KurisutinaCitation: Sergent C, Corazzol M, Labouret G, Stockart F, Wexler M, King JR, Meyniel F, Pressnitzer D. Bifurcation in brain dynamics reveals a signature of conscious processing independent of report. Nature Communications 12, 1149. Published 19 February 2021. DOI: 10.1038/s41467-021-21393-z. Published version; PMCID PMC7895979; PMID 33608533.
Reading completed 29 September 2026: the entire 19-page published paper, including all Methods, equations 1–9, five figures and captions, references and declarations; all 11 pages of Supplementary Information, including seven figures and complete captions; and both pages of the reporting summary. All figures, equation displays and reporting-form pages were visually inspected. No main or supplementary tables. The complete scientific package was acquired before substantive reading. The separately linked peer-review file, source-data spreadsheet and OSF data/code repository were identified but not read or audited. This is a critical reading, not a reanalysis or replication. Sources, hashes and scope are in papers/consciousness_connectomics/sergent2021_bifurcation.provenance.json.
Main conclusion
In awake adults hearing faint sounds, late EEG activity around 250–700 ms shows increased trial-to-trial variability near perceptual threshold. In the active auditory task, an explicit two-component mixture model explains these late distributions better than the tested unimodal alternatives and its inferred high/low states discriminate subsequent audibility reports. Related late dynamics remain when there is no required identification or audibility report about the sounds, despite the loss of the familiar central P300-like positivity. In that condition, the model predicts whether participants subsequently report thinking about sounds rather than something else.
This is evidence for a late processing signature that survives removal of a particular auditory task. It does not establish that the passive condition is free of tasks, reports, attention, memory or self-monitoring. Nor does fitting a mixture to scalp EEG establish a unique underlying dynamical mechanism, the necessity of frontal cortex, or a validated consciousness test for noncommunicating patients.
What was manipulated and what counted as consciousness
Twenty-five native French speakers entered the main experiment; two stopped and three met a predefined exclusion criterion of more than 25% artifact-contaminated EEG trials. The analyzed sample was 20 right-handed adults, ten women, mean age 23.4 years (range 21–29). Each completed active and passive sessions on different days, with order counterbalanced, ten in each order. The reporting summary says order assignment was randomized. Sample size followed EEG recommendations and previous studies; a prospective numerical power analysis is not reported. Most recruits were higher-education students.
Stimuli were 200-ms synthesized French vowels in continuous threshold-equalizing noise. Active stimulation levels were −13, −11, −9, −7 and −5 dB signal-to-noise ratio (SNR), plus noise alone. The planned session comprised 960 trials, 160 per stimulation level, across 20 blocks, lasting about 2.5–3 hours with breaks. The last ten participants additionally received −3 dB trials in the passive session. Thus even the physical stimulus range was not completely invariant across all participants and conditions.
In the active session participants identified the vowel and rated audibility on an 11-position scale, 0–10, displayed as a slider. The instructions distinguished nothing heard, doubtful presence and certain presence. The paper calls this a continuous rating approach, but its recorded response options were discrete and semantically anchored. Ratings of at least 30% were classified as “heard.” Results motivate this cutoff by a trough in the near-threshold bimodal distributions; Methods call it the median audibility at −9 dB. It is a criterion derived from these behavioral data, not an independently established ground truth about experience.
In passive sessions there was no identification or audibility decision on each sound. Instead, after stimulation, one of four tasks appeared randomly: a speeded click to a visual circle; a question about current thought content; a general-knowledge/arithmetic quiz; or a click-to-continue instruction. Thought probes offered sound, task, own thoughts, or nothing/sleepiness. Block-end questionnaires explicitly asked about hearing/recognizing vowels, attention and thought contents. Calling this a completely report-free or task-free experiment would therefore be inaccurate.
The response screen appeared 2–3 seconds after the sound; the principal 250–700 ms EEG interval precedes that screen. This reduces contamination by the actual response movement, but does not eliminate anticipation, covert decisions or memory formation. In passive sessions, choosing “sound” indicates sound-related current mental content. Choosing another option does not establish that the preceding sound was never consciously heard. Participants could hear it and later think of something else. Supplementary Fig. 5 also shows sound responses on noise-only trials, reinforcing that this is a thought-content measure rather than an exact vowel-detection label.
All participants were awake healthy adults. “Conscious/unconscious” here primarily means reported access to an individual auditory event, not experimentally induced global states of consciousness or unconsciousness.
Analysis and the meaning of bifurcation
The important methodological contribution is to compare distributions across trials rather than infer bifurcation from a nonlinear average response. A nonlinear but unimodal response can have a sigmoid mean. A mixture of a low state and a high state can produce a similar mean, but predicts increased variability near the intensity where the states are mixed.
EEG was recorded from 64 electrodes at 500 Hz. Main preprocessing included high-pass filtering, notch filtering, ICA and artifact rejection; about 94 ± 4% of trials remained. The authors analyzed evoked potentials and multivariate patterns. For the latter, regularized logistic classifiers learned highest-SNR stimulus-present versus stimulus-absent trials, using ten-fold validation with training and testing in different blocks. The classifier generalized to other SNRs and time points. The signed distance from its decision plane supplied the trial-level “projected activity.” Awareness reports were not its training labels, but this first stage was supervised by physical stimulus labels; the whole pipeline is not an entirely unsupervised discovery procedure.
For formal model fitting, projected activity was low-pass filtered at 10 Hz and averaged into 30-ms windows. Maximum-likelihood fits, independently for each participant and window, compared a two-parameter constant Gaussian null, a six-parameter nonlinear unimodal Gaussian and a seven-parameter mixture. An illustrative linear unimodal model was omitted from the formal comparison as a special case of the nonlinear model. Parameters were optimized with Nelder–Mead. Five-fold held-out log likelihood, again with blocks separated, entered group Bayesian model selection and protected exceedance probability (PEP).
The mixture has the form
[ p(x\mid s)=\beta(s),\mathcal N(x;\mu_{high}(s),\sigma^2) +(1-\beta(s)),\mathcal N(x;\mu_{noise},\sigma^2). ]
Here the mixing probability and high-state mean are logistic functions of SNR; the low-state mean is independent of SNR, and the two components share a standard deviation. The nonlinear unimodal alternative allows a logistic mean and a standard deviation linear in that mean. Although the opening prose description of the mixture's high-state mean calls it linear, equation 9 and its subsequent explanation specify a logistic function. The equations are the clearest definition of the tested model.
A preference for this mixture establishes a better description within these candidates. The study does not fit an explicit recurrent dynamical system, show attractor stability or hysteresis, identify circuit parameters, or exclude every possible continuous distribution with changing shape. The proposed neural “bifurcation” is therefore a mechanistic interpretation supported by these statistical signatures, not a uniquely identified dynamical mechanism. PEP is likewise a relative group model-selection quantity, not the probability that a theory of consciousness is true.
Principal results
Behavioral identification and audibility rose around −9 to −7 dB. Near threshold, audibility distributions were bimodal and variability peaked; Supplementary Fig. 1 gives all twenty participants' mean and variability curves. Mean EEG amplitudes were already nonlinear at early latencies, which by themselves could not separate the candidate models. The distinctive variability increase emerged later, around 250–300 ms.
Temporal generalization showed early patterns around 100 ms generalizing for about 50 ms, followed after about 250 ms by patterns extending across roughly 200-ms intervals. Neural and behavioral variability profiles first correlated significantly in the 200–250 ms window, t(19)=3.06, FDR-corrected p=.010, Cohen's d=.68; later tested windows beyond 250 ms had t>4.10, p<.01 and d>.9. These are associations between intensity-dependent profiles, not evidence that a particular local signal causes awareness.
Splitting trials by report, the 300–400 ms projected response increased with SNR on heard trials but remained near baseline on not-heard trials. The SNR-by-report interaction was F(1,19.4)=7.86, p=.011. The mixture's formal group preference was concentrated between about 250 and 700 ms, with PEP above 95% across most of this interval; earlier activity favored the nonlinear unimodal model. After about 700 ms the unimodal account again became favored. This formal comparison is displayed for active sessions in Fig. 3E; an equivalent passive-session PEP comparison is not shown in the main paper or SI.
The model's high/low likelihood ratio, fitted without awareness labels, discriminated heard from not-heard reports with peak AUC above .75 (approximately .75–.77 visually) in Fig. 3F. That curve combines intermediate SNRs −11, −9 and −7 dB. AUC is a ranking measure, not 75% classification accuracy, and this plot is not a separate performance estimate at every fixed SNR. The classifier feature extraction and formal model comparison are cross-validated; the final behavioral-prediction section does not separately specify that the mixture parameters for this association were fitted out of fold. It should not be described as an externally validated or fully nested clinical decoder.
In passive sessions, the broad central positivity associated with P300 largely disappeared. A late response nevertheless persisted with a different scalp distribution, including bilateral temporal positivity and central negativity. Cross-condition decoding was asymmetric: patterns learned from active trials captured much of the passive response, whereas patterns learned from passive trials did not capture all of the active response. This supports overlapping information with an additional task-associated component. It does not uniquely prove a strict subset relationship between neural circuits.
The passive variability peak occurred at higher stimulus intensities. For the plotted correlation with active-session audibility, the authors selected a shift of one or two SNR levels separately per participant to maximize neural–behavioral profile correspondence: nine used one level and eleven two. Methods calls these shifts one or two “dB,” but its examples and Fig. 5 caption indicate steps between SNR levels, which were spaced by 2 dB. This optimized alignment is descriptive support, not independent validation of a threshold shift.
Participants selected sound on 19% of thought probes overall, increasingly at higher SNR. Passive neural likelihood ratios predicted sound-versus-other reports above chance during much of the late interval, with the Fig. 5C curve visually around AUC .63–.65 near its peak. All SNRs were pooled for this analysis, and sound-report frequency itself depended on SNR. The reported performance therefore does not alone demonstrate equal-stimulus discrimination independent of intensity, attention or subsequent memory. It is still useful that awareness labels were withheld from fitting and then used for a behavioral check.
Localization, controls and supplementary evidence
Source estimation used a standard ICBM152 anatomy, a 15,002-vertex cortical mesh and sLORETA minimum-norm reconstruction, not each participant's individual MRI. At roughly 330 ms, both conditions showed inferred temporal and inferior prefrontal sources, with additional premotor/motor and supplementary motor activation during the active task. The authors explicitly acknowledge that the inferior prefrontal source could be leakage from temporal activity and that a more locally temporal account remains possible. The paper consequently cannot settle whether frontal activity is required for report-independent conscious access.
Supplementary Figs. 2–3 inspect activity distributions and intensity profiles conditional on report; their disappearance of the overall variability peak after splitting by report is consistent with a mixture explanation. These analyses also condition on a behavioral cutoff and should not be treated as independent proof of two ontological experience states. Supplementary Fig. 4 provides passive scalp and mean/variability profiles; Fig. 5 provides every thought-report category against SNR.
Supplementary Fig. 6 compares passive-first and passive-second groups of ten each. No difference survived correction, and neither evoked-potential nor prediction curves differed significantly. This weakens a simple carryover-from-active-task account but is not an equivalence test and cannot eliminate all covert monitoring. A passive-only control recruited ten new participants, analyzing seven after excluding three with poor EEG and absent auditory evoked responses even at highest SNR. It again showed a late response and above-chance thought-content prediction during a shorter interval. Excluding nonresponsive EEG recordings is particularly relevant to any proposed future use in nonresponding patients.
Supplementary Fig. 7 presents a five-person pure-tone control: 1,000 and 2,236.1 Hz, 200 ms, with a different SNR range. Four participants performed passive first and one active first. Late response differences and increased intermediate-intensity variability appeared for tones too. Its Fig. 7D uses interquartile range rather than standard deviation. These small controls broaden the evidence within the authors' study; they are not independent reproductions by this research project. The reporting summary's description of ten new passive-control participants refers to recruitment; Methods and Fig. S6 show only seven analyzed.
The reporting form documents randomization of trial conditions and session order, with participants and experimenters aware of active/passive instructions. Its older data-availability wording promises future deposition; the final article provides an OSF repository and source-data file. The repository's contents were not audited here.
Relation to the consciousness/connectomics research question
The authors interpret these results as supporting a late ignition-style prediction of global neuronal workspace accounts while separating conscious access from executive and response processes. Their proposed “global playground” is a conscious-processing network smaller than the task-engaged workspace. This is their theoretical interpretation. The experiment does not directly compare formal IIT and GNWT models, demonstrate global broadcast, or establish the anatomical extent of the relevant network.
Relative to Frässle et al., both studies show that removing required perceptual reports changes familiar neural correlates. Their validation differs: Frässle calibrates ocular proxies against rivalry reports before using them in passive viewing; Sergent fits a stimulus-conditioned neural distribution without awareness labels, then checks it against active ratings and intermittent passive thought reports. Neither route makes unreported experience directly observable. Sergent also makes the positive point that loss of a conventional P300-shaped scalp response does not entail loss of all late processing.
Relative to Cogitate, the relevant distinction is what participants must attend to and report. Cogitate's task-irrelevant trials still occur within an ongoing visual target task; Sergent's auditory stimuli have no identification task during passive sessions, but unrelated tasks and thought probes remain. These designs are different ways of reducing particular report demands, not interchangeable tests of consciousness in the complete absence of task context. Sergent concerns near-threshold auditory access; Cogitate concerns representations of visible visual contents and theory-specific temporal/connectivity predictions.
For connectomics, this is evidence about measured dynamics and their relation to reports under changing task demands. It supplies no connectome and no structural sufficiency test. An emulation benchmark motivated by the paper would need to reproduce conditional distributions, time courses and task effects, with validated behavioral relationships; matching mean activity or possessing an anatomical wiring diagram would not establish that those further constraints are met. That is an inference for research design, not an empirical demonstration of emulated consciousness.