The Science of Consciousness 2026 · San Diego, CA Presented October 13, 2026 · Poster #530 · Session 1

The Internal Dynamics of Decisions: From rats and primates to AI

Nathan M. Thornhill · ORCID 0009-0009-3161-528X
Institute for Complexity Science and Advanced Computing (ICSAC), Fort Wayne, Indiana
Brain or AI model? Round 1 of 8
A
B

Which one is the brain?

Schematic phase portraits, drawn in neutral until you answer. What is measured, and what sets how far each one wanders, is the share of the next state that arrives from outside rather than from the population itself.

The paperRead the preprintThe October 2026 preprint, 26 pages with the figures. 1560 KB, opens in a new tab.

The finding, in brief

Do two systems that behave alike reach a decision the same way? Behavior can’t say: a feedforward network can copy any recurrent brain’s behavior. Six networks wired two ways, trained to one score, still differ in their dynamics.

How much of each next moment arrives from outside, rather than being generated by the system itself: the chart from the middle of the poster (Figure 2 in the paper), plotted as first analyzed, at unmatched input dimension.
The axis
Different regimes
Brains run on themselves (autonomy). AI models run on what you feed them (input). Rat and monkey were recorded in different brain areas, doing unrelated jobs. Re-run at a matched input dimension (rat OFC, two Qwen models), the sides held. But the AI models’ label depends on how their input is counted, so the distances don’t compare.
Brain
75 of 115
Sessions
The rat decides. Then its cortex holds on longer.
Luo et al. 2025, 115 sessions from 12 rats, about 36,000 frontal and striatal units. The rise is positive in 10 of 12 rats by rat mean; a rat-level model gives p ≈ 0.1. Recovered, not discovered here. These are separate recordings from the rat cortex you play against, which is Schiereck et al. 2026.
AI model
0 of 54
Token-level tests
In the two models probed, the answer surfaces a few layers before the output.
Where it shows up at all under the readout used, the answer first becomes top-ranked around layer 25 (of 28 in Qwen, of 32 in Llama). The layers after it are feedforward stages, not a span in which a recurrent state could persist, so the test ran across generated tokens instead: two pre-registered passes, 30 tests in five AI models, then 24 in four. None found a significant commitment-locked rise. The estimator is noisy on short sequences, so a small effect is not ruled out.

How. A subspace state-space demix splits each next state into input drive and what the population generates itself, validated on synthetic systems with known answers before any real data. The six architectures were matched on size as well as score, and none of the 144 networks was dropped, so the signature tracks architecture, not task; by architecture, 3 of 3 on the delayed task. The estimator is N4SID (Van Overschee & De Moor 1994) and the input-aware demix is InputDSA (Huang et al. 2026), on Ostrow et al. 2023.

What this does not show. No claim about phenomenal experience or machine sentience. What is measured is decision commitment, which is not conscious access. The heaviest limitation is that cortex and the models differ in task and demand as well as wiring. The demix also cannot certify that cortical autonomy is recurrent rather than functionally feedforward (Goldman 2009), so the architecture claim rests on the six-network control, where the wiring is known.

Bio

Nathan M. Thornhill

Nathan M. Thornhill is a complexity science researcher in Fort Wayne, Indiana, working in information theory and pattern persistence. He is self-taught, and his earlier career was in healthcare: nursing assistant, then nursing-home administration, then ICU admissions.

What he studies is whether a system holds together and how you would measure that: pattern persistence in simple discrete systems, sleep stages from EEG, task success read off a model’s internal geometry. The analysis plans get written down and frozen first, and the measures that fail are reported next to the ones that work. The results so far are out as preprints and one book. He is a student member of the Association for the Scientific Study of Consciousness (ASSC).

When he isn’t researching he’s playing guitar, gardening, or with his wife and daughter.

How this work gets madeWho pays for it

No university support and no grant funding. A web design business pays for the research, ICSAC publishes it, and a tool the research needed has since turned into a product that brings in money of its own.

3Rivers WebTech 3riverswebtech.com

Websites, point of sale, and the ordinary technology work small businesses in Fort Wayne need. Pay once, own it forever, no monthly subscriptions. Client work buys the time this research takes.

ICSAC icsacinstitute.org

The Institute for Complexity Science and Advanced Computing, which publishes this work. Submissions go to an AI review panel with citation verification and double-blind preprocessing, and a curator checks every recommendation. The review stack is open source, so you can read the rubrics that evaluated your paper.

CiteStamp citestamp.com

Open any paper and walk its citations as an interactive live map. Every edge carries a type, so a paper can support yours, refute a result, extend a method or use your data. Millions of edges span indie preprints, Zenodo deposits and the journal literature. Human-signed claims and machine-inferred edges stay separate. Your AI assistant can query it without a key.

Follow the workElsewhere

Bluesky LinkedIn Newsletter Code ORCID

Data and pre-registrations

Every recording analyzed here is public and was collected by other laboratories. Nothing on this poster required new animal work.

  1. Luo TZ et al. 2025. Data from: Transitions in dynamical regime and neural mode underlie perceptual decision-making. Dryad. doi:10.5061/dryad.sj3tx96dm
  2. Neupane S, Fiete I, Jazayeri M. 2024. Data from: Mental navigation in the primate entorhinal cortex. DANDI Archive, dandiset 000897. dandiarchive.org/dandiset/000897
  3. Schiereck SS et al. 2025. Data from: The orbitofrontal cortex updates beliefs for state inference. Zenodo. doi:10.5281/zenodo.16997337

Pre-registrations. Both analysis plans were written down and frozen in advance, the six-network control before a single dynamical measure was computed and the token-level test before any of its results, and both accompany the October 2026 preprint as supplementary material: doi.org/10.20944/preprints202608.1095.v2. The measure that separates the six networks and the one that does not were both named in advance, and both are reported.

References

The sources the poster draws on. The preprint carries the full text and both pre-registrations.

Show allHide 20 references
  1. Belrose N, Ostrovsky I, McKinney L, Furman Z, Smith L, Halawi D, Biderman S, Steinhardt J. 2023. Eliciting latent predictions from transformers with the tuned lens. arXiv:2303.08112
  2. Block N. 1995. On a confusion about a function of consciousness. Behavioral and Brain Sciences 18(2):227–247. doi:10.1017/S0140525X00038188
  3. Doerig A, Schurger A, Hess K, Herzog MH. 2019. The unfolding argument: why IIT and other causal structure theories cannot explain consciousness. Consciousness and Cognition 72:49–59. doi:10.1016/j.concog.2019.04.002
  4. Doerig A, Schurger A, Herzog MH. 2021. Hard criteria for empirical theories of consciousness. Cognitive Neuroscience 12(2):41–62. doi:10.1080/17588928.2020.1772214
  5. Goldman MS. 2009. Memory without feedback in a neural network. Neuron 61(4):621–634. doi:10.1016/j.neuron.2008.12.012
  6. Huang A, Ostrow M, Singh SH, Kozachkov L, Fiete I, Rajan K. 2026. InputDSA: demixing then comparing recurrent and externally driven dynamics. ICLR 2026. arXiv:2510.25943
  7. Kleiner J, Hoel E. 2021. Falsification and consciousness. Neuroscience of Consciousness 2021(1):niab001. doi:10.1093/nc/niab001
  8. Lamme VAF. 2006. Towards a true neural stance on consciousness. Trends in Cognitive Sciences 10(11):494–501. doi:10.1016/j.tics.2006.09.001
  9. Luo TZ, Kim TD, Gupta D, Bondy AG, Kopec CD, Elliott VA, DePasquale B, Brody CD. 2025. Transitions in dynamical regime and neural mode during perceptual decisions. Nature 646(8087):1156–1166. doi:10.1038/s41586-025-09528-4
  10. [dataset] Luo TZ et al. 2025. Data from: Transitions in dynamical regime and neural mode underlie perceptual decision-making. Dryad. doi:10.5061/dryad.sj3tx96dm
  11. Mante V, Sussillo D, Shenoy KV, Newsome WT. 2013. Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature 503(7474):78–84. doi:10.1038/nature12742
  12. Mashour GA, Roelfsema P, Changeux J-P, Dehaene S. 2020. Conscious processing and the global neuronal workspace hypothesis. Neuron 105(5):776–798. doi:10.1016/j.neuron.2020.01.026
  13. Neupane S, Fiete I, Jazayeri M. 2024. Mental navigation in the primate entorhinal cortex. Nature 630(8017):704–711. doi:10.1038/s41586-024-07557-z
  14. [dataset] Neupane S, Fiete I, Jazayeri M. 2024. Data from: Mental navigation in the primate entorhinal cortex. DANDI Archive, dandiset 000897. dandiarchive.org/dandiset/000897
  15. Ostrow M, Eisen A, Kozachkov L, Fiete I. 2023. Beyond geometry: comparing the temporal structure of computation in neural circuits with dynamical similarity analysis. Advances in Neural Information Processing Systems 36. arXiv:2306.10168
  16. Schiereck SS, Pérez-Rivera DT, Mah A, DeMaegd ML, Hocker D, Ward RM, Savin C, Constantinople CM. 2026. The orbitofrontal cortex updates beliefs for state inference. Neuron 114(3):507–520.e8. doi:10.1016/j.neuron.2025.10.024
  17. [dataset] Schiereck SS et al. 2025. Data from: The orbitofrontal cortex updates beliefs for state inference. Zenodo. doi:10.5281/zenodo.16997337
  18. Soldado-Magraner J, Mante V, Sahani M. 2024. Inferring context-dependent computations through linear approximations of prefrontal cortex dynamics. Science Advances 10(51):eadl4743. doi:10.1126/sciadv.adl4743
  19. Tononi G, Boly M, Massimini M, Koch C. 2016. Integrated information theory: from consciousness to its physical substrate. Nature Reviews Neuroscience 17(7):450–461. doi:10.1038/nrn.2016.44
  20. Van Overschee P, De Moor B. 1994. N4SID: subspace algorithms for the identification of combined deterministic–stochastic systems. Automatica 30(1):75–93. doi:10.1016/0005-1098(94)90230-5