RESEARCH DOSSIER
Perceptual Decisions and Evidence Accumulation
Abstract. Perceptual decisions convert noisy sensory information into categorical choices and action commitments through computations distributed over time. This fundamental critical narrative research monograph examines psychophysical and neural evidence for accumulation-to-bound processes, traces historical experiments on motion and vibrotactile discrimination, compares diffusion, race, and leaky-competition models, and evaluates neural recordings, human electrophysiology, response-time distributions, confidence, and changes of mind. Its analysis separates behavioral evidence for integration from claims about unique neural implementation, assesses identifiability and causal inference, and identifies priorities for crossmodal generalization and human–artificial perceptual partnerships. The manuscript synthesizes published scholarship; it reports no original experiment or registered systematic review.
Keywords: perceptual decision-making; evidence accumulation; sequential sampling; drift-diffusion model; neural decision variable; random-dot motion; vibrotactile discrimination; reaction time; speed–accuracy tradeoff; confidence; postdecision evidence; changes of mind; lateral intraparietal cortex; human electrophysiology
The Scientific Problem of Perceptual Commitment
Perceptual decision-making concerns the transformation of uncertain sensory information into a categorical judgment, an action, and an estimate of the judgment's reliability. An observer inspecting a noisy visual field may decide which way dots move, whether a sound occurred, or which of two vibrations had the higher frequency. These tasks reveal the relationship among evidence quality, observation time, choice, and response latency. A decision is experimentally identified by an observable commitment, while the latent computation must be inferred through behavioral distributions, controlled stimulus perturbations, and neural measurements. The research program connects psychophysics to theories of sequential inference and action preparation Gold & Shadlen (2007) Palmer et al. (2005).
The principal scientific questions concern how successive samples are weighted, how much evidence is required before reporting a choice, how prior probabilities and rewards influence a decision policy, and how processing continues when the initial commitment is revised. A theory must specify the evidence variable, its temporal evolution, the stopping condition, and the mapping to an observed response. Alternative models can generate similar averages but predict different distributions of correct and incorrect response times. Accordingly, the relevant observations include trial-level accuracy, full reaction-time distributions, confidence reports, stimulus-aligned choice kernels, and, when available, simultaneously recorded neural dynamics Ratcliff & McKoon (2008) Forstmann et al. (2016).
This monograph critically integrates experimental psychology, computational modeling, and systems neuroscience. Its evidence is drawn from original primate and human perceptual tasks, formal model analyses, and focused methodological reviews. The synthesis treats an accumulator as a testable family of mechanisms rather than a literal anatomical container: evidence is a task-dependent representation, and a neural ramp is informative only when its relation to stimulus, movement, and choice is established. Neighboring subjects such as static signal detection, sensory transduction, motion perception, and general value-based choice have their own research questions. The distinctive subject here is the time-resolved formation and possible revision of perceptual commitments Ditterich (2006) Nienborg & Cumming (2009).
Evidence, Choice, and Time as Operational Constructs
Sensory evidence is information carried by a stimulus or neural response that favors one hypothesis over alternatives under a specified generative model. In a two-direction motion task, a momentary sensory sample can be described as favoring leftward or rightward motion, yet its reliability depends on stimulus coherence, sensory tuning, and noise correlations. A perceptual choice is typically a button press, saccade, reach, or verbal response, each imposing distinct motor and timing demands. Reaction time comprises perceptual encoding, decision formation, and response execution; separating them conceptually prevents a late action from being confused with weak sensory evidence Britten et al. (1992) Roitman & Shadlen (2002).
Behavioral evidence is most informative when difficulty changes predictably and the response mapping is held stable. Psychometric functions relate stimulus strength to the proportion of responses, whereas chronometric functions connect the same manipulation to decision latency. Palmer, Huk, and Shadlen showed that motion strength systematically changes both measures and that joint analysis constrains accounts of processing that accuracy alone leaves indeterminate Palmer et al. (2005). Response-terminated tasks allow observers to sample until commitment, whereas interrogation or response-signal tasks require a decision at an externally imposed time. These paradigms identify different properties of an accumulation hypothesis and should be compared with explicit assumptions about urgency and deadline effects.
Confidence is a further behavioral variable: an estimate, rating, wager, or willingness to decline a risky forced choice. Accuracy and confidence are related but do not coincide, because confidence depends on the information and criteria available when the estimate is elicited. In experiments with an opt-out choice, monkeys can reveal decision uncertainty behaviorally without a verbal probability report Kiani & Shadlen (2009). A model that describes speed and correctness may consequently require an additional account of post-decision information, confidence calibration, and the interval between response initiation and final report. The same observed accuracy can accompany distinct uncertainty states and decision policies Pleskac & Busemeyer (2010) van den Berg et al. (2016b).
Historical Foundations: Psychophysics to Neural Decisions
The contemporary perceptual-decision program inherits psychophysics' insistence on controlled stimuli and observable reports, together with mathematical analyses of discrimination under uncertainty. Signal detection theory supplied a formal account of how overlapping sensory distributions and response criteria jointly shape detection, though its familiar static form does not determine how evidence is processed over time. Sequential sampling models supplied that missing dimension by treating a decision as the result of repeated noisy observations. Historical priority belongs to several traditions in mathematical psychology, statistics, and neurophysiology; a modern diffusion model is therefore a refined explanatory framework, not an isolated discovery Ratcliff et al. (2016) Gold & Shadlen (2007).
The physiological research program advanced when investigators could record motion-sensitive neurons and behavioral discrimination in the same trained macaques. Newsome, Britten, and Movshon compared neuronal and perceptual performance during a matched visual discrimination task, establishing an important operational bridge between sensory responses and subjective choice Newsome et al. (1989). Britten and colleagues subsequently reported that the sensitivity of many middle-temporal neurons approached the animals' psychophysical sensitivity under tightly matched stimulus conditions Britten et al. (1992). These findings supported the possibility that limited and correlated sensory populations carry consequential decision information, but they did not by themselves establish the causal route from sensory firing to an eventual report.
Work on lateral intraparietal cortex extended the analysis from sensory representation to developing choice-related activity. Shadlen and Newsome documented responses associated with perceptual judgments, and Roitman and Shadlen connected neural dynamics to reaction times as animals discriminated motion Shadlen & Newsome (2001) Roitman & Shadlen (2002). Later work used temporal stimulus perturbations to challenge more specific models of integration Huk & Shadlen (2005). The sequence illustrates a methodological progression: matched psychophysical performance, response-time-dependent neural activity, and experimentally induced disturbances in sensory evidence. Each increases inferential specificity, although causal interpretation continues to depend on the architecture of the task and the circuits contributing to report.
Probabilistic Interpretation and Sequential Evidence
Suppose two hypotheses predict distinguishable distributions of momentary sensory observations. The log likelihood ratio for a sample measures how strongly that observation favors one hypothesis. Under conditional independence, successive log likelihood ratios can be summed, so accumulated evidence takes the form of a running balance between alternatives. The formal relation clarifies why accumulation improves discrimination when samples contain useful information and why repeated noisy observations produce variability in commitment time. Real neural inputs need not satisfy independence or stationarity, and conclusions about optimality therefore depend on those assumptions being justified in a particular experiment Bogacz et al. (2006) Drugowitsch et al. (2012).
Prior probability and asymmetric reward can alter the starting balance, decision criteria, or time-dependent policy. A prior favoring one motion direction is analytically distinct from a stimulus that carries more reliable evidence for that direction. Likewise, an urgency signal can accelerate commitment without changing sensory discriminability. Experiments that vary both stimulus reliability and incentives are needed to separate these effects. Models often admit multiple ways to reproduce an observed choice bias; interpreting fitted starting-point shifts as actual neural prior representations requires independent evidence. The Bayesian account is an exact normative specification under its assumptions and a hypothesis about psychological implementation when those assumptions are mapped onto neural activity Bogacz et al. (2006) Ratcliff & McKoon (2008).
The sequential probability ratio test provides a benchmark for deciding when additional information is worth its cost in a specified two-alternative setting. Evidence accumulation with boundaries can approximate or realize efficient choices when observations and costs satisfy the required conditions. Drugowitsch and colleagues demonstrated that a cost of continued sampling changes the theoretically desirable stopping policy, illustrating why a fixed threshold is not universally optimal Drugowitsch et al. (2012). Such analyses distinguish normative decision policies, formal cognitive algorithms, and observed neural trajectories. An empirical report of improved performance with more viewing time supports temporal information use; demonstrating that a particular likelihood-ratio algorithm implements that use requires much stronger model and perturbation evidence.
Diffusion Models and the Architecture of Reaction Time
The drift-diffusion model represents a decision variable evolving under systematic evidence and stochastic fluctuations until it reaches one of two bounds. Its main quantities typically include drift rate, boundary separation, starting position, and nondecision time. Drift expresses the average direction and quality of accumulated evidence, boundary separation characterizes the amount required for commitment, and nondecision time summarizes processes outside accumulation. These parameters are identified through the joint shapes of correct and error reaction-time distributions and choice proportions, rather than by a single mean latency. Ratcliff and McKoon provided a detailed account of this methodology in two-choice tasks Ratcliff & McKoon (2008).
Diffusion models are attractive because different experimental manipulations have distinguishable predictions under controlled conditions. Lower coherence generally reduces the informativeness of samples, while speed instructions may reduce response caution or alter urgency. Stimulus prevalence can bias the starting state or evidence interpretation. In the standard model, unbounded Gaussian noise is an approximation to an underlying sampling process, and estimates can depend on variability assumptions, contaminant responses, and participant-level heterogeneity. A good fit does not uniquely demonstrate continuous neural integration; analytically different processes may yield similar behavioral distributions Ratcliff et al. (2016) Forstmann et al. (2016).
Model adequacy should be assessed by posterior predictive or simulation-based recovery of complete distributional features: leading edges, tails, conditional accuracy by latency, and error times across difficulty levels. Fitting only mean reaction time and overall percent correct invites parameter trade-offs and weak identifiability. The possibility of across-trial drift and starting-state variability is especially consequential when one compares fast and slow errors. Formal analyses show that several commonly proposed two-choice architectures reduce to similar diffusion dynamics under limiting assumptions Bogacz et al. (2006). This equivalence is useful for identifying robust computations while warning against overinterpretation of a specific fitted algorithm as a uniquely established neural mechanism.
Race, Leaky Competition, and Alternative Accumulators
Accumulator theories differ in whether evidence feeds a single balance between alternatives or separate competing channels. A race model permits individual option signals to grow toward their own stopping criteria, whereas a leaky competing accumulator lets activation decay and makes alternatives inhibit one another. Usher and McClelland showed how leakage, competition, and nonlinear activation can describe perceptual-choice dynamics across different experimental paradigms Usher & McClelland (2001). Their model establishes a functional research family, not proof that one physiological circuit literally implements every equation. Identification requires manipulations that discriminate competition, leakage, urgency, shared noise, and response-stage effects.
The number of alternatives matters. A binary motion judgment can be approximated by a signed decision variable; a four-way choice requires representing several possible directions or differences among them. Churchland, Kiani, and Shadlen examined two- and four-choice tasks in macaques and reported LIP dynamics consistent with accumulation, including a different starting activity but comparable terminating levels Churchland et al. (2008). Multi-alternative models therefore need explicit assumptions about normalization, competition, and stopping conditions. Generalizing a well-fitting two-choice model to multiple choices by adding independent race units risks overlooking correlations, changing reward expectations, and the effects of the possible response set.
Comparison among models should focus on discriminating predictions instead of the sophistication of their mathematical notation. Diffusion, race, and competing-accumulator formulations can converge under parameter restrictions and diverge under time-varying inputs, multiple alternatives, or selective perturbations Bogacz et al. (2006) Ditterich (2006). A stronger evaluation uses parameter-recovery tests, held-out predictions across speed instructions, and measurements of how evidence arriving at different moments influences choice. When a model explains both behavior and neural signals, the correspondence must be evaluated against competing mapping functions, because an observed neural ramp may reflect motor preparation, urgency, or feedback in addition to sampled information.
Random-Dot Motion as an Experimental Laboratory
The random-dot motion paradigm offers a controlled distribution of stimulus strengths, usually varied through motion coherence. On each trial, an observer reports the global direction of a display containing noisy motion information. The experimenter can manipulate coherence, duration, direction priors, and rewards while preserving the response format. This feature allowed investigators to compare behavioral thresholds with motion-sensitive neurons and later to examine neural populations implicated in decision formation Newsome et al. (1989) Britten et al. (1992). Coherence is a task-specific experimental control rather than a universal unit of sensory uncertainty, and exact signal properties depend on dot-generation algorithms and display calibration.
In response-terminated variants, observers look at the moving dots until they are prepared to commit. Motion strength generally predicts both accuracy and decision time; easy trials often terminate more rapidly and with fewer errors Palmer et al. (2005). But a psychometric curve does not by itself specify the stopping rule. Chronometric analysis must account for anticipatory responses, very long lapses, motor delays, and whether zero-coherence trials effectively impose guessing. If an experimental response is a saccade rather than a keypress, its motor mapping may change neural activity near commitment. The causal object is the sequence of internal comparisons under a stated task, not any fixed behavioral gesture.
The motion task also supports temporal weighting tests: brief evidence pulses, shifts in coherence, and stochastic fluctuations can be embedded within a trial. Huk and Shadlen showed that short motion pulses affected subsequent LIP activity and behavior, providing a more direct test of accumulated memory for sensory information Huk & Shadlen (2005). Their evidence supported substantial temporal integration under their training and recording conditions. Comparisons among laboratories must nevertheless establish whether stimulus fluctuations, sensory adaptation, attention, and choice deadlines are genuinely comparable. Otherwise apparent disagreement about integration quality may arise from different evidence streams and response regimes rather than contradictory mechanisms.
Vibrotactile Comparisons and Memory-Guided Decisions
Perceptual decisions need not concern vision or continuously visible motion. In the classic vibrotactile comparison paradigm, trained monkeys feel two mechanical vibrations in sequence and decide which has the higher frequency. This task introduces a delay during which information about the first vibration must be retained, and a comparison stage when the second appears. Hernández, Zainos, and Romo recorded responses in medial premotor cortex that tracked stages of this process Hernández et al. (2002). The paradigm makes it possible to investigate how stored sensory information affects a later decision even when the two relevant signals never coincide in physical time.
Ventral premotor recordings further revealed activity associated with remembered sensory inputs, comparison, and forthcoming motor reports during sequential frequency discrimination Romo et al. (2004). These results are important because they show that a single broad cortical area may contain signals associated with several transformations in the perception-to-action chain. Attributing a neuron to a pure accumulator solely because firing changes through a trial would therefore be inadequate. The causal interpretation depends on stimulus identity, delay length, response selection, reward, and timing of neuronal selectivity. Delay-period activity may reflect working memory, prospective motor coding, or a mixture of functions.
Crossmodal generalization from random-dot motion to vibration requires separating an abstract computation from the representation it operates on. Both paradigms include evidence, comparison, uncertainty, and commitment, but their temporal structures differ substantially. In a simultaneous motion task, a stream can be sampled continuously; in delayed frequency comparison, a stored estimate and current sample must be combined. Evidence accumulation may occur within one or both intervals, during comparison, or at a subsequent action stage. The strongest crossmodal claim is consequently an experimentally constrained computational analogy, refined by recordings and perturbations, rather than the assumption that identical brain regions or exact parameters are shared O'Connell et al. (2012) Romo et al. (2004).
Sensory Neurons, Choice Probability, and Causal Direction
One of the field's central observations is that neural responses can predict behavioral choice even when external stimuli are held constant. This association is frequently quantified through choice probability or related measures of trial-level covariation. In early motion-discrimination experiments, the similarity between neuronal and psychophysical sensitivity suggested that sensory neuron populations could provide relevant information for behavior Britten et al. (1992). Yet a correlation between a sensory cell and choice can arise when sensory noise influences the decision, when decision-related signals feed back to sensory cortex, or when a common latent state modulates both recordings and choice.
Nienborg and Cumming investigated this interpretive problem using binocular disparity tasks, white-noise analysis, and recordings from visual cortical area V2. They showed a dissociation between patterns of stimulus influence and evolving choice-related neural activity, challenging a simple feedforward causal interpretation of choice correlations Nienborg & Cumming (2009). The study demonstrates why stimulus-to-choice kernels and neuron-to-choice association should be measured jointly. A neuron with strong choice probability need not carry proportionately strong causal weight in the decision computation. Conversely a neuron with weak individual association may contribute within a larger correlated population that a single-electrode experiment only partially observes.
Causal evidence can be strengthened through selective microstimulation, inactivation, transient stimulus pulses, or interventions on the time course of a candidate representation. Such interventions also require caution because they can change attention, stimulus encoding, motor planning, or arousal in addition to the targeted computation. Functional attribution is strongest when a manipulation produces the predicted behavioral change and the predicted intermediate neural dynamics with appropriate controls. Linking independent psychophysical and neural effects helps discriminate mechanisms; treating the combined evidence as a complete map of causal circuit flow is unwarranted without appropriate circuit-level measurements Huk & Shadlen (2005) Nienborg & Cumming (2009).
Parietal Decision Signals and Temporal Integration
Lateral intraparietal cortex became a focal research site because its activity is related to saccade planning and can also track evidence associated with the direction an animal will report. Shadlen and Newsome characterized relationships between LIP neural responses and perceptual decisions, while Roitman and Shadlen examined activity during reaction-time motion judgments Shadlen & Newsome (2001) Roitman & Shadlen (2002). The empirical significance lies in a signal whose evolving value covaries with sensory strength and choice, not in the name of a cortical area alone. The experiment's mapping between motion direction and eye-movement target is part of the mechanism under study.
Temporal integration predicts that a brief favorable sensory sample can influence subsequent neural activity even after the sample has disappeared. Huk and Shadlen applied evidence pulses and observed persistent behavioral and parietal effects, supporting a substantial integral of motion input in the tested task Huk & Shadlen (2005). This is more informative than averaging firing rates from many trials, because transient perturbations identify how information enters and persists in an ongoing process. Yet the inferred integration time scale can depend on the trial epoch and on which neural subpopulation is recorded, and integration observed in a signal does not establish that its recorded site alone performs the integration.
A contemporary circuit account must consider distributed interactions among sensory, parietal, frontal, and motor structures. A measured parietal trajectory may be a readout of a network state, a participant in recurrence, or an action-oriented representation downstream of partially separate inference. The same time-dependent buildup can satisfy a diffusion-model description without uniquely deciding among these anatomical alternatives. The explanatory standard therefore combines selective perturbation, sensory decoders, population covariance, and action-stage controls. This measured approach preserves the strong evidence for accumulation while treating the neural location and causal implementation as separately testable questions Gold & Shadlen (2007) O'Connell et al. (2012).
Human Electrophysiology and Supramodal Decisions
Human electroencephalography permits measurements of decision-related activity at millisecond resolution while observers perform perceptual tasks. O'Connell, Dockree, and Kelly reported a centroparietal positive potential with accumulation-like dynamics across sensory conditions, identifying a candidate signal that could be distinguished experimentally from sensory encoding and motor preparation O'Connell et al. (2012). The observation extends the empirical bridge beyond invasive primate recordings. It does not imply that the scalp potential corresponds to one precisely localized neuronal accumulator, because many spatially distributed sources contribute to a scalp waveform and inverse source localization remains underdetermined without additional anatomical constraints.
Kelly and O'Connell examined how stimulus strength and endogenous fluctuations influence the rate of such decision signals during motion discrimination Kelly & O'Connell (2013). Their work connected centroparietal buildup to external coherence and preparatory activity related to attentional state. This finding supports the view that difficulty and internal neural state shape the dynamics of decision formation. Nevertheless prestimulus attention, ongoing arousal, and expectation are correlated processes, and an EEG association does not identify a uniquely causal parameter. A stronger analysis aligns the signal to both stimulus onset and response time, controls motor response channels, and compares trialwise dynamics within difficulty conditions.
An important test of supramodal claims is whether the same neural signature predicts decisions when sensory modality and action requirements change. O'Connell and colleagues provided evidence for such generality within their task design O'Connell et al. (2012). The computational conclusion is that a decision-related signal can be expressed at a level of organization more abstract than particular receptors or movement muscles. Testing that conclusion in more complex choices requires substantial additional work: confidence reports, multiple alternatives, changing rewards, and continuous actions might recruit related but distinct population dynamics. Cross-paradigm convergence supports a family of transformations rather than a single universally identifiable brain signal.
Speed–Accuracy Policies and Adjustable Commitment
Perceptual systems operate under temporal constraints. A person can often increase the likelihood of a correct discrimination by taking longer to observe a weak signal, but waiting has opportunity costs and can delay consequential action. The speed–accuracy tradeoff is therefore an experimentally manipulable policy, not simply a flaw of human performance. In drift-diffusion terminology, a wider effective boundary often corresponds to greater caution; however urgency and changes in the evidence-independent component of activity can produce similar behavioral consequences Ratcliff & McKoon (2008) Hanks et al. (2014). Researchers must distinguish actual changes in sensory quality from changes in the willingness to commit.
Hanks, Kiani, and Shadlen directly compared LIP activity under speed-emphasis and accuracy-emphasis regimes in macaques. They observed similar terminating firing levels, with stronger evidence-independent activation in the faster regime, a finding that complicated simple interpretations based exclusively on lowering a neural threshold Hanks et al. (2014). The experiment illustrates why boundary parameters in cognitive models need not map one-to-one onto raw neuronal firing thresholds. A behavioral fit can capture a policy shift without proving that a particular neural criterion physically moved. Population coding, motor readiness, and baseline activation are part of the empirical mapping problem.
Normative analyses add the cost of observation, error penalties, and reward timing to the decision process. Drugowitsch and colleagues formalized the consequences of an explicit accumulation cost and showed why time-dependent stopping policies can be advantageous in suitable settings Drugowitsch et al. (2012). Psychological behavior may still deviate from an optimized policy because the observer has imperfect information about task statistics or uses stable heuristics across varying environments. A complete study should manipulate speed instructions separately from incentives, quantify intertrial intervals, and compare learning over time. Otherwise the estimated speed–accuracy policy can confound strategic adaptation, sensory attention, and fatigue.
Confidence, Metacognition, and Decision Certainty
Confidence is best defined operationally as a report or choice that indexes subjective probability of correctness or another explicitly specified criterion of certainty. Confidence may be assessed through verbal ratings, wagers, opting out, or later information search, and each format changes incentives and response demands. Kiani and Shadlen used an opt-out procedure during visual motion discrimination and found that parietal neural activity reflected aspects of decision certainty as well as developing choices Kiani & Shadlen (2009). This links confidence to perceptual evidence while leaving open whether all confidence judgments derive from the same state variables as the primary choice.
Two-stage dynamic signal detection theory proposes that evidence accumulation can continue after the first choice and that confidence is produced by reading the evolving process at a later point. Pleskac and Busemeyer developed this account to explain choice, latency, and confidence jointly Pleskac & Busemeyer (2010). The framework predicts effects of decision timing and the interval between choice and confidence elicitation. It also explains why a single abstract confidence measure may combine evidence strength, stopping policy, and post-decision processing. To evaluate it, experiments need to measure all three behavioral outcomes together and manipulate the timing of the confidence report.
Confidence can influence subsequent sampling rather than merely summarize a decision after it has been made. Balsdon, Wyart, and Mamassian examined confidence-dependent accumulation during perceptual inference Balsdon et al. (2020). Such findings complicate models that treat confidence as a passive readout of a fixed decision variable. A broader explanation may include feedback from an estimate of certainty to attention allocation, sampling termination, and later revision. Apparent overconfidence or underconfidence should therefore be assessed relative to an explicitly calibrated task and payoff structure; a numerical rating without its empirical accuracy relationship does not uniquely reveal metacognitive quality.
Postdecision Accumulation and Changes of Mind
An overt movement toward one option can begin before all relevant sensory information has completed processing. Resulaj, Kiani, Wolpert, and Shadlen demonstrated that human reaching trajectories sometimes reverse toward the alternative after an initial motion decision, even though the stimulus disappeared once the reach began Resulaj et al. (2009). Their account proposed that evidence already in the processing pipeline could continue influencing the ongoing action. This finding expands the object of decision research from a single discrete button press to a temporally extended episode of commitment, monitoring, and correction. Observed motor reversals provide particularly informative behavioral traces because their timing constrains the possible sources of later information.
Subsequent work showed that choices and confidence judgments can both change following initial commitment. Van den Berg and colleagues explained several such revisions through continued accumulation and processing delays rather than requiring a wholly separate correction mechanism van den Berg et al. (2016b) van den Berg et al. (2016a). The main theoretical benefit is parsimony: one evolving evidence state can affect an initial judgment, subjective certainty, and a later correction. However an observed course correction may also reflect motor dynamics, newly attended stimulus features, or a changed task rule. Testing the account requires controlling stimulus offset, movement biomechanics, and the distinction between changes in intention and mere trajectory curvature.
Postdecision processing has implications for measuring accuracy. An initial report and a final report can differ, and allowing a correction can improve or worsen outcomes depending on how additional evidence is weighted. The boundary at which an action starts is therefore task- and stage-specific rather than a universal end to inference. Researchers should identify the first commitment, subsequent evidence availability, confidence elicitation, and final action separately. These distinctions allow a model to specify when commitment becomes irreversible and when adaptive correction remains available. They also prevent changes of mind from being automatically classified as indecision or a breakdown of the original perceptual process Resulaj et al. (2009) Pleskac & Busemeyer (2010).
Multiple Alternatives and the Geometry of Choice
Binary tasks simplify the mathematics of evidence accumulation by representing preference between two hypotheses as a signed variable. Natural perceptual discrimination often involves several possible objects, locations, or actions. Multi-alternative choice requires representing comparative support for multiple candidates and specifying how competing streams interact. Churchland, Kiani, and Shadlen compared two- and four-choice motion decisions in macaques and reported activity patterns suggesting adjustments to the response-set size while preserving accumulation-related dynamics Churchland et al. (2008). These results establish the need for models that handle the geometry of alternatives and the allocation of representational resources.
Independent races, correlated races, and mutual-inhibition models make different assumptions about how evidence for one alternative affects another. In a race formulation, a threshold crossing may be sufficient for choice without a single difference signal; a coupled system can instead cause one candidate to suppress others. Models with divisive normalization or common noise may explain specific response-set effects, but separating them demands experiments that vary the number and similarity of alternatives. The mapping between an option and a planned eye movement can create additional correlations not predicted by abstract evidence models. Thus a neural preference for one saccade target cannot be interpreted independently of the response geometry Usher & McClelland (2001) Churchland et al. (2008).
General claims about a decision architecture should survive changes in possible actions, presentation order, response modality, and stimulus discriminability. The scaling of decision time with alternative count offers a relevant behavioral constraint, yet a single summary law is insufficient to choose between specific implementations. More informative experiments manipulate the relative strengths of multiple candidates while tracking the complete sequence of emerging commitments. Formal analyses also suggest that models mathematically equivalent in the two-choice limit can separate substantially once the alternatives multiply Bogacz et al. (2006) Ditterich (2006). The empirical goal is to distinguish reusable computational principles from task-specific mechanisms.
Experimental Designs and Identifiability
A defensible study of evidence accumulation specifies the physical stimulus, its trial-to-trial variation, the available response options, response deadlines, and reward contingencies before estimating internal processing parameters. Two-alternative forced-choice protocols control guessing probabilities more directly than an unqualified yes–no report, but introduce side and motor preferences that must be counterbalanced. Response-signal experiments constrain observation duration, whereas free-response trials let stopping time vary endogenously. Combining these designs may distinguish changing evidence quality from a varying stopping boundary. Such complementarity is central to comparing sequential-sampling explanations with response-stage accounts Palmer et al. (2005) Ratcliff et al. (2016).
Parameter identifiability is the degree to which distinct model settings produce distinguishable observable data under the actual design. Drift, nondecision time, starting bias, boundary settings, and across-trial variability can trade off if the data contain too few trials or narrow difficulty variation. The remedy includes experimental manipulations designed around discriminating predictions, simulation-based recovery, uncertainty intervals, and out-of-sample assessment. The fact that an optimizer returns stable parameters does not establish their uniqueness. Ratcliff and colleagues emphasize that entire response-time distributions and error patterns impose constraints that summary accuracy cannot provide Ratcliff & McKoon (2008) Ratcliff et al. (2016).
Model comparison is especially demanding when candidate models differ in flexibility. An additional urgency term or collapsing bound can improve in-sample fit merely by adding degrees of freedom. Researchers should predefine core comparisons, inspect predictive errors by condition, and evaluate whether more complex models generalize to untouched participants or paradigms. Formal equivalence results help identify which features cannot be distinguished under a given task Bogacz et al. (2006). A causal theory ultimately requires sensitivity to targeted manipulations, not just statistical preference among curve-fitting procedures. Transparent code, stimulus provenance, and complete condition-level data improve both replication and the interpretability of model selection.
Temporal Weighting and Psychophysical Kernels
A psychophysical kernel estimates how evidence presented at different moments influences the final report. In a stimulus with controlled temporal noise, a regression can relate fluctuations during successive time bins to choice while holding overall evidence constant. A flat kernel may be consistent with approximately uniform integration, whereas primacy or recency patterns may suggest early commitment, changing sensory gain, adaptation, or poststimulus selection. The kernel is a behavioral statistic contingent on stimulus autocorrelation, choice coding, and observer policy. It should not be interpreted as a literal neural impulse response without further controls Huk & Shadlen (2005) Ditterich (2006).
Brief motion pulses provide a complementary causal perturbation because the experimenter directly modifies evidence at a selected moment. Huk and Shadlen found persisting consequences of such perturbations for LIP activity and choice, an observation supporting temporal memory for recent sensory samples Huk & Shadlen (2005). When pulse effects decline late in a trial, possible explanations include terminated accumulation, lower attentional gain, sensory adaptation, or a change in urgency. Distinguishing these mechanisms requires jointly measuring response times, pulse timing relative to commitment, and whether late evidence remains available. Fixed stimulus durations and response-terminated trials can yield markedly different kernel shapes for this reason.
Temporal weighting also clarifies the relation between inference and action. An evidence pulse delivered after an internal stopping event might still reach the motor system too late to alter the first movement yet in time to influence later confidence or correction. Changes of mind experiments exploit precisely this temporal distinction Resulaj et al. (2009). The research aim is to estimate a causal timeline linking sensory encoding, accumulating evidence, stopping, and report execution. Methods that align all trials only to stimulus onset may obscure variability in the internal time of commitment; analyses aligned to action onset and trialwise model-inferred stopping times therefore offer complementary views.
Neural Population Dynamics and Distributed Computation
Single-unit recordings identify detailed temporal response properties but sample a small fraction of the neural population contributing to behavior. A perceptual decision may depend on coordinated activity across many cells, with correlations and shared fluctuations influencing the accuracy of the population code. The observation that single MT neurons can approach psychophysical sensitivity does not imply that the animal reads out exactly one cell; correlated pools can produce comparable behavioral limits Britten et al. (1992). Understanding the computation requires population-level estimates of information, readout weights, and temporal coupling between encoding and action-related representations.
Decision-related signals are distributed across regions that process sensory features, maintain task states, select potential actions, and monitor outcomes. LIP accumulation-like activity Roitman & Shadlen (2002), premotor comparison responses Romo et al. (2004), and a human centroparietal EEG component O'Connell et al. (2012) are complementary observations at different levels of measurement. The same functional term can describe distinct neural realizations, but similarity of waveform or correlation does not establish a one-to-one circuit homology. Task structure determines which populations are most predictive: a reach, saccade, and delayed verbal report recruit different response systems and may change the apparent timing of commitment.
Population analyses should ask whether a signal represents sensory evidence, an evolving categorical belief, prospective action, or a mixture that changes through time. Interventions can test whether altering a recorded population shifts perception, timing, or only the expression of an already-formed decision. Nienborg and Cumming's analysis cautions that strong choice modulation in early sensory neurons may contain feedback or state-related components Nienborg & Cumming (2009). Conversely the persistence of stimulus pulse effects in parietal activity strengthens a temporal integration interpretation under its own conditions Huk & Shadlen (2005). Together these studies motivate a distributed causal model that separates information flow from mere coactivation.
Individual Variability and Hierarchical Inference
Observers differ in sensory sensitivity, caution, attentional stability, motor latency, and capacity to learn task contingencies. Averaging across participants can obscure these differences and may create an apparent decision trajectory that no single observer exhibits. Hierarchical models can represent population-level tendencies while allowing person-specific parameters, but their conclusions depend on the assumed distribution of individual effects. A difference in estimated drift may reflect sensory encoding, attentional engagement, task comprehension, or model misspecification. Separate psychophysical calibration, eye tracking, and confidence measures improve interpretation of individual variation Forstmann et al. (2016) Ratcliff et al. (2016).
Reliability must be established separately for task measures and inferred parameters. A model fit that reproduces aggregate reaction times need not provide stable individual rankings on repeated sessions, especially with sparse trials and parameter correlations. Researchers can assess split-half and test–retest uncertainty, simulate recovery under the actual trial design, and compare fitted changes with independent behavioral indicators. Clinical or educational extrapolation requires stronger measurement validity than demonstrating a group-level experimental effect. The field's general-purpose model vocabulary should not conceal the dependence of parameter meaning on the stimulus set, response demands, and analysis assumptions Ratcliff & McKoon (2008) Forstmann et al. (2016).
Development, aging, expertise, and differences in sensory acuity can affect the observable speed–accuracy relation through multiple routes. Distinguishing them requires age-appropriate tasks and an explicit accounting of perceptual and motor stages. Apparent slowness cannot be attributed automatically to slower evidence accumulation; longer nondecision time or a more cautious criterion may produce similar response latency. Likewise a faster response may reflect better encoding, greater urgency, or impulsive guessing. Robust individual-differences research estimates these components jointly, reports uncertainty and repeatability, and avoids treating a single fitted parameter as a direct clinical or personality diagnosis Ratcliff et al. (2016).
Learning, Context, and Adaptive Decision Policies
Decision policies adapt to stimulus statistics, action costs, and reward contingencies. When one category becomes more probable, an observer may begin closer to the associated boundary or alter the interpretation of incoming evidence. Under asymmetric penalties, maintaining identical criteria for both choices may be inefficient. Formal optimality analyses indicate how priors and payoffs can justify such changes under explicitly stated conditions Bogacz et al. (2006). The experimental challenge is that stimulus statistics may be learned imperfectly and may drift during the session, while attention, fatigue, and motor readiness also change. Tracking choices trial by trial provides evidence that summary psychometric curves can conceal.
Learning in a perceptual task may improve early sensory encoding, stabilize the representation of the hypotheses, or tune a later stopping policy. If accuracy increases with experience, the gain need not occur at one fixed computational stage. Investigators should test whether the psychometric slope, chronometric function, confidence calibration, and evidence-weighting kernel change in parallel. A model that attributes all improvements to drift is a parsimonious starting hypothesis, but attention and task-specific strategy deserve independent measurement. Comparative experiments across input modalities and reward environments can reveal which aspects of adaptation are transferable Palmer et al. (2005) Kelly & O'Connell (2013).
Adaptive behavior also depends on whether time is costly and whether corrective actions remain available. Drugowitsch and colleagues modeled costs of accumulation, while postdecision research shows that commitment can have several partially reversible stages Drugowitsch et al. (2012) Resulaj et al. (2009). These perspectives converge on a temporally organized control problem: the observer chooses how long to inspect evidence, when to initiate an action, and whether to revise it. A full policy description must therefore include both evidence dynamics and the cost structure of action. Experimental manipulation of deadlines and reversal penalties offers stronger leverage than assigning a fixed cognitive style to a participant.
Replication, Model Criticism, and Contradictory Evidence
Research on accumulation enjoys an unusually productive exchange between formal theory, behavior, and electrophysiology, but the evidential strength of particular claims varies. Motion discrimination across sensory strengths, response-time distributions, and stimulus pulse perturbations supports temporal use of uncertain evidence Palmer et al. (2005) Huk & Shadlen (2005). The exact form of an internal accumulator is less securely identified, and competing mathematical implementations may converge on the same observables Bogacz et al. (2006). A reliable synthesis must preserve this distinction. Evidence for a computation at the behavioral level can be substantial while claims about its unique circuit location, noise distribution, or stopping criterion remain under active investigation.
Several findings place pressure on simple mechanistic readings. Nienborg and Cumming found that choice-related sensory activity could include influences beyond feedforward noise driving the decision Nienborg & Cumming (2009). Hanks and colleagues observed speed-policy adjustments in LIP without the straightforward change in terminal firing levels predicted by a literal lower-bound account Hanks et al. (2014). Meanwhile human EEG studies provide converging temporal signatures but cannot resolve the full underlying circuit architecture O'Connell et al. (2012). These results improve theoretical precision by requiring explicit separation of sensory encoding, recurrent feedback, action readiness, urgency, and evidence-based integration.
Replication quality depends on preserving the actual generative structure of a task, including stimulus sampling, eye-movement criteria, payoff rules, and analysis pipelines. Repeating a named paradigm with materially different dot algorithms or response timing can test generality but is not a literal replication of the original measurement conditions. Formal model recovery, preregistered comparisons, code availability, and shared trialwise data would strengthen the empirical basis of specific mechanistic claims. The field also benefits from reporting failed fits, participant exclusions, parameter uncertainty, and alternative mapping hypotheses, because published success on one well-studied task may overstate generality across perception, decisions, and populations Ratcliff et al. (2016) Forstmann et al. (2016).
Neural Mechanism, Representation, and Explanatory Levels
A full account of a perceptual decision connects at least four levels: physical stimulus statistics, encoded neural information, time-evolving representations that support choice, and overt motor or communicative commitment. Each link has a distinct method of validation. Psychophysics identifies how stimulus reliability relates to responses; encoding recordings describe available sensory information; model-based neural analyses estimate decision-variable dynamics; and controlled intervention identifies causal contributions. Gold and Shadlen framed decision formation as a central research problem spanning these levels Gold & Shadlen (2007). The resulting explanation becomes stronger as it specifies transformations between levels rather than merely assigning a psychological name to a brain area.
The inferential category of evidence requires special care. A motion-sensitive neuron can encode a physical feature without directly representing the posterior probability of the direction hypothesis. Conversely a movement-related neuron can predict choice while carrying little information about the instantaneous external stimulus. Decision-related signals may reflect a representation of accumulated sensory evidence, an action policy, or both within recurrently coupled networks Shadlen & Newsome (2001) Nienborg & Cumming (2009). Demonstrating a common computational signature across tasks is an important research result; determining whether it arises from shared neural mechanisms, convergent circuit dynamics, or an abstract descriptive model is a further question.
Investigators should describe evidence for each explanatory transition with its appropriate strength. Observed behavioral covariation supports an association; a well-fitting diffusion model supports a generative description conditional on assumptions; time-resolved perturbation supports a more causal inference; and intervention on a candidate circuit can constrain mechanisms if the manipulation is selective. Delays and feedback introduce additional ambiguity, especially when activity linked to decision formation persists after physical stimulation ends. The most productive research strategy combines these tools with predictions that discriminate alternative architectures and directly examines whether the same features survive changes in action demands Huk & Shadlen (2005) O'Connell et al. (2012).
Applications to Human–Artificial Perceptual Partnerships
Evidence accumulation provides a formal vocabulary for investigating decisions made by humans interacting with artificial perceptual systems. An interface might display sequential estimates, camera detections, a changing uncertainty indicator, or a recommendation that updates as new evidence arrives. The psychological questions concern whether the human observer changes sampling duration, confidence, decision criteria, and willingness to correct an initial judgment. These questions are empirical and can be studied with calibrated signals, recorded trajectories, controlled error rates, and model-based predictions. The existence of a mathematical accumulator within a machine is not itself evidence that a human–machine partnership will achieve improved judgment Forstmann et al. (2016) Gold & Shadlen (2007).
Artificial decision policies can be analyzed with principles resembling the sequential probability ratio test: accumulate informative observations, compare posterior support with a stopping condition, and balance delay against the costs of error. Such an analysis establishes a functional model of evidence use, subject to the system's actual sampling and computational architecture. In the Artificial Era, the scientific question extends to how human and artificial evidence streams are combined, how authority over commitment is allocated, and how both parties interpret uncertainty. Claims about psychological organization of artificial systems require explicit observables and tests; functional similarity does not settle every question about the nature or experience of the system.
A rigorous research design for joint perception could vary artificial reliability independently of human sensory clarity, randomize the timing of machine recommendations, and elicit initial and final choices with confidence reports. Predictions could then distinguish informative evidence integration from mere recommendation-following or generalized deference. Outcome measures should include calibration, decision latency, override rates, postdecision corrections, and robustness under distribution shift. Such studies should also test how explanations and displays affect attention to primary stimuli. The purpose of the present connection is to identify empirically tractable relationships between distinct decision agents, while the canonical core of this dossier remains the established human and comparative science of perceptual choice Pleskac & Busemeyer (2010) Balsdon et al. (2020).
Contemporary Evidence: Confidence, No-Report Dynamics, and Learned Categories
Recent experiments extend the temporal evidence framework by separating neural signatures of subjective confidence from its ordinary behavioral correlates. Dou and colleagues studied human motion decisions with EEG and reported that centroparietal positivity predicted subjective confidence even when task performance, reaction time, and evidence quality were controlled across complementary experiments. Their three experiments used samples of 25, 25, and 24 adults, respectively; those sample sizes limit population-wide generalization but support a focused dissociation under measured experimental conditions Dou et al. (2024). The result indicates that accumulation-related brain signals can carry information relevant to confidence beyond a trivial reflection of correctness. It also makes an important analytical demand: confidence and accuracy must be modeled as distinguishable dependent variables if their relationship is the subject of investigation.
A major unresolved question has been whether accumulation-like cortical responses are exclusively associated with preparing a required report. Stockart and colleagues combined human intracranial recordings from 29 participants across three preregistered experiments with immediate-response, delayed-report, and passive-viewing conditions. In ventral visual cortex, they observed signals associated with detection, stimulus intensity, and confidence, including effects under conditions without immediate overt reports Stockart et al. (2025). A leaky-accumulation model reproduced important behavioral and neural features. This evidence supports a role for temporal integration in visual perception beyond the execution of a response, while requiring attention to the authors' qualifications about possible covert decisions, differences in stimulus intensity, and the extent to which intracranial sampling generalizes to the whole brain.
Contemporary findings also sharpen the question of whether a choice and its confidence are computed serially or concurrently. Vivar-Lazo and Fetsch studied nonhuman primates reporting choice and confidence in a single eye movement; bounded-accumulator and population analyses in lateral intraparietal cortex supported concurrent updating under their task conditions Vivar-Lazo & Fetsch (2026). This does not erase evidence for postdecision accumulation in other paradigms, because different response schedules and informational conditions can support different architectures. In a separate 2026 human EEG study, Thoksakis and Ester found that centroparietal accumulation-related slopes tracked categorical evidence defined relative to a newly learned boundary, and that individual slopes covaried with estimated drift Thoksakis & Ester (2026). Their result extends evidence accumulation to learned representational transformations. Together the new studies favor explicit task-dependent accounts of confidence timing, neural location, and the nature of evidence, tested against alternatives rather than inferred from a shared waveform alone.
Outstanding Scientific Questions and Research Priorities
One unresolved problem concerns how to identify the physiological implementation of an accumulator when multiple coupled circuits display similar buildup dynamics. Simultaneous population recording, transient intervention, and explicit modeling of sensory and movement signals can help determine which activity is functionally necessary for commitment. A second problem is whether a stationary drift and fixed bound adequately describe decisions when evidence reliability changes rapidly within a trial. Response-time distributions alone may underconstrain these possibilities, so experimentally programmed changes in evidence quality should be analyzed together with temporal choice kernels and neural recordings Ratcliff et al. (2016) Huk & Shadlen (2005).
Another research priority concerns confidence and reversibility. A choice can be reported, its confidence updated, and the final action corrected at different moments. Two-stage and continuous-accumulation frameworks capture important observations but must be compared against alternatives involving discrete evaluation stages, memory reconstruction, and motor correction Pleskac & Busemeyer (2010) Resulaj et al. (2009) van den Berg et al. (2016a). A decisive experiment should manipulate precisely when information is available after commitment and predict both the direction of confidence changes and the latency of action revisions. Work across multiple sensory modalities and response effectors is essential to distinguish a general inferential process from familiar laboratory-specific circuits.
Finally, generalization across participants, development, culture, sensory ability, and artificial decision partners needs more systematic evidence. The usual psychophysical experiments involve heavily trained observers or tightly controlled college-laboratory settings; these designs offer internal precision while limiting immediate application to complex environments. Future programs should share exact stimulus-generation code, evaluate individual-parameter reliability, report uncertainty, and use cross-laboratory predictive benchmarks. The object of this research is to understand which temporal decision principles remain stable when information, incentives, and possible actions change. Such results will clarify the legitimate scope of accumulation models without erasing useful task-specific differences Forstmann et al. (2016) Drugowitsch et al. (2012).
Conclusion
Perceptual decisions arise through the temporally organized transformation of uncertain sensory information into commitments that can be measured through choice, response latency, confidence, and subsequent correction. A substantial body of evidence from visual motion and vibrotactile tasks, formal sequential-sampling theory, primate electrophysiology, and human EEG supports the proposition that information accumulated across time contributes to many perceptual judgments Newsome et al. (1989) Hernández et al. (2002) O'Connell et al. (2012). The scientific significance of the field is its ability to connect observable psychophysical performance with quantitative hypotheses about internal dynamics. Its findings illuminate why weak evidence takes longer to evaluate and why the timing of information can affect both choice and action.
The explanatory synthesis differentiates several nested questions: whether evidence is integrated behaviorally; which algorithm best predicts trial-level decisions; which neural signals track the evolving decision state; and which circuits causally implement the computation. Diffusion, race, and leaky-competition models define a productive family of hypotheses, with distinct assumptions about noise, boundaries, alternatives, and time. Experimentally induced evidence pulses and neural recordings constrain these hypotheses, while findings on feedback, urgency, and postdecision correction show how a purely one-stage threshold picture must be refined Usher & McClelland (2001) Hanks et al. (2014) Nienborg & Cumming (2009). The quality of the explanation follows the convergence of independent tests.
Perceptual evidence accumulation is thus a durable organizing principle for research on time-resolved choice, and its scientific value depends on precise operational definitions and reproducible methods. A complete research program jointly estimates stimulus reliability, decision time, action demands, neural representation, and confidence while evaluating alternative causal accounts. Extensions to adaptive human–artificial perception can be designed with the same methodological discipline, asking which evidence streams are trusted and when decisions are updated. The enduring scientific task is to determine how information becomes commitment across time and which elements of that transformation generalize between sensory modalities, organisms, and computational systems Gold & Shadlen (2007) Ratcliff et al. (2016).
References
- Balsdon, T., Wyart, V., & Mamassian, P. (2020). Confidence controls perceptual evidence accumulation. Nature Communications, 11, Article 1753. https://doi.org/10.1038/s41467-020-15561-w
- 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
- Britten, K. H., Shadlen, M. N., Newsome, W. T., & Movshon, J. A. (1992). The analysis of visual motion: A comparison of neuronal and psychophysical performance. Journal of Neuroscience, 12(12), 4745–4765. https://doi.org/10.1523/JNEUROSCI.12-12-04745.1992
- 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
- Ditterich, J. (2006). Stochastic models of decisions about motion direction: Behavior and physiology. Neural Networks, 19(8), 981–1012. https://doi.org/10.1016/j.neunet.2006.05.042
- Dou, W., Martinez Arango, L. J., Graham Castaneda, O., Arellano, L., Mcintyre, E., Yballa, C., & Samaha, J. (2024). Neural signatures of evidence accumulation encode subjective perceptual confidence independent of performance. Psychological Science, 35(7), 760–779. https://doi.org/10.1177/09567976241246561
- Drugowitsch, J., Moreno-Bote, R., Churchland, A. K., Shadlen, M. N., & Pouget, A. (2012). The cost of accumulating evidence in perceptual decision making. Journal of Neuroscience, 32(11), 3612–3628. https://doi.org/10.1523/JNEUROSCI.4010-11.2012
- 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
- Hanks, T., Kiani, R., & Shadlen, M. N. (2014). A neural mechanism of speed-accuracy tradeoff in macaque area LIP. eLife, 3, e02260. https://doi.org/10.7554/eLife.02260
- Hernández, A., Zainos, A., & Romo, R. (2002). Temporal evolution of a decision-making process in medial premotor cortex. Neuron, 33(6), 959–972. https://doi.org/10.1016/S0896-6273(02)00613-X
- Huk, A. C., & Shadlen, M. N. (2005). Neural activity in macaque parietal cortex reflects temporal integration of visual motion signals during perceptual decision making. Journal of Neuroscience, 25(45), 10420–10436. https://doi.org/10.1523/JNEUROSCI.4684-04.2005
- Kelly, S. P., & O'Connell, R. G. (2013). Internal and external influences on the rate of sensory evidence accumulation in the human brain. Journal of Neuroscience, 33(50), 19434–19441. https://doi.org/10.1523/JNEUROSCI.3355-13.2013
- Kiani, R., & Shadlen, M. N. (2009). Representation of confidence associated with a decision by neurons in the parietal cortex. Science, 324(5928), 759–764. https://doi.org/10.1126/science.1169405
- Newsome, W. T., Britten, K. H., & Movshon, J. A. (1989). Neuronal correlates of a perceptual decision. Nature, 341, 52–54. https://doi.org/10.1038/341052a0
- Nienborg, H., & Cumming, B. G. (2009). Decision-related activity in sensory neurons reflects more than a neuron's causal effect. Nature, 459, 89–92. https://doi.org/10.1038/nature07821
- 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
- Palmer, J., Huk, A. C., & Shadlen, M. N. (2005). The effect of stimulus strength on the speed and accuracy of a perceptual decision. Journal of Vision, 5(5), 376–404. https://doi.org/10.1167/5.5.1
- Pleskac, T. J., & Busemeyer, J. R. (2010). Two-stage dynamic signal detection: A theory of choice, decision time, and confidence. Psychological Review, 117(3), 864–901. https://doi.org/10.1037/a0019737
- 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
- Resulaj, A., Kiani, R., Wolpert, D. M., & Shadlen, M. N. (2009). Changes of mind in decision-making. Nature, 461, 263–266. https://doi.org/10.1038/nature08275
- Roitman, J. D., & Shadlen, M. N. (2002). Response of neurons in the lateral intraparietal area during a combined visual discrimination reaction time task. Journal of Neuroscience, 22(21), 9475–9489. https://doi.org/10.1523/JNEUROSCI.22-21-09475.2002
- Romo, R., Hernández, A., & Zainos, A. (2004). Neuronal correlates of a perceptual decision in ventral premotor cortex. Neuron, 41(1), 165–173. https://doi.org/10.1016/S0896-6273(03)00817-1
- 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
- Stockart, F., Msheik, R., Robin, A., Jurkovičová, L., Goueytes, D., Rouy, M., Mareček, R., Hoffmann, D., Mudrik, L., Roman, R., Brázdil, M., Minotti, L., Kahane, P., Pereira, M., & Faivre, N. (2025). Cortical evidence accumulation for visual perception occurs irrespective of reports. Nature Communications, 16, Article 8458. https://doi.org/10.1038/s41467-025-63255-y
- Thoksakis, A., & Ester, E. F. (2026). Neural measures of human decision-making track evidence accumulation in learned space. Journal of Neuroscience, 46(18), Article e0134262026. https://doi.org/10.1523/JNEUROSCI.0134-26.2026
- 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 den Berg, R., Zylberberg, A., Kiani, R., Shadlen, M. N., & Wolpert, D. M. (2016). Confidence is the bridge between multi-stage decisions. Current Biology, 26(23), 3157–3168. https://doi.org/10.1016/j.cub.2016.10.021
- van den Berg, R., Anandalingam, K., Zylberberg, A., Kiani, R., Shadlen, M. N., & Wolpert, D. M. (2016). A common mechanism underlies changes of mind about decisions and confidence. eLife, 5, e12192. https://doi.org/10.7554/eLife.12192
- Vivar-Lazo, M., & Fetsch, C. R. (2026). Neural basis of concurrent deliberation toward a choice and confidence judgment. Nature Neuroscience, 29(1), 159–170. https://doi.org/10.1038/s41593-025-02116-9