Representational Structure

Conscious-level representations inhabit structured, high-dimensional geometries with distinct roles for different layers; mid-depth, context-integrated states best align with behavior and brain signals.

Supporting Evidence (10)

Direct Fit to Nature: An Evolutionary Perspective on Biological and Artificial Neural Networks (2020) : Shows distributed, high-dimensional embeddings that capture world structure, consistent with geometric coding of content.

Shared functional specialization in transformer-based language models and the human brain (2024) : Demonstrates distinct representational geometries for embeddings vs. transformations and low cross-correlation, indicating separable functional subspaces.

Shared functional specialization in transformer-based language models and the human brain (2024) : Identifies embeddings as residual, context-accumulating spaces and attention-head transformations as context injectors, clarifying layered geometry.

Interpreting and improving natural language processing in machines with natural language processing in the brain (2019) : Finds middle layers best capture long-range context, consistent with mid-depth integration.

Brains and algorithms partially converge in natural language processing (2022) : Shows inverted-U mapping across layers to brain activity; middle layers align best.

Artificial neural network language models predict human brain responses to language even after a developmentally realistic amount of training (2024) : Relates lower perplexity to higher neural predictivity with layer-wise structure, supporting the role of contextualized representations.

TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction (2025) : Builds multimodal embeddings by layer-grouping, preserving low- and high-level features in a designed geometry that predicts fMRI.

In vivo ephaptic coupling allows memory network formation (2023) : RSA reveals matched representational dissimilarities across regions at the electric-field level, implicating a field-defined geometry.

Beyond dimension reduction: Stable electric fields emerge from and allow representational drift (2022) : Connects latent connectivity components to cortical patches via a kernel, reconstructing electric-field structure that organizes information flow.

Fusions of Consciousness (2023) : Formalizes experience transitions in Markov polytopes (qualia kernels), providing an abstract geometric framing of representational dynamics.

Contradictory Evidence (3)

On the Potential of Microtubules for Scalable Quantum Computation (2025) : Posits discrete quDit units in microtubules as fundamental information carriers, a radically different substrate and geometry from distributed embeddings.

Consciousness and Human Brain Organoids: A Conceptual Mapping of Ethical and Philosophical Literature (None) : Links synchronized assemblies in organoids to memory-like representations, suggesting lower-dimensional synchrony as sufficient in some contexts.

The neural architecture of language: Integrative modeling converges on predictive processing (2021) : Architecture alone (with readout) explains substantial brain variance, raising the possibility that some geometry arises from structural priors rather than learned conscious computations.

Access and reportability depend on re-representation, gating, and linear readout of integrated states, not merely on the existence of internal encodings.

Supporting Evidence (7)

Concepts of Consciousness (2002) : Distinguishes A-conscious (representational, reportable) content from phenomenal content, emphasizing the role in reasoning.

The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness (2017) : Proposes an internal attention schema that shapes what can be accessed and reported.

The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention (None) : Implements an attention schema and short-term memory that organize information into report- and control-relevant formats.

Recurrent neural networks with explicit representation of dynamic latent variables can mimic behavioral patterns in a physical inference task (2022) : Finds that linearly decodable latent states predict primate-like behavior, linking decodability to access/action.

Artificial neural network language models predict human brain responses to language even after a developmentally realistic amount of training (2024) : Layer-wise linear mappings from model states to fMRI suggest that accessible, linearly readable structure aligns with brain responses.

Palatable Conceptions of Disembodied Being (2025) : Shows how a context buffer structures what is causally available for the next-token 'report' in LLMs.

Auditing Language Models for Hidden Objectives (2025) : SAE reveals stable features on control tokens that encode reward-model biases, indicating structured representational channels for report behavior.

Contradictory Evidence (3)

Sensory Horizons and the Functions of Conscious Vision (None) : Emphasizes many parallel encodings that do not reach awareness, warning against equating decodability with conscious access.

Key concepts and current views on AI welfare (2025) : Argues that evaluative representations are crucial for sentience/agency, which may not be captured by linear decodability alone.

In vivo ephaptic coupling allows memory network formation (2023) : Shows field-level shared content without establishing its reportability, suggesting distinctions between shared representation and conscious access.

Conscious contents are temporally buffered and serialized from parallel encodings; mid-timescale integrations show higher temporal autocorrelation and better align with report.

Supporting Evidence (7)

Shared functional specialization in transformer-based language models and the human brain (2024) : Embeddings have higher temporal autocorrelation than transformations, indicating temporally smoothed integrative states.

Sensory Horizons and the Functions of Conscious Vision (None) : Highlights tension between parallel encodings and serial experience, implying temporal selection mechanisms.

Palatable Conceptions of Disembodied Being (2025) : Context window makes immediate past causally available for next-token generation, a serializing temporal buffer.

Interpreting and improving natural language processing in machines with natural language processing in the brain (2019) : Middle layers are optimal for long-range context (>15 words), consistent with mid-timescale integration.

The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention (None) : A 10-timestep memory buffer enables trajectory inference, illustrating the role of short-term temporal integration in control.

How do you feel? Interoception: the sense of the physiological condition of the body (2002) : Interoceptive information is re-represented along an anterior gradient, consistent with temporal and hierarchical integration.

Beyond dimension reduction: Stable electric fields emerge from and allow representational drift (2022) : Field reconstructions from latent kernels imply temporally stable patterns that can support serial readout despite unit-level drift.

Contradictory Evidence (2)

Artificial neural network language models predict human brain responses to language even after a developmentally realistic amount of training (2024) : Neural predictivity saturates with perplexity, raising questions about how increased temporal integration translates to measurable brain alignment.

TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction (2025) : Layer-group averaging compresses temporal/hierarchical detail, potentially obscuring fine-grained dynamics even as prediction remains strong.

Cross-area coordination is implemented by mesoscale fields/connectivity kernels in biology and by attention/residual routing in transformers, yielding integrated representational structure.

Supporting Evidence (5)

Beyond dimension reduction: Stable electric fields emerge from and allow representational drift (2022) : Connectivity kernels map latent components to cortical patches to reconstruct electric fields that structure information flow.

In vivo ephaptic coupling allows memory network formation (2023) : Only the electric-field representational dissimilarity matrices matched across regions, suggesting field-level coordination.

Cytoelectric coupling: Electric fields sculpt neural activity and “tune” the brain’s infrastructure (2023) : LFPs influence ensemble selection with storage tied to connectivity patterns, supporting mesoscale coordination mechanisms.

Shared functional specialization in transformer-based language models and the human brain (2024) : Attention-head transformations inject contextual information into a residual stream, paralleling coordination mechanisms.

Brains and algorithms partially converge in natural language processing (2022) : Layered coordination yields best alignment at intermediate depths, consistent with integrated information routing.

Contradictory Evidence (2)

Conscious artificial intelligence and biological naturalism (None) : Predictive processing emphasizes synaptic message passing and error minimization; field mechanisms are not required in such accounts.

In Search of a Biological Crux for AI Consciousness (2024) : Functionalist readings imply that biological specifics (e.g., fields) may be unnecessary if functional roles are preserved.

Evaluative representations modulate which contents enter conscious access and guide report and action.

Supporting Evidence (6)

Key concepts and current views on AI welfare (2025) : Links agency and sentience to evaluative representations, highlighting value-coded structure.

A composite and multidimensional heuristic model of consciousness (2025) : Frames gating and control as functional dimensions of conscious content.

Consciousness defined: requirements for biological and artificial general intelligence (None) : Proposes a network linking perception, memory, self, desire, and decision, embedding evaluative factors.

Conscious artificial intelligence and biological naturalism (None) : Predictive processing casts perception as value-laden inference via prediction-error minimization.

Auditing Language Models for Hidden Objectives (2025) : Reveals stable features encoding reward-model biases at the control token, an evaluative structure shaping report.

How do you feel? Interoception: the sense of the physiological condition of the body (2002) : Interoceptive pathways instantiate homeostatic value signals that are re-represented cortically.

Contradictory Evidence (2)

The neural architecture of language: Integrative modeling converges on predictive processing (2021) : High neural predictivity from architecture and language modeling does not explicitly require evaluative representations.

Consciousness and Human Brain Organoids: A Conceptual Mapping of Ethical and Philosophical Literature (None) : Memory-like synchrony in organoids is discussed without clear evaluative structure, complicating necessity claims.

Architecture and training constrain representational structure: intermediate layers and contextualization increase brain predictivity up to a saturation point.

Supporting Evidence (4)

The neural architecture of language: Integrative modeling converges on predictive processing (2021) : Transformer architecture (GPT-2) explains most explainable neural variance; architecture contributes independently of training.

Artificial neural network language models predict human brain responses to language even after a developmentally realistic amount of training (2024) : Lower perplexity predicts higher brain alignment until a plateau, linking training quality to representational geometry.

Brains and algorithms partially converge in natural language processing (2022) : Inverted-U alignment across layers underscores the architectural placement of optimal representations.

TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction (2025) : Engineered layer-grouped multimodal embeddings improve brain prediction, illustrating designable geometry.

Contradictory Evidence (1)

Recurrent neural networks with explicit representation of dynamic latent variables can mimic behavioral patterns in a physical inference task (2022) : RNNs with explicit latent state representations achieve primate-like behavior, indicating multiple architectures can instantiate suitable geometries.