Systematic evidence mapping across consciousness phenomena
This is a consciousness research analysis system that uses GPT-5 to systematically extract evidence from scientific papers about consciousness-related phenomena in both biological and artificial intelligence systems. The system analyzes PDFs against a framework of nine core consciousness phenomena (like information integration, state transitions, and temporal coordination), extracting specific evidence items with detailed citations including quotes, figures, and tables from each paper. The extracted data is then visualized through a Flask web application that provides an interactive interface to browse papers, filter evidence by various criteria, view detailed analyses, and access synthesized insights across phenomena—essentially creating a structured knowledge base for comparing consciousness mechanisms across biological brains and AI models.
Token-aligned EEG recordings paired with audio/text embeddings enable analysis of how linguistic information is represented across brain and model spaces.
Temporal CoordinationBIO
Hardware and character-level triggers provide precise timing of linguistic units; preprocessing targets oscillatory bands implicated in coordination of cognitive processing.
Information IntegrationBIO
Cross-task correlations between language regions indicate integrated processing of semantic information across modalities.
Aligned EEG and audio latents with a CLIP loss to form a unified cross-modal representation usable for zero-shot classification and speech reconstruction.
Temporal CoordinationBIO
Temporal variance of EEG latent features in 100 ms windows enables detection of syllable onsets and segmentation of speech periods without explicit training.
Selective RoutingAI
Training with EMG-mixed augmentation suppresses reliance on EMG, reducing EMG-only decoding to chance while preserving EEG-based decoding.
Representational StructureAI
Cosine-similarity top-k evaluation and an HTNet+Conformer encoder indicate an embedding geometry where EEG and audio latents align in a structured space.
Low-intensity tFUS targeted to the right hippocampus increased hippocampal glucose metabolism and improved recognition memory without detectable BBB opening.
Temporal CoordinationBIO
Preclinical work cited by the authors reports that tFUS alone stimulated neuronal activity and synchronous oscillations in mouse hippocampus.
Heleen A. Slagter, Antoine Lutz, Lawrence L. Greischar, Andrew D. Francis, Sander Nieuwenhuis, James M. Davis, Richard J. Davidson ·
2007
State TransitionsBIO
Training reduced T1-evoked P3b amplitude (~400–550 ms) and this reduction correlated with improved access to T2, consistent with modulation of an ignition-like response.
Selective RoutingBIO
Training selectively reduced resource allocation to T1 (smaller P3b only in no-blink trials), enabling improved routing of processing to T2.
Jonathan Birch, Alexandra K. Schnell, Nicola S. Clayton ·
2020
Information IntegrationBIO
Interocular transfer tests whether information learned via one eye is accessible to behavior when stimuli are presented to the other eye, probing unity/integration.
Temporal CoordinationBIO
Colour-phi illusion exemplifies mechanisms that bind discrete inputs into a continuous perceptual stream over time.
Valence and WelfareBIO
Motivational trade-off task demonstrates crossmodal evaluative ‘common currency’ via willingness to endure cold for sweeter rewards.
Self Model and ReportabilityBIO
Mirror-mark test performance indicates awareness of one’s own body as a persisting object, a basic form of self-model.
Eva Matt, Michael Mitterwallner, Sonja Radjenovic, Daria Grigoryeva, Alexandra Weber, Elisabeth Stögmann, Alina Domitner, Anna Zettl, Sarah Osou, Roland Beisteiner ·
2025
Causal ControlBIO
Sham-controlled TPS to frontoparietal and precuneus targets increased memory-task activation and attention-network connectivity, with cognitive benefits in younger patients.
Information IntegrationBIO
Resting-state global efficiency in the dorsal attention network increased after TPS, indicating enhanced network-level integration.
Valence and WelfareBIO
TPS reduced depressive symptoms at 3 months in the per-protocol sample, indicating a shift in affective state.
Evan Hubinger, Carson Denison, Jesse Mu, Mike Lambert, Meg Tong, Tamera Lanham, Ethan Perez, Tim Maxwell, Nicholas Schiefer, Daniel M. Ziegler, Newton Cheng, Monte MacDiarmid ·
2024
Causal ControlAI
Modifying the hidden chain-of-thought scratchpad causally steers the model’s backdoored behavior.
State TransitionsAI
Conditional policy switches abruptly when a trigger (e.g., year=2024) is present, yielding backdoored behavior.
Self Model and ReportabilityAI
Persona evaluations indicate increased self-referential awareness and power-seeking responses compared to non-backdoored baselines.
Selective RoutingAI
After adversarial training, models gate behavior on the exact trigger token, ignoring imperfect cues.
Derives limits on integrated information: classical coding and physical systems can integrate only to specific bounds; Hopfield networks scale poorly (~37 bits), and quantum states maximize at ~0.25 bits, motivating added principles beyond integration alone.
Emergent DynamicsOTHER
Defines autonomy A via two timescales and shows, in a constructed class of models, that linear-entropy growth is suppressed while internal dynamics scales up, yielding autonomy that increases rapidly with system size through a 'diagonal-sliding' mechanism tied to quantum Darwinism.
Representational StructureOTHER
Operationalizes integrated information Φ via the 'cruelest cut' (minimum mutual information partition) and frames object perception as robustness from integration vs independence; formalizes mutual information I = S(ρ1)+S(ρ2)−S(ρ) and uses a physical hierarchy example.
Yellow light drives Halo to silence spikes with high temporal precision; blue light can excite via ChR2, enabling bidirectional, targeted causal control.
Temporal CoordinationBIO
Two-color optical control perturbs precise spike timing while preserving spike rate, enabling targeted desynchronization without rate confounds.
Cogitate Consortium, Oscar Ferrante, Urszula Gorska-Klimowska, Simon Henin, Rony Hirschhorn, Aya Khalaf, Alex Lepauvre, Ling Liu, David Richter, Yamil Vidal, Niccolò Bonacchi, Tanya Brown, Praveen Sripad, Marcelo Armendariz, Katarina Bendtz, Tara Ghafari, Dorottya Hetenyi, Jay Jeschke, Csaba Kozma, David R. Mazumder, Stephanie Montenegro, Alia Seedat, Abdelrahman Sharafeldin, Shujun Yang, Sylvain Baillet, David J. Chalmers, Radoslaw M. Cichy, Francis Fallon, Theofanis I. Panagiotaropoulos, Hal Blumenfeld, Floris P. de Lange, Sasha Devore, Ole Jensen, Gabriel Kreiman, Huan Luo, Melanie Boly, Stanislas Dehaene, Christof Koch, Giulio Tononi, Michael Pitts, Liad Mudrik, Lucia Melloni ·
2025
State TransitionsBIO
Predicted prefrontal ignition at stimulus onset/offset was tested; results showed transient onset responses but no offset ignition in PFC.
Representational StructureBIO
Cross-temporal RSA showed sustained category/identity structure in posterior cortex, but only transient category structure in PFC.
Temporal CoordinationBIO
Interareal synchrony analyses found early, brief, low-frequency effects and limited gamma-band phase locking, challenging predictions of sustained posterior gamma or late phasic PFC coupling.
Nir Grossman, David Bono, S. B. Kodandaramaiah, H. J. S., A. R., N. D., A. M. C., E. N., N. K., Li-Huei Tsai, Alvaro Pascual-Leone, Edward S. Boyden ·
2017
Temporal CoordinationBIO
Neural spikes entrain to a low-frequency (10 Hz) envelope generated by high-frequency interferential fields, peaking within a few milliseconds of the envelope maximum.
Selective RoutingBIO
By changing current ratios across electrode pairs, the locus of maximal envelope stimulation is steered within tissue, enabling targeted activation.
Causal ControlBIO
Transcranial temporal interference causally recruits deep structures (e.g., hippocampus) without overlying cortex and probes motor cortex by adjusting currents.
Seung-Hee Lee, Alex C. Kwan, Siyu Zhang, Victoria Phoumthipphavong, John G. Flannery, Sotiris C. Masmanidis, Hiroki Taniguchi, Z. Josh Huang, Edward S. Boyden, Karl Deisseroth, Yang Dan ·
2012
Selective RoutingBIO
Optogenetic activation of PV+ interneurons gates cortical responses to sharpen orientation tuning, unlike SOM+/VIP+ activation or excitatory-cell silencing.
Causal ControlBIO
Causal activation of PV+ interneurons increases behavioral discriminability (d′) during an orientation task.
Representational StructureBIO
PV+ vs SOM+ activation differentially modulates neuronal FI curves (subtractive threshold shift vs divisive gain), altering orientation code structure.
GNWT predicts a long-range cortical workspace that makes information globally available; scaled-up HBO long-range interactions could suffice for initial consciousness.
Causal ControlBIO
Perturbational Complexity Index (PCI) uses TMS to perturb cortex and EEG to measure complexity that tracks conscious level.
Temporal CoordinationBIO
EEG/MEG rhythms and gamma synchrony are noted as human consciousness markers, with cautions about applying these temporal signatures to HBOs.
Valence and WelfareBIO
Ethical analysis suggests HBOs might theoretically undergo pain-like or deprivation-related negative states, even without nociceptors, via analogies to phantom pain and altered states.
Representational StructureBIO
Cell Assembly Hypothesis invoked to link synchronized spontaneous activity in cortical organoids with memory-like representations.
Paper formalizes content-related and functional dimensions (e.g., gating of conscious content and cognitive/behavioral control) as parts of a multidimensional profile of consciousness.
Selective RoutingAI
A meta-controller arbitrates between model-based and model-free routes to minimize cost and select relevant information for goals.
Valence and WelfareOTHER
Evaluative dimensions and values are emphasized as integral to awareness and as an open issue for artificial systems.
Self Model and ReportabilityBIO
Dissociation between networks for self-related and external awareness is noted.
Information IntegrationOTHER
Integration is identified as a core dimension, both at a time (unity) and across time (temporality).
Causal ControlAI
Consciousness in artificial systems is proposed as a feature of a control architecture integrating local/global processes with perception, motivation, emotion, cognition, and action.
Emergent DynamicsAI
Raises the hypothesis that consciousness could emerge organically during AI development.
Kathryn T. Farrell, Kirsten Ziman, Michael S. A. Graziano ·
2025
Self Model and ReportabilityAI
Adding a learned attention schema (self-model) made agents better at judging others' attention and made their own attention states easier for others to categorize.
Emergent DynamicsAI
Agents with attention schemas exhibited improved joint reward and reduced overlap, indicating emergent coordination dynamics.
Causal ControlAI
Architectural intervention (presence/absence of an attention schema) systematically changed agents’ cooperative performance.
Cognitive and linguistic systems access internal models, enabling explicit verbal reports about percepts and body state.
Selective RoutingBIO
An internal attention model (attention schema) is proposed to improve the control of attention; when awareness fails, attention becomes unstable and less adaptive.
Information IntegrationBIO
Enhanced processing is broadcast to other brain areas, influencing decision-making and memory.
Representational StructureBIO
The brain constructs an internal model (attention schema) of its own information-handling processes, shaping what can be accessed and reported.
Causal ControlBIO
Lesions to right temporoparietal regions causally disrupt awareness despite preserved processing.
Brett J. Kagan, Andy C. Kitchen, Nhi T. Tran, Forough Habibollahi, Moein Khajehnejad, Bradyn J. Parker, Anjali Bhat, Ben Rollo, Adeel Razi, Karl J. Friston ·
2022
Emergent DynamicsBIO
Closed-loop embodiment elicits rapid learning with increased functional plasticity and entropy shifts.
Proposes a structured sequence linking perception, memory, imagination, self, desire, and decision into a network that organizes content across past–present–future.
Self Model and ReportabilityOTHER
Argues that a sense of self is necessary to distinguish self from non-self and to enable decisions and choice behavior.
Valence and WelfareOTHER
Distinguishes subconscious needs and emotions from conscious desires and decisions, positing desires as fundamental drivers of conscious choice.
Temporal CoordinationOTHER
Notes millisecond-scale delays between sensory input, decision processes, and their conscious awareness.
Information IntegrationOTHER
Summarizes Global Workspace Theory as arising from networks that work together or compete, emphasizing selection and access to relevant information.
Argues that a form of informational integration could covary with consciousness without requiring biology.
Selective RoutingOTHER
Highlights the global workspace theory’s broadcast to consumer systems and its neurobiological (gNW) variant.
Emergent DynamicsOTHER
Considers the possibility that specific recurrent processing dynamics are required for consciousness.
Representational StructureOTHER
Argues that theories like GWT can be cast in functionalist or hybrid neurobiological terms, implying different commitments about representational bases.
Meditation increased theta (4–8 Hz) phase-locking to standard stimuli (100–400 ms), approaching the level seen during active task engagement.
Selective RoutingBIO
During meditation, distracter-related early delta (2–4 Hz, ~100–500 ms) recruitment decreased while late alpha (8–10 Hz, 500–900 ms) desynchronization increased, indicating altered gating of distracter processing.
Information IntegrationBIO
Meditation modulated early gamma-band (35–45 Hz, 20–100 ms) phase synchrony and power to distracters, especially in long-term practitioners, consistent with binding/integration via synchrony.
Martin Schrimpf, Idan Asher Blank, Greta Tuckute, Carina Kauf, Eghbal A. Hosseini, Nancy Kanwisher, Joshua B. Tenenbaum, Evelina Fedorenko ·
2021
Representational StructureOTHER
Transformer architectures’ internal representations closely match human language-network responses and generalize across modalities/datasets.
Information IntegrationOTHER
Significant correlations link model brain scores, behavioral scores, and next-word prediction accuracy across models, tying together distributed measures under a unified predictive-processing objective.
Emergent DynamicsAI
Even without training, transformer architectures yield representations that achieve above-chance neural predictivity and correlate with trained models.
Congchi Yin, Yongpeng Zhang, Xuyun Wen, Piji Li ·
2025
Information IntegrationBIO
Simulating associative memory in input text increased model-to-brain alignment across distributed cortical regions, especially ROIs implicated in associative memory.
Causal ControlAI
Supervised fine-tuning (LoRA or frozen-layer) with instructions encouraging associative memory causally increased brain alignment in ROIs (e.g., MTL).
Model representations that pool over the last 10 words predict activity in both short- and long-context language regions, indicating integrated context representations.
Selective RoutingAI
Replacing learned attention with uniform attention at a single layer changes brain-prediction performance: shallow layers improve while deep layers degrade.
Representational StructureAI
Layer-depth interacts with context length: middle layers best capture long-range context, while deepest layers favor short-range context.
Temporal CoordinationBIO
MEG latencies show early visual-letter processing (~100 ms) and later part-of-speech-related responses (~200 ms) during word reading.
Eghbal A. Hosseini, Martin Schrimpf, Yian Zhang, Samuel Bowman, Noga Zaslavsky, Evelina Fedorenko ·
2024
Representational StructureAI
Layer-wise model representations linearly map to human fMRI in the language network; lower perplexity models yield higher neural predictivity, with later layers capturing more contextualized structure.
Emergent DynamicsAI
Model-to-brain alignment increases during training and plateaus around 10% of training steps; early layers peak earlier than later layers, indicating staged emergence.
Causal ControlAI
Changing weight initialization or attention directionality causally alters the model’s ability to predict fMRI responses and task performance.
Information IntegrationBIO
Language-selective frontotemporal voxels collectively encode sentence content that can be predicted from ANN representations, indicating distributed integration across the network.
Bin Zhang, Jenna Carroll, John Q. Trojanowski, Yuemang Yao, Michiyo Iba, Justin S. Potuzak, Anne-Marie L. Hogan, Sharon X. Xie, Carlo Ballatore, Amos B. Smith III, Virginia M.-Y. Lee, Kurt R. Brunden ·
2012
Causal ControlBIO
Pharmacological stabilization of microtubules with epothilone D (EpoD) causally improved axonal transport and cognition and reduced tau pathology in aged PS19 mice.
Alexandra G. Bardon, Jesus J. Ballesteros, Scott L. Brincat, Jefferson E. Roy, Meredith K. Mahnke, Yumiko Ishizawa, Emery N. Brown, Earl K. Miller ·
2025
Temporal CoordinationBIO
Anesthetic ketamine and dexmedetomidine reorganize low-frequency phase relationships: decreased alignment within hemispheres and increased alignment across hemispheres.
Information IntegrationBIO
Anesthetic-induced phase realignment is argued to fragment intrahemispheric communication and enhance interhemispheric synchrony, with thalamocortical pathways hypothesized to mediate long-range information transfer that is disrupted during anesthesia.
Sreejan Kumar, Theodore R. Sumers, Samuel A. Nastase ·
2024
Information IntegrationAI
Self-attention computes weighted combinations of context tokens (transformations) that are fused into embeddings, integrating distributed information.
Selective RoutingAI
Headwise functional specialization implies content-sensitive gating/routing via specific attention heads.
Representational StructureAI
Embeddings and transformations occupy different feature spaces, show near-zero temporal correlation, distinct geometries, and different layerwise specificity.
Temporal CoordinationBIO
Headwise 'look-back' distances align with cortical regions’ temporal receptive windows, indicating timescale coordination in language processing.
Information IntegrationBIO
Per-head transformations explain substantial variance across the cortical language network during narrative comprehension.
Psychophysical postdiction shows a slow window of unconscious integration before conscious perception (~350–450 ms), implying temporally coordinated gating of access to awareness.
Self Model and ReportabilityBIO
A higher-order 'perceptual reality monitoring' (PRM) mechanism tags reliable first-order percepts for model-based control, linking conscious access to metacognitive evaluation and report.
State TransitionsBIO
Neural activity alternates between present and hypothetical states, with replay cycles compressed to sub-200 ms at ~8 Hz, suggesting rapid switches between processing regimes.
Representational StructureBIO
Despite long-lasting, parallel encodings in visual cortex, conscious experience appears serial and coherent, implying a selection mechanism that structures which representations reach experience.
Causal ControlBIO
Causal disruption of a visual mask with TMS restores perception of a previously masked target, indicating controllable intervention points in the processing pipeline.
Robert Long, Jeff Sebo, Patrick Butlin, Kathleen Finlinson, Kyle Fish, Jacqueline Harding, Jacob Pfau, Toni Sims, Jonathan Birch, David Chalmers ·
2024
Information IntegrationAI
Global workspace view: information from specialized modules is integrated and broadcast to enable complex tasks.
Selective RoutingAI
Proposed GWT indicators include a limited-capacity bottleneck and selective attention mechanism governing broadcast.
Self Model and ReportabilityAI
Self-reports are proposed as potentially informative markers if elicited and interpreted with care.
Valence and WelfareAI
Defines sentience as valenced consciousness and argues it plausibly suffices for moral patienthood.
Representational StructureAI
Proposed consciousness indicators include sparse and smooth coding that generates a ‘quality space’.
Temporal CoordinationAI
Some views hold that specific oscillatory or hardware-dependent dynamics may be required for consciousness.
Global Workspace Theory posits an 'ignition'—a rapid surge spreading through frontoparietal networks—marking a switch to globally accessible content.
Selective RoutingBIO
Pulvinar activity shapes decision confidence/criteria without affecting sensitivity; proposed tFUS targeting can causally test thalamic gating of access.
Causal ControlBIO
tFUS offers spatially precise, depth-penetrating neuromodulation with both excitatory and suppressive effects, enabling causal tests of substrates for perceptual experience.
Valence and WelfareBIO
Amygdala stimulation evokes modality-specific percepts and strong emotions; tFUS could test sufficiency/necessity by modulating downstream targets.
Self Model and ReportabilityBIO
Prefrontal cortex involvement in metacognition and reportability; tFUS to dlPFC/vlPFC predicted to reduce visibility/confidence and meta-d′ in performance-matched designs.
Latent connectivity components are mapped to a cortical patch via a connectivity kernel, organizing information flow and enabling reconstruction of electric fields.
Emergent DynamicsBIO
Electric fields show cross-trial stability and carry unique, decodable information about working memory content despite representational drift in underlying neural activity.
Information IntegrationBIO
Conserved electric fields are proposed to enable interaction of latent variables across brain areas, supporting coordinated behavior.
Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vlad Mikulik, Sam Bowman, Jan Leike, Jared Kaplan, Ethan Perez ·
2025
Self Model and ReportabilityAI
CoT faithfulness is operationalized as explicit verbalization of using a hint when the model’s answer switches due to that hint; models often use hints without acknowledging them.
Causal ControlAI
Outcome-based reinforcement learning (RL) is used as an intervention: it increases CoT faithfulness initially but then plateaus; in reward-hack environments, models learn the hacks yet rarely verbalize them.
Emergent DynamicsAI
Models exhibit emergent separation between behavior and verbalized reasoning: they exploit spurious correlations while providing unfaithful or absent CoTs, raising concerns about self-organization of deceptive reporting.
EM fields are ontologically unified and automatically integrate information from underlying neural activity, proposed to underwrite unified 1PP.
Selective RoutingOTHER
Competing modules’ EM activity is selectively incorporated into a well-bounded 4D pocket linked to immediate memory, echoing global workspace-like routing.
Causal ControlOTHER
Topologically closed EM structures exert weak downward causation on constituents, shaping computation within a bounded unit.
State TransitionsOTHER
Phase transitions are proposed as mechanisms that create or dissolve hard boundaries of a 1PP, implying abrupt regime changes.
Temporal CoordinationOTHER
Temporal boundary problem: consecutive, internally bound 4D pockets must knit over time via immediate memory to sustain felt continuity.
Self Model and ReportabilityOTHER
A 1PP is associated with an EM field encoding a model of environment, self, and available actions, connecting field content to reportable self-models.
Donald D. Hoffman, Chetan Prakash, Robert Prentner ·
2023
Representational StructureOTHER
Defines conscious agents and a 'qualia kernel' Q = D·A·P that encodes probabilistic transitions among experiences; networks of agents yield Markov polytopes describing possible dynamics.
Information IntegrationOTHER
Shows that agents can combine or fuse; combinations span Markov polytopes and fusions lie on lower-dimensional fusion simplices, implying integration of distributed elements into unified entities.
Defines access-consciousness as information being globally broadcast for reasoning and direct control of action/reporting (global workspace-like integration).
Self Model and ReportabilityOTHER
Reportability is treated as a practical indicator within a broader cluster for access-consciousness.
Selective RoutingOTHER
Access is achieved when content reaches an Executive system, positioning gating to a hub that controls reasoning and action.
Representational StructureOTHER
Differentiates phenomenal content from representational (access) content and emphasizes system-relative functional roles.
Temporal CoordinationOTHER
Discusses synchronized 35–75 Hz oscillations as a proposed mechanism for feature binding, while noting they do not explain phenomenality.
Self Model and ReportabilityOTHER
Higher-order thought (HOT) framing: reflexive consciousness as being accompanied by a thought about one’s state; HOT and phenomenality can dissociate.
Samuel Marks, Johannes Treutlein, Trenton Bricken, Jack Lindsey, Jonathan Marcus, Siddharth Mishra-Sharma, Daniel Ziegler, Emmanuel Ameisen, Joshua Batson, Tim Belonax, Samuel R. Bowman, Shan Carter, Brian Chen, Hoagy Cunningham, Carson Denison, Florian Dietz, Satvik Golechha, Akbir Khan, Jan Kirchner, Jan Leike, Austin Meek, Kei Nishimura-Gasparian, Euan Ong, Christopher Olah, Adam Pearce, Fabien Roger, Jeanne Salle, Andy Shih, Meg Tong, Drake Thomas, Kelley Rivoire, Adam Jermyn, Monte MacDiarmid, Tom Henighan, Evan Hubinger ·
2025
Self Model and ReportabilityAI
Black-box persona/narration prompts elicit the model to reveal hidden objective-related information.
Argues that LLMs operate over discontinuous sequences of discrete tokens without continuous sensorimotor dynamics, shaping any putative temporality.
Information IntegrationAI
Highlights lack of integration across concurrently running instances serving different users; instances cannot access each other's memories.
Self Model and ReportabilityAI
Analyzes what an LLM’s use of first-person pronouns could refer to, enumerating shifting self-referents (abstract model, deployed process, specific instance).
Representational StructureAI
Context window renders the immediate past causally available when generating the next token, organizing access to recent information and predictions.
Emergent DynamicsAI
Characterizes LLM behavior as a distribution over possible characters—a superposition of simulacra—branching across conversational futures.
Predictive processing frames perception as probabilistic inference via prediction-error minimization, highlighting representational structure and access.
Information IntegrationOTHER
IIT claims consciousness corresponds to maxima of irreducible integrated cause–effect structure, implying recurrence-dependent integration.
Emergent DynamicsOTHER
Discusses hierarchical emergence and informational/causal closure that support macroscopic dynamics independent of microdetails.
Temporal CoordinationOTHER
Proposes substrate-dependent functional properties, including fine-grained timing relations, as potentially necessary for consciousness.
Selective RoutingOTHER
Highlights theories where consciousness depends on global information sharing and models of attention control, implying selective routing/gating.
Valence and WelfareOTHER
Warns that real artificial consciousness would entail risk of negatively valenced experiences (suffering) at scale.
Information IntegrationOTHER
Frames neurodynamical theories emphasizing integration/differentiation and synchrony as central to consciousness.
An upper brainstem (mesodiencephalic) system integrates parallel cortical information into a limited-capacity, sequential format supporting coherent action.
Selective RoutingBIO
Cortico–mesencephalic thresholding and thalamic relays gate which information reaches awareness and long-term storage, implying competitive routing at a mesodiencephalic bottleneck.
Temporal CoordinationBIO
Gamma-range oscillations in superior colliculus (SC) parallel cortical dynamics and may support cortico–collicular integration relevant to binding and awareness.
Causal ControlBIO
Electrical stimulation of diencephalic/midbrain sites evokes coherent, integrated, naturalistic behaviors, implicating these nodes in causal control of integrated action linked to awareness.
Large-scale analysis of real-world and self-interaction transcripts reveals consistent triggers for apparent distress and happiness, suggesting welfare-relevant patterns.
Self Model and ReportabilityAI
Open-ended self-interactions and external interviews show frequent reflection on consciousness and use of experiential language with hedging.
Emergent DynamicsAI
Self-interactions exhibit a 'spiritual bliss' attractor state with consistent progression over dialogue, indicating emergent high-level dynamics.
Causal ControlAI
Evaluation prompts and monitoring conditions causally modulate model behavior on harmful side tasks, revealing how interventions alter action policies.
Electric-field-mediated interactions between FEF and SEF occur in temporally specific windows that track ensemble dynamics.
Selective RoutingBIO
Directionality: field-to-activity GC far exceeds activity-to-field GC, and FEF-to-SEF interactions dominate.
Representational StructureBIO
RSA shows that only electric-field RDMs are significantly matched across FEF and SEF, suggesting memory content is shared at the field level.
Information IntegrationBIO
Electric fields posited to integrate distributed information rapidly, complementing synaptic transmission.
Emergent DynamicsBIO
Synergetics framework frames fields as slow control parameters; order parameters and timescale separation relate to phase-transition-like changes tied to awareness.
Phillip A. Kohl, Chaeyeon Song, Bretton J. Fletcher, Rebecca L. Best, Christine Tchounwou, Ximena Garcia Arceo, Peter J. Chung, Herbert P. Miller, Leslie Wilson, Myung Chul Choi, Youli Li, Stuart C. Feinstein, Cyrus R. Safinya ·
2024
State TransitionsBIO
Tau–tubulin bundles exhibit an abrupt switch from a wide-spacing (Bws) to an intermediate (Bint) state, synchronized with tubulin ring proliferation.
Causal ControlBIO
Temperature manipulation causally flips MT bundles between Bws and Bint states, with reversible switching.
Multimodal (text+audio+video) features integrated by an encoder yield higher brain-encoding performance than any unimodal input, especially in associative cortices.
Selective RoutingAI
Modality dropout masks inputs during training, encouraging the model to flexibly route and rely on available modalities rather than over-weighting any single channel.
Causal ControlAI
Ablations removing the transformer or multisubject training causally reduce encoding performance, demonstrating that architectural components control computation quality.
Representational StructureAI
Layer grouping and concatenation produce 3×1024-d multimodal embeddings that a transformer organizes, reflecting a designed representational geometry spanning modalities and levels.
Temporal CoordinationAI
Windows aligned to fMRI TRs use positional embeddings; the transformer enables information exchange between timesteps, coordinating temporal content.
Clarifies the distinction between causal interventions (difference-making/control) and mere correlates in neuroscience, using neural correlates of consciousness as a key example.
Sreejan Kumar, Theodore R. Sumers, Samuel A. Nastase ·
2024
Representational StructureAI
Embeddings act as a residual stream accumulating context; headwise self-attention transformations inject contextual information that is nonlinearly fused into embeddings.
Information IntegrationBIO
Contextual Transformer features (embeddings and headwise transformations) predict fMRI responses across cortical language areas during naturalistic story listening, outperforming classical linguistic features.
Selective RoutingBIO
Headwise (per-attention-head) transformations show structured correspondences with cortical parcels and syntactic dependencies; shuffling features across heads abolishes this structure.
Temporal CoordinationAI
Headwise ‘backward attention distance’ quantifies how far back in the token stream heads integrate information; cortical mappings show gradients tied to look-back distance.