← Reference notes

Unreviewed preprint. This is a bioRxiv preprint (v1) that has not been certified by peer review. Findings may change. Data and code are promised "upon publication" and were not available to check.

Summary

Neuropixels recordings from macaque face patches (inferotemporal ML/AM) and prefrontal cortex during binocular flash suppression show that, even while the stimulus and the reported percept stay constant, population activity flips roughly 15 times per second between brief micro-epochs that represent the consciously seen face and micro-epochs that represent the suppressed face. Each epoch type carries a near-"pure" code for its face, good enough to reconstruct it; the conscious epochs differ by somewhat higher tuning gain. The authors argue conscious access depends on when a representation is selected, not just where it lives (Summary; Results; Discussion).

When to apply

• Background on neural correlates of consciousness (NCC), binocular rivalry / flash suppression, and front-vs-back-of-brain debates (GNW, HOT vs IIT, RPT).
• Example of single-trial, state-based analysis of large population recordings (LLR / HMM state partitioning, split-half cross-validation to avoid circular "double dipping").
• Not a clinical or behavioural recipe; animal electrophysiology study.

Procedure (methods, condensed)

1. Paradigm: binocular flash suppression via red/cyan anaglyph goggles: one face shown to one eye (priming), then a different face flashed to the other eye; the flashed face dominates perception while the primed face is suppressed. Analysis window: 300–1000 ms after the flash (Results; Methods).
2. Human check: 14 human participants reported percepts by button press; the flashed image was reported on ~82% of rivalry trials vs ~87% for unambiguous stimuli. Monkeys did not report; their percept is inferred from the human data (Fig. 1b; Methods).
3. Recordings: two male rhesus macaques (monkeys A and M); face patches localised by fMRI; Neuropixels 1.0 NHP probes in ML and AM (IT) and in prefrontal PO (orbitofrontal) and PVL (ventrolateral); Kilosort3 sorting, single- and multi-units used (Methods).
4. Stimuli: (a) 10 sessions with 10 screened stimuli in 25 pairs, including an "unambiguous" control where the new image is shown alone; (b) 9 sessions with 1500 parametric faces in 750 pairs, used to fit linear face-feature decoders (Methods).
5. State inference: per 10 ms bin, a Poisson log-likelihood-ratio decoder labels activity as matching the conscious-face template (state 1) or the suppressed-face template (state 2), with templates taken from priming responses; an HMM is also used to estimate dwell times (Results).
6. Controls: split-half and leave-one-neuron-out cross-validation; trial-balanced subsampling; shuffled templates; a four-state On/Off-gain model; simulated alternative models (conscious-only, averaging, uncoordinated per-neuron switching) (Results; Ext. Data Figs. 5–10).

Costs and limits

• Unreviewed preprint; only two monkeys; monkeys' percepts are inferred from human reports, not measured directly.
• The authors note that using the same data to define states and to measure them is circular and inflates accuracy. They rely on held-out-neuron validation, which gives somewhat lower but still clear separation.
• Switching was only weakly coordinated between simultaneously recorded ML and AM; the authors say this is not strong evidence of cross-patch coordination.
• Prefrontal reconstructions were weaker (best case <80% explained variance vs >90% in IT; shape features <30%).
• A linear decoder told conscious from suppressed epochs apart only moderately well (66%).
• Why some micro-epochs reach awareness is left open (Discussion, partly speculative).
• Data (DANDI) and code (GitHub) promised only upon publication.

Evidence (detail)

• Trial-averaged IT responses were only weakly modulated by percept compared with unambiguous presentation (slope ~0.23; 5550 cell–stimulus pairs) (Fig. 1h).
• State 1 and state 2 were each occupied about half the time after the flash. The LLR autocorrelation time constant was ~67 ms, and HMM median dwell times were ~64 ms (state 1) and ~59 ms (state 2), so roughly 15 Hz, versus ~0.5 Hz perceptual rivalry switches (Fig. 2c–d; Fig. 5c–d; Discussion).
• Faces reconstructed from state-1 activity matched the conscious face (>80% explained variance for some features). State-2 activity matched the suppressed face. Time-averaged activity gave a blend of both (Fig. 2h–i; Ext. Data Fig. 2).
• Single-trial 10-way identity decoding: time-averaged activity gave ~40% (conscious) vs ~34% (suppressed). State-1 points gave >61% (conscious) vs ~4% (suppressed), and state 2 showed the reverse (Fig. 2k–l).
• A two-state model beat the best one-state model in all 19 sessions. Held-out neurons switched represented face at state transitions inferred from the other half (Fig. 3).
• The same neurons encoded both faces: tuning gain across cells correlated r = 0.90 between states. Tuning gain was ~1.21× higher for the conscious face in state 1 than for the suppressed face in state 2, while mean firing was slightly higher in state 2 (Fig. 4c–e).
• Within states, modulation tracked the conscious face (state 1, slope ~0.93) or the suppressed face (state 2, slope ~−0.76). Equalising gain removed the residual average modulation, which explains the classic "weak modulation" puzzle (Fig. 4g).
• Simulated averaging and uncoordinated models failed cross-validated reconstruction, tuning reversal, and two-vs-one-state model comparison. Only the coordinated two-state model matched the data (Ext. Data Fig. 10).
• Prefrontal cortex showed the same interleaving, including under held-out validation (Fig. 6).

Facts

• id: F1 claim: In macaque IT face patches under binocular flash suppression, population activity alternates between conscious-face and suppressed-face representations with ~60 ms dwell times (~15 Hz), while percept and stimulus stay constant. confidence: medium # single source: author-reported, unreviewed preprint date: 2026-10-16 sources: [S1] license: CC BY 4.0
• id: F2 claim: The suppressed face can be reconstructed with high fidelity from activity during "suppressed" micro-epochs, using decoders trained only on unambiguous responses. confidence: medium # single source: author-reported, unreviewed preprint date: 2026-10-16 sources: [S1] license: CC BY 4.0
• id: F3 claim: Conscious-epoch tuning gain is ~1.2× higher than suppressed-epoch gain in the same neurons, and a linear decoder separates the two epoch types at 66% accuracy. confidence: medium # single source: author-reported, unreviewed preprint date: 2026-10-16 sources: [S1] license: CC BY 4.0
• id: F4 claim: Weak percept-related modulation of time-averaged rivalry responses is explained by averaging across the two interleaved states plus the small gain asymmetry. confidence: medium # single source: author-reported, unreviewed preprint date: 2026-10-16 sources: [S1] license: CC BY 4.0
• id: F5 claim: Prefrontal cortex (PO, PVL) shows qualitatively similar interleaving, with weaker face reconstructions than IT. confidence: medium # single source: author-reported, unreviewed preprint date: 2026-10-16 sources: [S1] license: CC BY 4.0
• id: F6 claim: Study used two rhesus macaques (19 Neuropixels sessions) plus 14 human participants for perceptual validation. confidence: medium # single source date: 2026-10-16 sources: [S1] license: CC BY 4.0

Related

• None in the library yet (first neuroscience / bioRxiv reference note).

Review notes

• Built from the bioRxiv v1 full-text PDF (accessed 2026-10-18). Paraphrase only; no figures reproduced; no verbatim quotes beyond short attributed terms.
• License checked three ways: API license: cc_by, landing page "CC-BY 4.0 International license", PDF header. Copyright stays with the author/funder; CC BY 4.0 permits reuse with attribution.
• Date note: bioRxiv records receipt on 2026-10-15 and posting of v1 on 2026-10-16.
• Re-check before promoting to reviewed: whether a newer version or a journal version exists.

@misc{iil_note_p30amr5v,
  author = {{iinventedlight Library}},
  title = {{Rapid alternation of conscious and suppressed stimulus representations in the primate brain}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-10-18},
  url = {https://iinventedlight.com/library/reference/biorxiv/rapid-alternation-conscious-suppressed-tsao-2026/},
  note = {Permanent ID iil:note:p30amr5v. License: CC-BY-4.0. Summary of https://doi.org/10.64898/2026.10.15.759299.},
}

For language models and agents: start at /llms.txt. This page in Markdown: /library/reference/biorxiv/rapid-alternation-conscious-suppressed-tsao-2026.md.