with sufficient time resolution to identify when stopping occurs. The challenge is to retain explicit computational predictions while capturing the evolving contexts of choice Hawkins et al. (2015) Krajbich et al. (2010).

A third priority is measurement and generalization. Parameter labels remain attractive because they promise interpretable components of mental performance, but cross-task reliability, recoverability in realistic samples, and prediction under interventions require more systematic testing. Independent replications should publish complete distributions and trial-wise data where ethical and legally permissible, including unsuccessful model fits. Comparative evidence should be accumulated across domains and populations rather than declaring one model universally correct on the basis of a single benchmark. This agenda connects methodological transparency with substantive theories of psychological organization Lerche et al. (2017) Palminteri et al. (2017).

Conclusion

Sequential sampling models provide a rigorous account of choice as an evolving and terminating process. Their central scientific contribution is to connect selected alternatives with the times at which commitments occur, allowing a joint theory of accuracy, errors, speed, and response-time distributions. Historical sequential-testing ideas developed into diffusion processes, independent races, leaky competition, and dynamic preference models, each with a different account of how evidence is represented and how action becomes obligatory. The strongest findings establish that temporally structured evidence and response policy shape observable decisions in many laboratory settings, while the exact generative mechanism remains a question for targeted comparison Stone (1960) Ratcliff & McKoon (2008) Brown & Heathcote (2008).

The evidence also establishes conditions for responsible interpretation. A fitted drift parameter does not automatically identify a particular sensory or cognitive mechanism; an estimated threshold requires experimental evidence for its relation to caution; a neural accumulation-like signal does not uniquely identify a stochastic differential equation. Robust explanation depends on selective intervention, trial-level or distributional fit, parameter and model recovery, and independent measurements where available. Recent efforts to distinguish absolute from relative evidence show how methodological questions can become direct empirical tests Voss et al. (2004) Kvam (2025) O'Connell et al. (2018).

The field's long-term research trajectory is toward stronger integration of mathematical specifications, controlled temporal experiments, computational validation, and mechanistic measurements. In comparative psychological science, including Artificial Era research, sequential sampling supplies a precise framework for posing questions about evidence, stopping, and selected action. Its explanatory scope expands when each proposed correspondence is operationalized, experimentally discriminated, and reproduced across appropriate conditions. The resulting synthesis preserves the central achievement of mathematical psychology: scientific insight grows when a psychological claim implies an inspectable distribution of possible observations Forstmann et al. (2016) Wilson & Collins (2019).

References

  • Bogacz, R., Brown, E., Moehlis, J., Holmes, P., & Cohen, J. D. (2006). The physics of optimal decision making: A formal analysis of models of performance in two-alternative forced-choice tasks. Psychological Review, 113(4), 700–765. https://doi.org/10.1037/0033-295X.113.4.700
  • Brown, S. D., & Heathcote, A. (2008). The simplest complete model of choice response time: Linear ballistic accumulation. Cognitive Psychology, 57(3), 153–178. https://doi.org/10.1016/j.cogpsych.2007.12.002
  • Busemeyer, J. R., & Townsend, J. T. (1993). Decision field theory: A dynamic-cognitive approach to decision making in an uncertain environment. Psychological Review, 100(3), 432–459. https://doi.org/10.1037/0033-295X.100.3.432
  • Churchland, A. K., Kiani, R., & Shadlen, M. N. (2008). Decision-making with multiple alternatives. Nature Neuroscience, 11, 693–702. https://doi.org/10.1038/nn.2123
  • Forstmann, B. U., Ratcliff, R., & Wagenmakers, E.-J. (2016). Sequential sampling models in cognitive neuroscience: Advantages, applications, and extensions. Annual Review of Psychology, 67, 641–666. https://doi.org/10.1146/annurev-psych-122414-033645
  • Gold, J. I., & Shadlen, M. N. (2007). The neural basis of decision making. Annual Review of Neuroscience, 30, 535–574. https://doi.org/10.1146/annurev.neuro.29.051605.113038
  • Hawkins, G. E., Forstmann, B. U., Wagenmakers, E.-J., Ratcliff, R., & Brown, S. D. (2015). Revisiting the evidence for collapsing boundaries and urgency signals in perceptual decision-making. Journal of Neuroscience, 35(6), 2476–2484. https://doi.org/10.1523/JNEUROSCI.2410-14.2015
  • Krajbich, I., Armel, C., & Rangel, A. (2010). Visual fixations and the computation and comparison of value in simple choice. Nature Neuroscience, 13(10), 1292–1298. https://doi.org/10.1038/nn.2635
  • Krajbich, I., & Rangel, A. (2011). Multialternative drift-diffusion model predicts the relationship between visual fixations and choice in value-based decisions. Proceedings of the National Academy of Sciences, 108(33), 13852–13857. https://doi.org/10.1073/pnas.1101328108
  • Kvam, P. D. (2025). The Tweedledum and Tweedledee of dynamic decisions: Discriminating between diffusion decision and accumulator models. Psychonomic Bulletin & Review, 32(2), 588–613. https://doi.org/10.3758/s13423-024-02587-0
  • Lee, D. G., & Pezzulo, G. (2026). Choice-induced preference change under a sequential sampling model framework. Scientific Reports, 16, Article 14455. https://doi.org/10.1038/s41598-026-44610-5
  • Lerche, V., Voss, A., & Nagler, M. (2017). How many trials are required for parameter estimation in diffusion modeling? A comparison of different optimization criteria. Behavior Research Methods, 49(2), 513–537. https://doi.org/10.3758/s13428-016-0740-2
  • Milosavljevic, M., Malmaud, J., Huth, A., Koch, C., & Rangel, A. (2010). The drift diffusion model can account for the accuracy and reaction time of value-based choices under high and low time pressure. Judgment and Decision Making, 5(6), 437–449. https://doi.org/10.1017/S1930297500001285
  • O'Connell, R. G., Dockree, P. M., & Kelly, S. P. (2012). A supramodal accumulation-to-bound signal that determines perceptual decisions in humans. Nature Neuroscience, 15, 1729–1735. https://doi.org/10.1038/nn.3248
  • O'Connell, R. G., Shadlen, M. N., Wong-Lin, K., & Kelly, S. P. (2018). Bridging neural and computational viewpoints on perceptual decision-making. Trends in Neurosciences, 41(11), 838–852. https://doi.org/10.1016/j.tins.2018.06.005
  • Palminteri, S., Wyart, V., & Koechlin, E. (2017). The importance of falsification in computational cognitive modeling. Trends in Cognitive Sciences, 21(6), 425–433. https://doi.org/10.1016/j.tics.2017.03.011
  • Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review, 85(2), 59–108. https://doi.org/10.1037/0033-295X.85.2.59
  • Ratcliff, R. (2006). Modeling response signal and response time data. Cognitive Psychology, 53(3), 195–237. https://doi.org/10.1016/j.cogpsych.2005.10.002
  • Ratcliff, R., & McKoon, G. (2008). The diffusion decision model: Theory and data for two-choice decision tasks. Neural Computation, 20(4), 873–922. https://doi.org/10.1162/neco.2008.12-06-420
  • Ratcliff, R., Smith, P. L., Brown, S. D., & McKoon, G. (2016). Diffusion decision model: Current issues and history. Trends in Cognitive Sciences, 20(4), 260–281. https://doi.org/10.1016/j.tics.2016.01.007
  • Ratcliff, R., Thapar, A., & McKoon, G. (2006). Aging, practice, and perceptual tasks: A diffusion model analysis. Psychology and Aging, 21(2), 353–371. https://doi.org/10.1037/0882-7974.21.2.353
  • Ratcliff, R., & Tuerlinckx, F. (2002). Estimating parameters of the diffusion model: Approaches to dealing with contaminant reaction times and parameter variability. Psychonomic Bulletin & Review, 9(3), 438–481. https://doi.org/10.3758/BF03196302
  • Roe, R. M., Busemeyer, J. R., & Townsend, J. T. (2001). Multialternative decision field theory: A dynamic connectionist model of decision making. Psychological Review, 108(2), 370–392. https://doi.org/10.1037/0033-295X.108.2.370
  • Shadlen, M. N., & Newsome, W. T. (2001). Neural basis of a perceptual decision in the parietal cortex (area LIP) of the rhesus monkey. Journal of Neurophysiology, 86(4), 1916–1936. https://doi.org/10.1152/jn.2001.86.4.1916
  • Smith, P. L., & Ratcliff, R. (2025). Sequential-sampling models of decision making. In Diffusion Process Models of Decision Making (pp. 24–55). Cambridge University Press. https://doi.org/10.1017/9781009652667.004
  • Stone, M. (1960). Models for choice-reaction time. Psychometrika, 25(3), 251–260. https://doi.org/10.1007/BF02289729
  • Usher, M., & McClelland, J. L. (2001). The time course of perceptual choice: The leaky, competing accumulator model. Psychological Review, 108(3), 550–592. https://doi.org/10.1037/0033-295X.108.3.550
  • Van Zandt, T. (2000). How to fit a response time distribution. Psychonomic Bulletin & Review, 7(3), 424–465. https://doi.org/10.3758/BF03214357
  • Voss, A., Rothermund, K., & Voss, J. (2004). Interpreting the parameters of the diffusion model: An empirical validation. Memory & Cognition, 32(7), 1206–1220. https://doi.org/10.3758/BF03196893
  • Wiecki, T. V., Sofer, I., & Frank, M. J. (2013). HDDM: Hierarchical Bayesian estimation of the drift-diffusion model in Python. Frontiers in Neuroinformatics, 7, Article 14. https://doi.org/10.3389/fninf.2013.00014
  • Wilson, R. C., & Collins, A. G. E. (2019). Ten simple rules for the computational modeling of behavioral data. eLife, 8, e49547. https://doi.org/10.7554/eLife.49547