Vassiki Chauhan, Krystal McCook, Mariam Latif, Alex White
4 min
How does the human brain process multiple written words simultaneously? While previous research has established that the brain processes single words in specific regions, it remains unclear whether these regions can handle multiple words at once or if they are limited by a serial bottleneck. This study investigates whether the 'simultaneous suppression' effect—a phenomenon where neural responses to multiple stimuli are weaker when presented simultaneously versus sequentially—applies to written words.
The researchers used fMRI to measure BOLD responses while participants viewed rapid sequences of character strings. To isolate the effect of word count from visual stimulation, they kept the total number of visual elements constant across conditions (zero, one, or two words, with the remainder being illegible 'false font' strings). They analyzed activity in text-selective regions of the ventral temporal cortex (such as the VWFA) and core language regions in the frontal and temporal lobes. They also examined whether the lexical frequency of the words—a proxy for the difficulty of word recognition—interacted with the presentation format.
The study found that BOLD responses in reading-related brain regions increased linearly with the number of words presented, regardless of whether they were presented sequentially or simultaneously. Contrary to the researchers' initial hypothesis, there was no evidence of simultaneous suppression in these regions. However, behavioral performance was significantly worse for simultaneous word presentation, and the sensitivity of ventral temporal regions to word frequency was attenuated in the simultaneous condition. This suggests that while the brain can detect multiple letter strings in parallel, the subsequent process of lexical access—identifying the words—is subject to interference or a serial bottleneck.
This research provides a mechanistic insight into the reading network, suggesting a hybrid model of processing. It appears that the brain performs parallel sublexical processing (detecting letter strings) but relies on a serial process for lexical identification. This distinction helps reconcile conflicting theories about whether reading is a purely serial or parallel process, indicating that the answer depends on the stage of processing being measured.
We have learned much about the brain regions that support reading by measuring neuronal responses to single words, but we know little about how the brain processes multiple words simultaneously. This fMRI study fills that gap by varying the number of English words presented while holding the amount of visual stimulation constant. We adapted the “simultaneous suppression” paradigm, which has demonstrated that the response to multiple stimuli presented simultaneously is typically smaller than the sum of responses to the same stimuli presented sequentially. On each trial, participants viewed a rapid sequence of three frames. Each frame contained two character strings, most of which were pseudoletters with visual features matched to familiar letters. The experimental conditions differed in the number of words in the sequence: zero words, one word, two words sequentially, or two words simultaneously. Behavioral task accuracy was worse for detecting two words presented simultaneously than sequentially. BOLD responses increased linearly with the number of words presented in several reading-related regions of the left hemisphere: text-selective occipito-temporal regions, the STS, the intraparietal sulcus, and the inferior frontal sulcus. In all of those regions, responses did not differ between sequential and simultaneous presentation of two words. Nonetheless, the sensitivity of ventral temporal text-selective regions to the lexical frequencies of two words was attenuated by simultaneous presentation. To account for these patterns of activity and task performance, we suggest that the reading network can detect two strings of letters simultaneously, but there is interference during lexical access.
Sam: [confident, direct] They used high-precision eye-tracking. Fixation breaks were rare and saccades were essentially absent. And the behavioral data corroborated the neural picture—accuracy was lower on simultaneous trials, which is exactly what you'd predict if lexical access is the bottleneck, not visual detection.
Alex: [thoughtful] So it's a functional constraint on meaning retrieval, not a sensory constraint on vision. Did the bottleneck localize specifically to the ventral temporal pathway, or did they see it in frontal regions too?
Sam: [measured, guiding] The linear summation—the parallel detection signature—was consistent across text-selective regions in the occipito-temporal cortex, the superior temporal sulcus, and the inferior frontal sulcus. But the lexical frequency attenuation was specific to the ventral temporal pathway. Frontal regions showed sensitivity to word count but not the same dampening of the frequency effect. That localizes the bottleneck to the transition between visual detection and lexical identification, which aligns with the ventral temporal cortex's role as the orthographic lexicon.
Alex: [reflective] So the architecture is genuinely distributed—parallel input processing across a broad network—but the rate-limiting step is localized. That has real implications for how reading models are built. Most serial models treat the whole system as a queue. [[RP_SECTION:implications-for-reading-models|Implications for Reading Models]]
Sam: [calm, concluding] And that's the key takeaway. The evidence here argues for decoupling visual detection from lexical identification in reading models. The front-end is parallel; the back-end is serial. The bottleneck isn't a design flaw—it's probably a trade-off. Running full lexical retrieval in parallel across multiple words simultaneously would be computationally expensive, and the system appears to have resolved that by serializing at exactly the stage where retrieval demand is highest. What the paper leaves open is whether that serialization is strict—one word at a time—or whether there's some partial overlap, and whether the degree of bottleneck varies with reading skill or task demands. Those are the natural follow-on questions.
Alex: [grounded] A parallel front-end feeding a serial back-end, with the bottleneck localized to the lexical access stage. That's a cleaner decomposition than most reading models assume. Thanks for listening to ResearchPod.