Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Multisensory Integration and Causal Binding

Abstract. Multisensory integration is the perceptual coordination of evidence arriving through distinct sensory modalities. Causal binding is the inference that different sensory measurements arise from a common environmental source, and causal segregation is the inference that they arise from separate sources. This fundamental research monograph examines how these operations interact in spatial localization, visual–haptic discrimination, audiovisual speech, perceptual event segmentation, temporal correspondence and naturalistic scenes. A structured critical synthesis traces the historical development of crossmodal research from psychophysical and superior-colliculus experiments to reliability-weighted maximum-likelihood estimation, Bayesian causal inference, hierarchical cortical models and recent methodological revision. Primary behavioral and neurophysiological studies are analyzed alongside research on temporal recalibration, development, attention, observer-specific response rules, implicit spatial representations and computational parameter identifiability. Studies published through October 2026 show why a single illusion magnitude or temporal binding window can reflect several interacting factors, including unimodal sensory noise, environmental priors, causal inference and decision policy. This dossier evaluates the resulting evidence with explicit attention to experimental design, cross-task generalization, alternative mechanisms and the distinction between computational predictions and biological implementation. Its synthesis identifies future research priorities in neural causal testing, source correspondence in crowded environments, reliable individual measurement, learning, and comparative evaluation of artificial sensor-fusion systems.

Keywords: multisensory integration; causal binding; Bayesian causal inference; common-cause inference; sensory reliability; cue combination; maximum-likelihood estimation; auditory–visual integration; visual–haptic integration; McGurk effect; sound-induced flash illusion; temporal binding window; ventriloquism; superior colliculus; cortical hierarchy; crossmodal recalibration; perceptual learning; sensor fusion; Artificial Era

The Scientific Problem of Multisensory Perception

The perceptual world is organized around objects and events, although the nervous system receives partly independent patterns of light, sound, mechanical pressure, and bodily movement. Multisensory integration concerns how information from two or more sensory modalities changes a perceptual estimate, discrimination, or action relative to the evidence furnished by each modality alone. Causal binding concerns the additional inference that the signals belong to the same external event or source. The two problems interact because adding information is useful only when the relevant signals can legitimately be treated as evidence about a shared property. A scientifically adequate account therefore describes sensory precision, source correspondence, spatial and temporal relations, prior expectations, task instructions, and response criteria together. Noppeney (2021) develops this broader framework through inference, attention, and learning.

An observer watching a distant speaker receives visual articulation before an acoustic waveform arrives; touching a visible object adds haptic estimates whose precision may differ from optical estimates. These situations can support unified perception despite discrepancies in modality, propagation speed, and encoding delay. The central scientific question is how a perceptual system determines when differences are tolerable evidence about one cause and when they indicate separate causes. Binding is measurable through localization attraction, categorical responses, simultaneity judgments, discrimination improvement, or neural interactions, but none of these measures is a direct readout of an abstract binding variable. Körding et al. (2007) supplies an explicit computational treatment of common-cause inference in audiovisual localization.

This dossier treats multisensory perception as an experimentally tractable family of estimation and causal-inference problems. It surveys foundational animal physiology, psychophysics, audiovisual speech, computational inference, cortical systems, temporal recalibration, development, and applications to artificial perception. Each research tradition addresses a distinct aspect of the explanatory system. The analysis distinguishes a normative model, which specifies what an idealized observer would do under assumptions, from a process model specifying mechanisms or response strategies. van Atteveldt et al. (2014) emphasizes flexible neural operations, whereas Stein & Stanford (2008) addresses multisensory convergence and modulation at the single-neuron level. Their differing levels of analysis are complementary when their empirical predictions are explicitly matched.

Definitions, Boundaries, and Levels of Explanation

Sensory cue combination is a mathematical operation on evidence about a property, such as location or size. Perceptual fusion denotes a unified resulting experience or decision. Causal binding refers to attributing two or more inputs to a common source; segregation denotes maintaining multiple candidate sources. Temporal binding refers to the influence of timing on perceived unity and often is operationalized by simultaneity or temporal-order judgments. Crossmodal recalibration describes an experience-dependent adjustment of perceived alignment or sensory representation. These terms should be tied to measurement procedures rather than treated as interchangeable labels for one hidden faculty. Vroomen & Keetels (2010) distinguishes mechanisms underlying perceived synchrony, and Fujisaki et al. (2004) experimentally documents shifts after repeated audiovisual lag.

The primary dependent variables are frequently estimates of location, thresholds, psychometric slopes, choice probabilities, reaction-time distributions, and errors under conflicting sensory cues. Neural studies additionally quantify spikes, local field potentials, scalp electrophysiology, hemodynamic signals, and stimulus-driven connectivity. A neural response exceeding the response to one modality alone establishes an interaction under particular comparison conditions; it does not by itself prove a common conscious percept or identify a Bayesian algorithm. Meredith & Stein (1986b) demonstrates convergence of visual, auditory, and somatosensory signals in superior colliculus cells. Driver & Noesselt (2008) reviews why crossmodal influences also occur in structures traditionally studied as sensory specific.

Three analytical levels must be held apart. At the computational level, a theory describes goals, priors, likelihoods, losses, and the causal structure to be inferred. At the algorithmic level, it describes weighting, sampling, selection, attention, or temporal alignment procedures. At the physiological level, it describes circuits whose responses implement parts of those computations. A statistically successful model can characterize behavior without uniquely establishing its neuronal implementation. Similarly, finding a cortical region responsive to two modalities does not specify its contribution to the participant's perceptual decision. Rohe & Noppeney (2015) advances the link between hierarchical cortical representations and Bayesian causal inference, providing a test case for this multi-level interpretation.

Historical Formation of the Field

The modern field emerged through converging experimental traditions rather than one founding experiment. Psychophysics established procedures for measuring sensory judgments and their uncertainty; auditory and visual research investigated how signals were localized and grouped; and comparative neuroscience demonstrated the convergence of sensory pathways in orienting systems. Early audiovisual speech research then made the problem especially striking by showing that seeing a talker's articulation can alter what listeners report hearing. In McGurk & MacDonald (1976), incongruent lip movements and speech sounds produced systematic perceptual reports different from the presented auditory syllable. The finding was influential because it made audiovisual interaction observable as a content change, rather than merely a speed benefit.

In subsequent decades, experiments showed that sound could reshape judgments of apparent visual motion and even the perceived number of flashes. The correspondence by Sekuler et al. (1997) documented a sound-dependent visual-motion phenomenon, while Shams et al. (2000) showed that auditory beeps can produce an additional perceived flash. These demonstrations undermined a simple, universal hierarchy in which vision always overrides other senses. The direction of influence depends on what is being judged, how reliable the available measurements are, and whether the signals are treated as belonging together. An illusion demonstrates a robust conditional phenomenon, but the same perceptual response may be generated by multiple computational and physiological mechanisms.

A quantitative change in explanation came with the application of statistically principled cue weighting. Ernst & Banks (2002) compared measured visual and haptic uncertainty with a maximum-likelihood combination model; Alais & Burr (2004) extended closely related reasoning to audiovisual localization, including conditions where visual localization was degraded. Körding et al. (2007) integrated source-cause inference with spatial estimation, formalizing why integration should vary with the evidence of one or two external sources. The later research program examined model-selection rules, cortical organization, temporal correlation, developmental change, and adaptation. The historical sequence is scientifically useful because progressively stronger experimental designs separate signal reliability from correspondence beliefs and response strategy, rather than assuming that any crossmodal effect establishes the complete mechanism.

Neural Convergence and the Superior Colliculus

A major physiological foundation is the study of neurons whose firing changes when stimuli from different modalities co-occur. Meredith & Stein (1986b) recorded multisensory convergence within the superior colliculus, a structure important for orienting behavior. Its observations showed that combined visual, auditory, and somatosensory stimulation can alter responses relative to unimodal conditions. This animal work provided direct evidence of neural multisensory interaction, while leaving open how a particular spike pattern relates to human conscious unity or statistically optimal perceptual estimation. It established that the combination of signals is a property of identifiable neural circuits under controlled stimulation, not solely a linguistic description of behavioral reports.

Spatiotemporal arrangement proved decisive. Meredith et al. (1987) demonstrated that the temporal relation between stimuli shapes whether superior colliculus neurons enhance or suppress their responses; Meredith & Stein (1986a) examined the importance of spatial relations for multisensory neuronal activity. These findings are commonly summarized through principles of spatial and temporal coincidence and through sensitivity to the relative effectiveness of unimodal inputs. Their application requires care: the size of a combined neural response depends on baseline strength, stimulus geometry, recording conditions, and the metric chosen for enhancement. A large proportional gain when single-modality inputs are weak does not directly establish a universal law governing perceptual accuracy in every cortical task.

Subcortical orienting circuits offer particularly tractable systems for investigating multisensory convergence and action selection. Their evidence cannot be transferred wholesale to cortical operations that support speech recognition, abstract object identity, or strategic decisions. Stein & Stanford (2008) synthesizes the neural integration tradition and explains how physiological changes can influence salience and behavior. A rigorous cross-level analysis asks whether a manipulation of spatial coincidence or temporal lag changes both neural signals and independently measured behavioral predictions. It also examines whether the effects survive controls for alerting, arousal, motor readiness, and changes in single-modality responses. Those controls help isolate integrative computation from more general consequences of stimulus intensity or salience.

Reliability-Weighted Cue Combination

When two measurements estimate the same continuous quantity, a familiar normative model assumes unbiased Gaussian errors, independent noise conditional on the underlying value, and a single shared source. The combined estimate is the sum of each cue estimate multiplied by a weight proportional to its precision, the reciprocal of its variance. For visual and haptic estimates, the predicted visual weight is haptic variance divided by the sum of visual and haptic variances; the predicted combined variance is the product of the two variances divided by their sum. These equations yield a clear experimental test: first measure each cue's precision, then compare audiovisual or visuo-haptic behavior against the prediction made without fitting new weighting parameters. Ernst & Banks (2002) performed this kind of quantitative comparison.

Reliability weighting clarifies why a sensory modality can dominate one task and become subordinate in another. Alais & Burr (2004) found that clear visual targets strongly influenced perceived sound location, whereas severe visual blur permitted auditory influence over perceived visual position. Intermediate conditions showed combination rather than categorical capture, and joint localization precision could improve. The term statistically optimal is conditional: the benchmark assumes a specified loss function, noise distribution, independence structure, and common source. Any systematic bias, correlated errors, mismatched coordinate frames, or uncertain causal correspondence can alter the optimal weighting rule. A model matching one set of psychophysical thresholds is informative precisely because its assumptions can be manipulated and challenged in other conditions.

Maximum-likelihood fusion should therefore be regarded as a conditional component of a broader theory. When sensory cues plausibly indicate different events, averaging their locations may create an estimate that corresponds to neither event. When their errors are correlated, treating them as independent exaggerates the informational gain from combination. When an observer must identify a categorical speech sound rather than a continuous position, the relevant likelihoods, latent variables, and decision costs are different. Körding et al. (2007) places weighting inside a causal mixture model; Noppeney (2021) situates cue reliability within a dynamically adapted perceptual system. The general empirical requirement is to measure unimodal uncertainty at the time of the multisensory test, rather than assuming fixed reliability from sensory labels alone.

The Common-Cause Inference Problem

A pair of sensory signals can represent a single event, as when the sound and sight of an impact arise together, or different events, as when one person speaks while another moves. Binding becomes a statistical inference over alternative causal structures. In the basic two-cue case, the observer considers a common-cause hypothesis and a separate-causes hypothesis. Under the common-cause hypothesis, the measurements share one latent environmental variable; under separate causes, they are generated by distinct latent variables. The posterior probability of a common cause depends on the prior probability of that hypothesis and the relative likelihood of the observations under each structure. Körding et al. (2007) used this conceptual architecture to explain nonlinear audiovisual localization responses.

Spatial separation, relative timing, temporal correlation, and learned source regularities inform this inference. A small difference in estimated location may be compatible with a shared cause when measurement noise is substantial; the same difference becomes diagnostic of separate causes when both senses have high precision. Thus a spatial discrepancy is interpreted jointly with sensory uncertainty, rather than against one universal separation threshold. The prior for common causation additionally reflects task and environmental expectations and can be estimated only under a specified observer model. Parise et al. (2012) experimentally demonstrated that correlated fluctuations in the temporal structure of audio and visual signals can influence whether their combined localization approaches the statistically optimal precision expected under integration.

An important implication is that a single behavioral index, such as the magnitude of visual capture of sound, cannot uniquely identify a tendency to bind. Strong attraction can result from large common-cause prior probability, low auditory precision, systematic unisensory biases, or a response-level rule. Conversely weak attraction can arise from very precise auditory information even when the observer assigns substantial probability to common causation. Odegaard et al. (2015) analyzed biases in visual, auditory, and audiovisual spatial perception and illustrates why separate measurement of each modality is essential. Model-based research must therefore identify which latent parameter is being estimated and show that the experimental design can distinguish it from alternatives.

Bayesian Models of Integration and Segregation

A Bayesian causal-inference model organizes the task in three stages: specifying distributions of possible environmental causes, describing the probabilistic transformation from causes to sensory measurements, and calculating posterior beliefs conditional on the observed evidence. Each causal configuration generates a potentially different posterior estimate of location or another task-relevant variable. When cues share one source, their likelihoods contribute jointly; when separate sources are inferred, their estimates are maintained separately. The observer then selects an action or report according to a loss function. Körding et al. (2007) demonstrated that this architecture predicts patterns beyond the linear averaging implied by a fixed fusion model, especially when audiovisual discrepancies become sufficiently large.

Normative inference does not dictate one final response rule. Under model averaging, an observer combines estimates derived under alternative causal hypotheses, weighted by their posterior probabilities. Under model selection, the observer chooses the most probable causal structure and responds within it. Under probability matching, the observer samples a structure according to posterior probabilities before responding. These strategies can produce distinguishable response distributions even if they use the same sensory likelihoods. Wozny et al. (2010) compared such rules in audiovisual spatial judgments from more than one hundred participants and reported behavior for many observers that was more consistent with probability matching than with the other examined strategies. This finding challenges any presumption that minimizing squared localization error exhausts human perceptual objectives.

What counts as optimal depends upon both the generative model and the cost of errors. Wrongly binding two independent causes can be costly in navigation or threat localization, while missing a common cause can sacrifice precision. Observers may also value learning, exploration, or robust performance under changing distributions. A description such as near-Bayesian behavior therefore requires specification of the exact parameterization, competing decision rules, and goodness-of-fit procedure. Noppeney (2021) treats binding, attention, learning, and decision-making as interacting demands. The task of scientific explanation is to determine which assumptions are psychologically justified, how model parameters are identified, and where predictions remain stable when observers encounter novel stimulus arrangements.

Causal Priors, Likelihoods, and Identifiability

The prior probability of a shared source is a theoretically meaningful construct only when it is separated from the likelihoods and from response biases. In practical experiments, a high estimated common-cause prior may compensate for poorly estimated unisensory noise or for an inappropriate model of the spatial distribution of sources. Calibration trials that isolate vision and hearing, repeated location sampling, and varying spatial discrepancies help constrain those parameters. Odegaard et al. (2015) argues through measured spatial biases that audiovisual judgments are shaped by features of sensory representations not captured by an assumption of unbiased Gaussian cues. Reliable model comparison requires sensitivity analyses to measurement distortions and plausible alternative distributions.

A hierarchy of hypotheses can also include more than two causes and several cues. Natural auditory scenes frequently contain multiple speakers, reflections, and moving objects; vision can display several plausible sound sources. Under these conditions, the observer must infer which auditory stream corresponds to which visible entity and whether any should be paired. The elegant two-cue common-versus-separate model gives a foundation, while real-world correspondence requires multiple assignments, structured temporal trajectories, and potentially combinatorial search. Noppeney (2021) notes the increased importance of attention when many time-varying signals compete for binding. A model useful for simple flashes and tones should be evaluated anew before being generalized to conversational scenes or complex movement.

Fitting a good-looking integration curve is insufficient evidence that each fitted parameter captures a unique psychological cause. Simulation-based parameter recovery tests ask whether known parameters can be recovered from synthetic data with the intended design. Model recovery tests ask whether competing algorithms can be distinguished under plausible noise. Hierarchical fits can stabilize estimates for individuals, yet pooling assumptions need scrutiny when observer strategies genuinely vary. Posterior predictive checks should examine response distributions, localization bias, variance, and behavior on previously unfit conditions. Wozny et al. (2010) provides an instructive example of alternative decision-rule comparison, while Körding et al. (2007) establishes the significance of explaining nonlinear cue interactions. A strong research program reports not only a preferred model but also cases in which it fails.

Spatial Correspondence and Crossmodal Localization

Spatial congruence is a powerful clue to common environmental origin, although it is rarely sufficient on its own. In a typical ventriloquism experiment, the participant localizes a sound presented near a visual stimulus whose position is independently manipulated. The attraction of perceived sound location toward the visual cue is measured as a function of disparity. Alais & Burr (2004) combined this paradigm with changes in the precision of visual localization. Their finding that auditory influence can become dominant under degraded vision makes sensory reliability a falsifiable element of explanation, rather than a post hoc appeal to visual supremacy. The experiment also shows why judgments should be characterized along a continuum of conflict conditions.

Spatial information can differ in format across modalities. Retinal position, head-centered acoustic direction, hand-centered touch, and body-relative proprioception are represented in coordinate systems that depend on gaze and posture. Transformations between these systems can introduce bias and uncertainty before the alleged integration operation begins. A shift in reported audiovisual location may therefore reflect a problem in reference-frame conversion, a common-cause inference, or a weighted estimate after binding. Meredith & Stein (1986a) established spatial dependence of multisensory neural interactions in a nonhuman model. Odegaard et al. (2015) demonstrated the importance of accounting for systematic spatial biases when describing human unimodal and bimodal perception.

The most informative designs perturb several determinants independently. They change sensory precision while holding physical disparity constant, manipulate learned source probability while holding reliability constant, and sample enough source positions to detect central tendencies or spatial priors. They also separate report modalities, because asking for the sound position can yield different results from asking for the visual position in the same event. Körding et al. (2007) examined audiovisual localization tasks that combine uncertainty about cause and uncertainty about location. The resulting asymmetry and nonlinear dependence across discrepancies provide sharper evidence for causal inference than a one-point demonstration of spatial attraction.

Temporal Correspondence, Synchrony, and Binding Windows

The brain encounters sensory signals at different times because of physical transmission, receptor dynamics, and central processing. Perceived synchrony is consequently an inference about event timing under uncertainty. Experimental tasks include audiovisual simultaneity judgment, temporal-order judgment, temporal ventriloquism, and the assessment of how lag changes an illusion or identification performance. The temporal binding window is an operational range of delays within which a particular measure indicates a substantial tendency to combine or treat signals as belonging together. It is task-dependent and can be asymmetric across visual-leading and auditory-leading configurations. Vroomen & Keetels (2010) distinguishes window breadth, compensation for timing variability, longer-term adaptation, and immediate crossmodal shifts as separable mechanisms.

A temporal window inferred from an observer's simultaneity reports should not be equated with a fixed physiological interval in which all sensory integration occurs. Decision criteria, the visibility of a visual onset, auditory acuity, trial order, stimulus intensity, attention, and the response options can all change the measured psychometric curve. Similarly, susceptibility to the sound-induced flash illusion depends on the temporal relationship of beeps and flashes but also on stimulus reliability and the observer's interpretation of cause. Shams et al. (2000) established the striking phenomenology of auditory alteration of flash count; Shams et al. (2005) explored the resulting responses with an ideal-observer framework. These experiments are informative when treated as model constraints rather than interchangeable direct measures of neural binding strength.

Individual differences in temporal sensitivity can relate to the propensity to experience audiovisual illusions. Stevenson et al. (2012) investigated associations between measures of the multisensory temporal binding window and susceptibility to audiovisual illusions, showing a research route from simple psychophysics to person-level variability. Such correlations are scientifically useful but can reflect shared sensory thresholds or decision strategies. Establishing the cause of a broader binding window requires longitudinal or experimental interventions and careful checks of task reliability. More generally, the correspondence judgment should incorporate both time and space, because signals separated in one dimension can still provide strong source evidence through matched patterns in another.

Audiovisual Speech as a Binding Experiment

Spoken communication provides a rich test of causal binding because the relevant auditory waveform and visible articulatory movements contain related temporal structure and information about a common linguistic event. In the classic experiment by McGurk & MacDonald (1976), mismatching a visible articulatory gesture with an auditory syllable produced reports of sounds differing from the physical acoustic stimulus. The effect demonstrated that visual speech can influence the perceived phonetic category. Its theoretical importance lies in the fact that successful listening depends on integrating partial evidence about the speech-generating event rather than reading out a complete auditory representation in isolation. The original demonstration remains a landmark, while the prevalence of any particular response depends on stimuli, speakers, recording quality, language experience, instructions, and the listener's sensory abilities.

Audiovisual speech differs from simple light-and-tone localization because binding involves evolving, high-dimensional signals whose properties correspond at phonetic and articulatory levels. A visible mouth opening predicts the timing and some identity information of an imminent sound; however, physical synchrony and statistical covariation do not guarantee that the two streams belong together in a crowded scene. Source identification, attention to a talker, and learned speech regularities all shape correspondence. van Wassenhove et al. (2005) used psychophysics and electrophysiology to show that visual speech can facilitate the early processing of auditory speech, with effects related to how informative the visual articulatory pattern is about the auditory target. This result broadens integration beyond a late, consciously reported categorical illusion.

One experimental challenge is to distinguish phonetic fusion from ordinary response selection. Participants may report a syllable because of an altered percept, a categorical expectation, or a rule for choosing among presented alternatives. Studies strengthen their inferences through diverse consonant contrasts, graded acoustic noise, independent lipreading and hearing assessments, reaction times, and testing of stimulus-generalization. McGurk & MacDonald (1976) shows a clear condition-specific phenomenon, but does not establish one universal model of audiovisual speech processing. Noppeney (2021) places naturalistic speech among a broader set of perceptual inference demands in which attention, reliability, causal beliefs, and temporal prediction must be considered jointly.

Sound-Induced Visual Illusions and Event Segmentation

The sound-induced flash illusion offers a compact demonstration of causal binding across vision and hearing. In the experiment by Shams et al. (2000), auditory beeps accompanying a visual flash changed the number of flashes participants reported seeing. The task is distinctive because the relevant dependent variable is apparent event numerosity rather than a shift in spatial location. It therefore tests whether auditory temporal structure contributes to the segmentation of visual events. The effect gives empirical grounds for studying perception as the construction of event representations from multiple streams. Yet its magnitude depends on timing, luminance, sound properties, response method, and observer-specific characteristics; laboratory protocols must specify these parameters to make comparisons meaningful.

An important theoretical question is whether the brain treats beeps and flashes as evidence from one event sequence and then combines their information about numerosity. Shams et al. (2005) compared human responses with an ideal-observer account across conditions displaying different degrees of influence and found broadly consistent patterns. Such agreement supports statistical integration as a useful explanatory model. It does not by itself identify a particular circuit or guarantee that one fixed algorithm explains audiovisual localization, speech fusion, and numerosity judgments. A model of illusory flashes must represent event count, timing uncertainty, prior beliefs about audiovisual co-occurrence, and the cost of conflicting judgments in a manner appropriate to that task.

The visual stream-bounce phenomenon shows a related effect on causal interpretation of motion. Sekuler et al. (1997) reported that sound can alter the perceived motion of intersecting moving objects, changing whether observers experience a crossing or a collision-like bounce. A short sound near the point of intersection is informative because collision events often generate abrupt acoustic transients. Here auditory information can shape the interpretation of an ambiguous visual trajectory without providing another position estimate of the same kind. Together with the flash illusion, this demonstrates that multisensory integration includes constraints on event segmentation and causal interpretation. A scientific explanation should identify the common features of the tasks while preserving the different latent variables that each observer must infer.

Temporal Correlation as Evidence of Common Cause

The temporal structure of a signal can provide stronger correspondence information than the mere fact that two stimuli occurred near each other. A sequence of acoustic intensity changes may covary with a visible motion trajectory; independent events are less likely to exhibit a consistent shared pattern. Parise et al. (2012) tested this idea by manipulating audiovisual temporal correlation in localization tasks and found the precision advantage associated with statistically optimal combination when audiovisual signals were temporally correlated. The result supports the use of fine-grained temporal relations as evidence for common causation. It also helps explain how an observer may bind signals across a modest spatial mismatch when their fluctuations track each other.

Correlation-based correspondence involves estimating structure across time, which creates a different experimental demand from a single synchrony judgment. Researchers must control common amplitude envelopes, rhythmic predictability, stimulus energy, and the possibility that one modality simply alerts the observer to an event in the other. Demonstrating a correlation-dependent precision benefit requires unimodal baselines and explicit tests of the predicted combined variance, rather than faster responses alone. Parise et al. (2012) provides a model case; Colonius & Diederich (2004) illustrates that speeded orienting has its own temporal interaction models. These observations motivate mechanistic studies of crossmodal temporal tracking without assuming that measured correlation always corresponds to a genuinely shared environmental source.

Natural scenes present a further challenge because physically independent signals may be correlated through external regularities, such as a machine generating several synchronized outputs, and common-source streams may be intermittently occluded or masked. Binding under these conditions requires representation of temporal persistence, the possibility of missing cues, and confidence that correlation will generalize. Noppeney (2021) examines how attention helps approximate correspondence judgments when multiple competing streams are present. A robust source-inference theory should therefore combine instantaneous likelihoods with structured predictions over time and test whether the predicted grouping changes as evidence accumulates.

Temporal Recalibration and Crossmodal Learning

Repeated exposure to an audiovisual lag can change the delay at which an observer reports the signals to be synchronous. Fujisaki et al. (2004) experimentally demonstrated shifts in subjective audiovisual simultaneity following minutes of exposure to a consistent lag; the exposure also changed temporal tuning of an auditory-induced visual illusion. The finding establishes plasticity in measured temporal alignment and supports the possibility that experience recalibrates perceptual timing mechanisms. It does not mean that all adjustment occurs at one processing stage or that the perceptual system maintains a permanently correct clock for every modality. Recalibration can be stimulus-dependent, task-dependent, and sensitive to the experienced distribution of discrepancies.

A causal account of recalibration must distinguish several possible adaptations. The observer might change the expected physical lag between two correlated signals, alter the internal timing of a modality, modify the decision criterion for reporting simultaneity, or change the likelihood assigned to a common cause. These mechanisms can generate similar psychometric shifts in one task yet differ on transfer to new stimuli or response formats. Vroomen & Keetels (2010) reviews several distinct solutions to intersensory timing disparities, including flexible windows, compensation, longer-term shifts, and immediate attraction of temporal estimates. Cross-task transfer and recovery after exposure become critical measurements for discriminating the alternatives.

Learning and causal inference may be dynamically coupled. If persistent mismatches are judged to have arisen from one event, updating sensory alignment can be useful; if they come from unrelated events, recalibration could introduce a harmful distortion. Conversely a newly calibrated internal representation may change future evidence about whether cues belong together. Noppeney (2021) frames the problem as adaptive multisensory perception across timescales. This interaction suggests testable manipulations of common-cause probability during the training phase, followed by independent measurements of unisensory timing, multisensory judgments, and the retention of recalibration. The empirical relationship between recalibration, causal belief, and long-term stability remains an active target for computational and physiological investigation.

Selective Attention and Perceptual Grouping

Attention influences which sensory information enters a correspondence judgment and which competing sources receive priority. In a crowded scene, several visual objects and auditory streams may occur at overlapping times. Selecting a plausible pair reduces the combinatorial problem of matching all candidate sources, though selection can introduce bias toward already salient or expected events. Noppeney (2021) emphasizes attentional mechanisms as approximations to causal binding under naturalistic complexity. A well-controlled experiment manipulates spatial attention, feature-based attention, or task relevance while independently measuring stimulus salience, so that an apparent integration change can be attributed to an explicit component of the processing chain.

A facilitation in reaction time under combined stimulation can reflect enhanced sensory evidence, changes in alertness, faster response preparation, or statistical advantages from having two chances to detect a target. Colonius & Diederich (2004) modeled the effects of temporal integration on saccadic responses and showed why their temporal structure matters. Speed advantages are thus informative only when compared against plausible statistical facilitation baselines and relevant alternatives. Behavioral experiments often need both accuracy and complete response-time distributions. Neural response modulation may coexist with unchanged perceptual reports or with changes in action that follow from attention rather than explicit perceptual fusion. Stein & Stanford (2008) places response enhancement and behavioral benefit in a common physiological framework while retaining the need for task-specific controls.

Research on attention also raises a key question about apparent automaticity. Crossmodal influences can arise under divided attention or when one modality is designated irrelevant, but this does not establish that correspondence is indifferent to task demands. Changes in visual quality, competing speaker number, expectancy, and temporal predictability may alter both selection and integration. Driver & Noesselt (2008) describes multisensory influences at cortical stages traditionally labeled unimodal, indicating that selection can operate in distributed networks. A complete account must examine how attentional allocation affects sensory likelihoods, priors, and the final decision separately, rather than use the word automatic as an all-purpose description of any early or involuntary interaction.

Cortical Hierarchies and Neural Mechanisms

Neural explanations of multisensory perception require measurements that reveal how information and causal interpretations change across processing stages. Classical descriptions of early sensory cortex as entirely modality-specific have been challenged by evidence of crossmodal influences on supposedly unisensory regions. Driver & Noesselt (2008) reviews anatomical, electrophysiological, and neuroimaging observations that indicate multiple routes for interactions, including feedback and attentional modulation. Such observations call for a network-based explanation while preserving the distinction between crossmodal modulation and a jointly computed perceptual estimate. A sound-induced response change in visual cortex does not automatically imply that a cell represents a complete audiovisual event or calculates the posterior probability of a shared cause.

One influential human neuroimaging experiment by Rohe & Noppeney (2015) compared sensory signals with predictions from Bayesian causal-inference models and reported a hierarchy of representations across cortical regions. The research associated lower-level representations more closely with modality-specific estimates and higher-level representations with integrated causal estimates. This is a substantive bridge between behavioral model predictions and neural data. However, the inferential distance between hemodynamic pattern similarity and neuronal computation remains important: multivariate representations demonstrate decodable information under particular analysis assumptions, while identifying the actual circuit computation requires complementary temporal measurements and interventions. A cortical hierarchy may also represent changing task and decision demands alongside sensory evidence.

The physiology surveyed by Stein & Stanford (2008) and the operations cataloged by van Atteveldt et al. (2014) point toward distributed and flexible mechanisms, rather than a single region that universally binds all modalities. Proposals include feedforward convergence, recurrent interactions, predictive feedback, divisive normalization, and network-level reweighting. Their relative contributions can vary with stimulus timing, task relevance, reliability, training history, and the modality pair being tested. The next generation of explanation depends on linking local neural dynamics to explicit generative models and to observer-level choices using interventions that distinguish integration, segregation, and decision formation. Differences in experimental species, anesthesia, tasks, and response modalities must be retained when integrating results across laboratories.

Development and Lifespan Change

Multisensory integration changes with the development of sensory precision, motor experience, task understanding, and expectations about the physical environment. Children may respond to multiple cues without combining them in the variance-minimizing manner characteristic of an ideal observer with stable sensory estimates. In Gori et al. (2008), young children showed limited integration of visual and haptic form information under the studied task conditions, supporting the idea that developing systems may use information from different modalities in distinctive ways. The result should be interpreted for its particular age groups, form judgments, and methods, rather than converted into a claim that young children have no multisensory interactions. Early orienting, speech, and event perception involve different demands from precision-weighted form discrimination.

Developmental change can reflect multiple components of the causal-binding system. A child may possess a strong preference for audiovisual correspondence but have highly variable hearing or vision, or may integrate effectively under one stimulus arrangement while struggling with a demanding explicit temporal-order judgment. Apparent age differences in integration therefore require appropriately matched unisensory measures, age-sensitive response formats, and reliability assessments. Stevenson et al. (2012) documents individual variability in temporal binding windows and susceptibility to illusions in an adult research setting; it provides an example of why observer-level measurements matter without directly establishing a developmental pathway. Longitudinal experiments should test whether improvements in sensory fidelity, causal priors, attention, and decision strategies follow distinct trajectories.

The same conceptual problem appears in aging research. Differences in sensory thresholds, neural timing, mobility, and confidence can alter combined perceptual responses without a corresponding change in the observer's fundamental preference to bind signals. The methodological analysis by Zhu et al. (2026a) shows how measured illusion size or window width can have different latent explanations, and it cautions against interpreting group differences as specific gains or losses of integration. The strongest lifespan designs measure modality-specific sensory reliability, use multiple stimulus disparities and task types, estimate uncertainty, and test the stability of model parameters across sessions. Whether any particular latent causal-inference parameter has a characteristic life-course trajectory remains an empirical question requiring direct measurement rather than inference from age-group means alone.

Touch, Proprioception, and Crossmodal Reference Frames

Vision and touch provide complementary estimates of an object's dimensions, surface properties, and position relative to the body. A hand exploring a visible object generates tactile and proprioceptive information in a coordinate frame different from the retinal location of its optical image. Successful combination requires calibration, reference-frame transformation, and a judgment that the observations belong to the same object. Ernst & Banks (2002) provided a landmark psychophysical demonstration of nearly optimal visual–haptic cue combination for the studied judgments. The predicted reliability-weighted estimate is a consequence of assumptions about common origin and independent measurement noise. Changing the object, posture, sensory precision, or temporal arrangement could require different weights and a revised model.

In natural action, proprioception and touch are generated through self-movement, and their timing is correlated with motor commands. This creates opportunities to predict forthcoming sensory consequences and disambiguate external events from self-produced signals. However, a behavioral advantage during reaching or grasping may reflect motor learning, efference-related prediction, ordinary attention, or information integration. These explanations make partially overlapping predictions, so experiments must manipulate visual feedback and haptic reliability separately from movement demands. Stein & Stanford (2008) identifies multisensory integration within orienting and action-relevant systems; extending its principles to complex manual behavior requires direct tests of relevant circuits and the changes in movement accuracy attributable to sensory evidence.

The scientific boundary with sensorimotor integration is therefore functional. The present dossier studies how the system infers correspondence and combines sensory evidence across modalities. Research on control policies, action selection, and sensorimotor adaptation treats the broader mapping from bodily state estimates and predicted outcomes to movement. Cross-references are scientifically productive because action actively changes which sensory observations become available, and the changing observations can revise the probability that cues share a cause. Noppeney (2021) describes multisensory perception as an adaptive inferential process that operates in real scenes. The resulting dynamic loop can be tested by comparing passive exposure, active exploratory movement, and manipulations of whether sensory signals are consistent with a single external event.

Experimental Tasks, Calibration, and Measurement

Experimental design begins with an explicit construct and the distinct observations that would support it. Localization tasks measure spatial estimates, forced-choice tasks measure discriminability, simultaneity judgments measure a response to temporal correspondence, and phonetic tasks measure speech categorization. The fact that all may be altered by combining modalities does not make their dependent variables interchangeable. Alais & Burr (2004) demonstrates how precision and bias in each modality can be measured and used to predict combined localization. McGurk & MacDonald (1976) demonstrates that audiovisual speech alters categorical content, which requires a different representational and measurement model. Each experimental paradigm should identify both a target mechanism and the most plausible nonintegrative alternative explanation.

Stimuli need physical calibration. Visual onset timing can depend on display technology and software frame synchronization; acoustic onset depends on buffering, hardware, earphones, and propagation; haptic devices introduce mechanical latencies. Psychometric functions for temporal-order judgments are particularly sensitive to these timing errors. Researchers should report physical asynchronies at the participant's sensors, stimulus dimensions, masking conditions, block order, randomization, and exclusion criteria. Vroomen & Keetels (2010) examines the importance of distinct timing processes for intersensory synchrony, while Fujisaki et al. (2004) demonstrates that the experimental history of temporal alignment changes future reports. Reproducing a temporal effect requires accounting for both physical stimulus timing and adaptive changes induced by the testing procedure itself.

A useful design samples several conflict sizes in both directions, and combines unimodal, synchronous, asynchronous, correlated, and uncorrelated conditions. The investigator can separately manipulate cue reliability and the environmental probability of common origin. Response modalities should be counterbalanced where practical, and confidence or secondary judgments can enrich the evidence without being conflated with location estimates. Körding et al. (2007) used varied localization conflicts to constrain causal models; Parise et al. (2012) used temporal correlation to study common-cause evidence. These studies illustrate how theoretically informative manipulations improve identification of the intended process. A well-designed experiment also documents the range of naturalistic contexts over which the resulting parameter estimates can reasonably be generalized.

Computational Evaluation and Reproducibility

Computational evaluation needs to establish that an apparent preference for one model is not an artifact of the modeler's assumptions. Conventional Gaussian sensory likelihoods, spatially uniform source priors, and fixed noise can simplify inference while failing to describe particular observer data. Liu, S., et al. (2026) developed flexible semiparametric approaches for examining the shapes of auditory and visual noise and prior distributions in multisensory causal inference. Their results indicated eccentricity-dependent sensory noise and nonstandard prior shapes, with findings involving audiovisual recalibration and noise changes during multisensory conditions. This primary research provides concrete reasons to inspect generative assumptions instead of treating successful fits to a canonical Bayesian specification as conclusive evidence about underlying perceptual mechanisms.

The measurement critique is explicit in Zhu et al. (2026a), which argues that illusion strength and temporal binding windows are composite outcomes of sensory precision, prior expectations, and binding processes. Model parameters therefore need recovery simulations; competing algorithms need model-recovery experiments; and fitted predictions need tests on observations that were not used during calibration. Reporting a single best-fit parameter without uncertainty obscures the possibility of practical nonidentifiability. A thorough analysis compares held-out predictive performance, response distributions, and robustness to plausible assumptions about prior and noise shape. Findings should be reported separately for each measured task before pooling is used to support a claim about a general trait of multisensory binding.

Replication also concerns variability across observers and stimuli. Wozny et al. (2010) compared model-averaging, model-selection, and probability-matching accounts, showing that the rules used to translate causal uncertainty into reports are themselves empirical hypotheses. A source replication should specify whether the goal is to reproduce a qualitative illusion, a numerical cue-weighting relation, or a complete fitted decision rule. These are different standards of success. Open stimulus-generation code, timing logs, preregistered primary contrasts, hierarchical uncertainty estimates, and accessible model specifications reduce analytical flexibility. The aim of reproducible multisensory science is to recover a stable pattern of observations and an explicit range of plausible mechanisms, with experimental tests that can distinguish the mechanisms when the available data permit it.

Contemporary Research and the Revision of Ideal-Observer Assumptions

A major direction of recent multisensory research is to test the assumptions that make classic Bayesian models tractable. The original causal-inference framework by Körding et al. (2007) explained important localization effects by considering whether signals share a source. More recent evidence asks how well the underlying likelihood and prior distributions correspond to actual human behavior. In Liu, S., et al. (2026), participants' audiovisual localization data were analyzed with flexible semiparametric models, allowing the investigators to estimate sensory noise and prior forms before constructing more restricted models. Their results suggested eccentricity-dependent noise with peripheral flattening and spatial priors with more complex shapes than standard assumptions. The substantive advance is methodological: models should earn their predictive accuracy through direct constraints on sensory representations, rather than inherit all parameters from conventional simplicity.

The question of whether causal inference appears only in explicit localization judgments has also become more precise. Friemel & Rohe (2026) investigated implicit audiovisual spatial representations and examined evidence for causal-inference processes in representations that were not simply a direct answer to an explicit common-cause question. The study extends the explanatory reach of causal inference to the structure of spatial perception, but its findings remain tied to the tasks and measures used. Converging experiments will need to establish which implicit signals are upstream of a report, which correlate with its selection, and which change causally when source assignments are manipulated. A model that predicts both neural or implicit measurements and overt behavior would be more strongly constrained than one fitted to overt localization alone.

Research has additionally broadened the latent attributes under investigation. Zhu et al. (2026b) examined flash–beep interactions across temporal and numerical relations through a Bayesian causal-inference framework, connecting event count to the question of whether audiovisual signals belong together. The outcome matters because causal binding is not confined to continuous spatial coordinates. Related theoretical work by Liu, J., et al. (2026) addresses the coupling between causal inference and crossmodal recalibration, proposing a unified framework that connects immediate source judgments with longer-term changes. These developments remain heterogeneous: experimental demonstrations constrain actual behavior, whereas a unified framework must be evaluated by its ability to predict interventions and the transfer of adaptation across new tasks.

Individual Differences and Interpretive Cautions

Individuals differ in unimodal precision, temporal discrimination, confidence, susceptibility to perceptual illusions, and strategy in ambiguous tasks. Stevenson et al. (2012) investigated relations between measured temporal binding windows and audiovisual illusion susceptibility. Such results motivate studies of stable observer-specific differences, but they do not establish one single dimension of integrative capacity. Temporal-order decisions and illusion reports can share noise sources or criteria. The appropriate measurement question is whether an individual's observed behavior can be decomposed into sensory reliability, prior expectations, causal-source inference, and downstream decision rule with sufficiently precise and repeatable estimates.

Comparisons involving neurodevelopmental conditions, sensory impairment, aging, or clinical groups require special attention to this decomposition. If one group performs differently on a McGurk or flash-count task, the result may reflect changes in the clarity of visual speech, hearing thresholds, temporal discrimination, prior learning, task comprehension, or response policy. Zhu et al. (2026a) demonstrates through computational arguments why descriptive group differences in multimodal readouts can misattribute their sources. Before a causal or clinical interpretation, investigators should report modality-specific baselines, stimulus conditions, uncertainties, and the sensitivity of fitted latent parameters to alternative models. The strength of a conclusion is measured by what has been ruled out experimentally, not by the vividness of the illusion.

A personalized psychophysical assessment could ultimately distinguish some patterns of sensory precision and causal inference, but its diagnostic and treatment relevance requires independent validity studies in the intended population. Correlation with a condition does not establish a clinical biomarker. Even a reliable group difference may have overlapping individual distributions that limit classification. Liu, S., et al. (2026) shows that inferential assumptions about sensory noise and prior form can substantially change a model's characterization of behavior. The research agenda should therefore include transparent uncertainty estimates, prospective validation, diverse and representative samples, cross-task convergence, and demonstration that the chosen measurement adds information beyond simpler sensory tests. This monograph addresses general perceptual science rather than diagnosis or individualized clinical advice.

Naturalistic Multisensory Scenes and Artificial Perceptual Systems

Real perceptual environments contain multiple candidate causes moving through time. During a conversation, several faces, voices, gestures, echoes, and background events can overlap; moving between rooms changes visual access and acoustic reflections. Successful perception requires source selection, temporally persistent grouping, sensory reliability estimation, and revisions when new evidence arrives. Noppeney (2021) presents this adaptive perspective, incorporating attention, inference, and learning within a single account of multisensory function. The challenge is to build tasks preserving sufficient naturalistic complexity for validity while maintaining controllable variables that identify a mechanism. Virtual environments and carefully synchronized video–audio recordings offer useful possibilities when their latency, spatial geometry, and stimulus quality are calibrated.

The same computational structure has practical significance for artificial sensor fusion. A robotic system combining camera images, microphone signals, haptic measurements, and motion sensors must decide which readings correspond to one environmental event and how uncertainty should affect their use. The Bayesian common-cause versus separate-causes formulation by Körding et al. (2007) provides a formal design option: represent competing assignments, estimate sensor reliabilities, and compare predicted outcomes under each assignment. Temporal correlation approaches such as Parise et al. (2012) offer another correspondence signal, while the model-assumption critique of Liu, S., et al. (2026) warns that a preselected noise family can mischaracterize sensor behavior. These are functional analogies and design hypotheses that require engineering validation under realistic system failures, distribution shifts, and ambiguous sources.

Within the Artificial Era, multisensory causal binding provides a defined area for comparisons between biological and artificial information-processing systems. Such comparisons can examine observable competencies: source association, conflict detection, adaptive weighting, uncertainty calibration, crossmodal learning, and recovery from misleading inputs. A similar output pattern does not by itself establish identical internal algorithms or subjective experience. Formal experiments can compare systems using matched stimuli and measures while retaining each system's different physical sensors, learning histories, and representations. van Atteveldt et al. (2014) and Noppeney (2021) offer conceptual frameworks for considering flexible processing operations; the scientific contribution of comparison emerges when claims are formulated as measurable hypotheses, supported by reproducible results, and restricted to the properties actually tested.

Critical Synthesis and Open Research Questions

The most durable empirical result is that perception can be systematically influenced by information arriving through another sensory modality. McGurk & MacDonald (1976) established content changes in audiovisual speech; Ernst & Banks (2002) demonstrated precision-sensitive combination of visual and haptic estimates; Shams et al. (2000) showed that sound alters reported visual event numerosity. These results support distinct forms of interaction with different target variables. The account becomes deeper when source inference is added: Körding et al. (2007) formalized the choice between common and separate causes, and Parise et al. (2012) showed that shared temporal structure supplies correspondence information. A synthesis therefore needs several linked components—encoding reliability, correspondence evidence, causal-structure uncertainty, adaptive calibration, and the decision or action used to report the result.

The major explanatory dispute concerns how much human behavior is captured by the simplest normative calculations and which mechanisms instantiate the computations. Wozny et al. (2010) demonstrated that response strategies beyond straightforward model averaging can matter. Rohe & Noppeney (2015) linked cortical representations with levels of causal inference, while Friemel & Rohe (2026) examined implicit spatial representations. Liu, S., et al. (2026) questioned assumed likelihood and prior shapes, and Zhu et al. (2026a) clarified why the same behavioral index can arise from distinct parameter combinations. These findings direct attention toward identifiability and hierarchical explanation. Agreement between a model and observed choices supports that model under the tested conditions; causal interpretation becomes stronger when manipulations independently test its components and reveal effects in multiple complementary measurements.

Several open questions organize the next research program. How does the perceptual system assign many sensory streams to several moving sources without an exhaustive combinatorial search? Which patterns of neural activity are necessary for binding, which merely reflect attention or decisions, and how can interventions disentangle them? How are sensory precision, source priors, and decision costs jointly learned across environments? Which experimental tasks recover stable and transferable person-level parameters, and which are dominated by response conventions? How does active movement improve source identification and recalibrate conflicting cues? How can comparable tests be developed for artificial sensor-fusion systems without conflating functional performance with unmeasured experiential properties? Noppeney (2021) and Zhu et al. (2026a) provide complementary grounds for posing these questions as specific programs of empirical and methodological research.

Conclusion

Multisensory integration and causal binding form a coherent psychological research object because perceptual systems must both combine uncertain observations and determine which observations can legitimately describe a shared environmental cause. The empirical foundation ranges from neural convergence and orienting responses Meredith & Stein (1986b) to quantitative psychophysical tests of cue combination Ernst & Banks (2002), audiovisual speech fusion McGurk & MacDonald (1976), temporal event illusions Shams et al. (2000), and explicit causal-inference modeling Körding et al. (2007). Together, these traditions establish that crossmodal perception depends on reliability, correspondence, temporal structure, task, and learning. Their common explanatory framework is the estimation of environmental structure from uncertain sensory evidence, with alternative causal assignments remaining possible whenever the signals could have multiple sources.

The central methodological conclusion is that a visible interaction or an illusion magnitude describes the outcome of several interacting factors. Direct inference about a binding mechanism requires separately measured unimodal precision, defensible generative assumptions, discriminating experiments, and alternative response rules. Wozny et al. (2010) illustrates how distinct decision policies can fit causal-inference tasks; Liu, S., et al. (2026) shows that the shape of sensory likelihoods and priors can depart from traditional assumptions; and Zhu et al. (2026a) identifies the interpretive risks of treating composite perceptual readouts as isolated mechanisms. Physiological results become most valuable when tied to intervention and behavior, and developmental or group comparisons become most defensible when they model the sensory baselines from which differences arise.

The field now has a precise, expandable research architecture: experimentally calibrate sensory evidence, identify source-correspondence variables, compare causal models, measure neural implementation, establish adaptive timescales, and examine generalization to naturalistic settings. The framework applies to vision, hearing, touch, action, and carefully defined problems of artificial sensor fusion, although each application requires its own empirical validation. Noppeney (2021) emphasizes the adaptive interplay of inference, attention, and learning; Rohe & Noppeney (2015) offers a neurocomputational link; and Liu, J., et al. (2026) proposes a framework for jointly studying causal inference and recalibration. The enduring result is a scientifically testable account of how multiple sensory streams contribute to coherent judgments about events and sources, together with a clear program for resolving the mechanisms that remain uncertain.

References

  • Alais, D., & Burr, D. (2004). The ventriloquist effect results from near-optimal bimodal integration. Current Biology, 14(3), 257–262. https://doi.org/10.1016/j.cub.2004.01.029
  • Colonius, H., & Diederich, A. (2004). Multisensory interaction in saccadic reaction time: A time-window-of-integration model. Journal of Cognitive Neuroscience, 16(6), 1000–1009. https://doi.org/10.1162/0898929041502733
  • Driver, J., & Noesselt, T. (2008). Multisensory interplay reveals crossmodal influences on 'sensory-specific' brain regions, neural responses, and judgments. Neuron, 57(1), 11–23. https://doi.org/10.1016/j.neuron.2007.12.013
  • Ernst, M. O., & Banks, M. S. (2002). Humans integrate visual and haptic information in a statistically optimal fashion. Nature, 415, 429–433. https://doi.org/10.1038/415429a
  • Friemel, F., & Rohe, T. (2026). Implicit causal inference in audiovisual spatial representations. Cognition, 271, Article 106413. https://doi.org/10.1016/j.cognition.2025.106413
  • Fujisaki, W., Shimojo, S., Kashino, M., & Nishida, S. (2004). Recalibration of audiovisual simultaneity. Nature Neuroscience, 7, 773–778. https://doi.org/10.1038/nn1268
  • Gori, M., Del Viva, M., Sandini, G., & Burr, D. C. (2008). Young children do not integrate visual and haptic form information. Current Biology, 18(9), 694–698. https://doi.org/10.1016/j.cub.2008.04.036
  • Körding, K. P., Beierholm, U., Ma, W. J., Quartz, S., Tenenbaum, J. B., & Shams, L. (2007). Causal inference in multisensory perception. PLOS ONE, 2(9), Article e943. https://doi.org/10.1371/journal.pone.0000943
  • Liu, J., Huang, C.-C., & Zeng, F. (2026). Coupling causal inference and cross-modal recalibration: A unified framework for adaptive multisensory perception. Biology, 15(5), Article 376. https://doi.org/10.3390/biology15050376
  • Liu, S., Holland, T., Ma, W. J., & Acerbi, L. (2026). Distilling noise characteristics and prior expectations in multisensory causal inference. PLOS Computational Biology, 22(5), Article e1014251. https://doi.org/10.1371/journal.pcbi.1014251
  • McGurk, H., & MacDonald, J. (1976). Hearing lips and seeing voices. Nature, 264, 746–748. https://doi.org/10.1038/264746a0
  • Meredith, M. A., & Stein, B. E. (1986). Spatial factors determine the activity of multisensory neurons in cat superior colliculus. Brain Research, 365(2), 350–354. https://doi.org/10.1016/0006-8993(86)91648-3
  • Meredith, M. A., & Stein, B. E. (1986). Visual, auditory, and somatosensory convergence on cells in superior colliculus results in multisensory integration. Journal of Neurophysiology, 56(3), 640–662. https://doi.org/10.1152/jn.1986.56.3.640
  • Meredith, M. A., Nemitz, J. W., & Stein, B. E. (1987). Determinants of multisensory integration in superior colliculus neurons. I. Temporal factors. The Journal of Neuroscience, 7(10), 3215–3229. https://doi.org/10.1523/JNEUROSCI.07-10-03215.1987
  • Noppeney, U. (2021). Perceptual inference, learning, and attention in a multisensory world. Annual Review of Neuroscience, 44, 449–473. https://doi.org/10.1146/annurev-neuro-100120-085519
  • Odegaard, B., Wozny, D. R., & Shams, L. (2015). Biases in visual, auditory, and audiovisual perception of space. PLOS Computational Biology, 11(12), Article e1004649. https://doi.org/10.1371/journal.pcbi.1004649
  • Parise, C. V., Spence, C., & Ernst, M. O. (2012). When correlation implies causation in multisensory integration. Current Biology, 22(1), 46–49. https://doi.org/10.1016/j.cub.2011.11.039
  • Rohe, T., & Noppeney, U. (2015). Cortical hierarchies perform Bayesian causal inference in multisensory perception. PLOS Biology, 13(2), Article e1002073. https://doi.org/10.1371/journal.pbio.1002073
  • Sekuler, R., Sekuler, A. B., & Lau, R. (1997). Sound alters visual motion perception. Nature, 385, 308. https://doi.org/10.1038/385308a0
  • Shams, L., Kamitani, Y., & Shimojo, S. (2000). What you see is what you hear. Nature, 408, 788. https://doi.org/10.1038/35048669
  • Shams, L., Ma, W. J., & Beierholm, U. (2005). Sound-induced flash illusion as an optimal percept. NeuroReport, 16(17), 1923–1927. https://doi.org/10.1097/01.wnr.0000187634.68504.bb
  • Stein, B. E., & Stanford, T. R. (2008). Multisensory integration: Current issues from the perspective of the single neuron. Nature Reviews Neuroscience, 9, 255–266. https://doi.org/10.1038/nrn2331
  • Stevenson, R. A., Zemtsov, R. K., & Wallace, M. T. (2012). Individual differences in the multisensory temporal binding window predict susceptibility to audiovisual illusions. Journal of Experimental Psychology: Human Perception and Performance, 38(6), 1517–1529. https://doi.org/10.1037/a0027339
  • van Atteveldt, N., Murray, M. M., Thut, G., & Schroeder, C. E. (2014). Multisensory integration: Flexible use of general operations. Neuron, 81(6), 1240–1253. https://doi.org/10.1016/j.neuron.2014.02.044
  • van Wassenhove, V., Grant, K. W., & Poeppel, D. (2005). Visual speech speeds up the neural processing of auditory speech. Proceedings of the National Academy of Sciences, 102(4), 1181–1186. https://doi.org/10.1073/pnas.0408949102
  • Vroomen, J., & Keetels, M. (2010). Perception of intersensory synchrony: A tutorial review. Attention, Perception, & Psychophysics, 72(4), 871–884. https://doi.org/10.3758/APP.72.4.871
  • Wozny, D. R., Beierholm, U. R., & Shams, L. (2010). Probability matching as a computational strategy used in perception. PLOS Computational Biology, 6(8), Article e1000871. https://doi.org/10.1371/journal.pcbi.1000871
  • Zhu, H., Beierholm, U., & Shams, L. (2026). Analyzing multisensory integration: Do's and don'ts. Neuroscience & Biobehavioral Reviews, 191, Article 107008. (Online publication October 2, 2026.) https://doi.org/10.1016/j.neubiorev.2026.107008
  • Zhu, H., Zhang, Y., Beierholm, U., & Shams, L. (2026). Crossmodal interaction of flashes and beeps across time and number follows Bayesian causal inference. Psychonomic Bulletin & Review, 33, Article 58. https://doi.org/10.3758/s13423-026-02857-z