Integrated information and predictive processing theories of consciousness: An adversarial collaborative review (2509.00555v1)

Abstract: As neuroscientific theories of consciousness continue to proliferate, the need to assess their similarities and differences -- as well as their predictive and explanatory power -- becomes ever more pressing. Recently, a number of structured adversarial collaborations have been devised to test the competing predictions of several candidate theories of consciousness. In this review, we compare and contrast three theories being investigated in one such adversarial collaboration: Integrated Information Theory, Neurorepresentationalism, and Active Inference. We begin by presenting the core claims of each theory, before comparing them in terms of (1) the phenomena they seek to explain, (2) the sorts of explanations they avail, and (3) the methodological strategies they endorse. We then consider some of the inherent challenges of theory testing, and how adversarial collaboration addresses some of these difficulties. More specifically, we outline the key hypotheses that will be tested in this adversarial collaboration, and exemplify how contrasting empirical predictions may pertain to core and auxiliary components of each theory. Finally, we discuss how the data harvested across disparate experiments (and their replicates) may be formally integrated to provide a quantitative measure of the evidential support accrued under each theory. We suggest this approach to theory comparison may afford a useful metric for tracking the amount of scientific progress being made in consciousness research.

• The paper presents an adversarial review comparing IIT, NREP, and AI-C by analyzing their theoretical constructs and empirical predictions using a Bayesian framework.
• The analysis highlights distinct methodological strategies, such as phenomenology-first versus model-first approaches, underscoring challenges in operationalizing consciousness.
• The review demonstrates the potential of Bayesian evidence accumulation to adjudicate between competing neuroscientific theories and guide future research.

Comparative Analysis of Integrated Information, Neurorepresentationalism, and Active Inference Theories of Consciousness

Introduction

This review provides a comprehensive, adversarially collaborative comparison of three leading neuroscientific theories of consciousness: Integrated Information Theory (IIT), Neurorepresentationalism (NREP), and Active Inference (AI-C). The analysis is situated within the context of the INTREPID Consortium, a structured adversarial collaboration under the Accelerating Research on Consciousness (ARC) initiative. The review systematically delineates the core claims, explanatory targets, methodological strategies, and empirical predictions of each theory, and proposes a Bayesian framework for theory adjudication based on multi-experiment evidence accumulation.

Theoretical Foundations

Integrated Information Theory (IIT)

IIT posits that consciousness is identical to the maximally irreducible cause-effect structure instantiated by a physical substrate. The theory is grounded in a set of phenomenological axioms (existence, intrinsicality, information, integration, exclusion, composition) and corresponding physical postulates. The central construct is the quantification of integrated information ($\Phi$), which is intended to capture both the quantity and quality of conscious experience. IIT explicitly rejects functionalism, asserting that input-output equivalence does not guarantee phenomenological equivalence, and instead emphasizes the necessity of specific causal architectures.

Neurorepresentationalism (NREP)

NREP is a predictive processing-based theory that interprets consciousness as arising from hierarchically organized, multimodal neural representations. It is rooted in representationalism, positing that conscious experience is constituted by inferentially constructed world-models that integrate sensory, proprioceptive, and contextual information. NREP emphasizes the role of 'superinference' across multiple representational levels, and highlights the importance of multimodal richness, situatedness, unity, dynamics, and intentionality as haLLMarks of conscious experience. The theory is agnostic regarding the precise computational mechanisms, focusing instead on the representational content and its integration.

Active Inference (AI-C)

AI-C is derived from the free energy principle, providing a normative account of sentient behavior as (approximately) Bayes-optimal belief updating and policy selection. The minimal theory of consciousness implicit in active inference (AI-C) claims that changes in conscious content are necessarily driven by changes in posterior beliefs about hidden states, with a privileged role for the interface between continuous sensorimotor and discrete decision-making hierarchies. AI-C is process-theoretic, emphasizing the computational and inferential machinery underlying conscious access, and is compatible with a broad range of functional architectures.

Comparative Analysis

Explananda

All three theories target phenomenal consciousness, i.e., the qualitative character of experience. IIT and NREP derive their explananda from reflective analysis of subjective experience, while AI-C focuses on the computational properties underlying conscious access and content. Notably, IIT aims to account for both global and local properties of conscious states via a unidimensional $\Phi$ scale, whereas NREP and AI-C reject a simple scalar ordering, instead positing multidimensional or context-dependent gradations.

Explanans

IIT offers an identity theory: conscious experience is the cause-effect structure of the substrate. NREP and AI-C both ground their explanations in inferential architectures, but differ in emphasis. NREP focuses on representational content and its integration, while AI-C foregrounds the process of active inference and policy selection. AI-C uniquely posits that active sampling (overt or covert) is necessary for changes in conscious content, a claim not shared by IIT or NREP.

Methodological Strategies

IIT adopts a phenomenology-first approach, deriving axioms from introspection and operationalizing them in physical terms. NREP blends philosophical representationalism with empirical and computational neuroscience, focusing on the neural basis of representations and their integration. AI-C employs a model-first approach, constructing process-theoretic models to generate and test predictions about conscious phenomena. Each approach entails distinct challenges in translating theoretical constructs into empirically testable predictions, particularly given the underdetermination by auxiliary hypotheses.

Empirical Predictions and Adversarial Testing

The INTREPID Consortium has designed a suite of experiments to test key differentiating predictions:

• Role of Inactive Neurons: IIT predicts that inactivation of already-inactive neurons should alter conscious experience due to changes in the cause-effect structure, a claim not shared by NREP or AI-C. NREP predicts that suppression of background activity (but not inactive neurons per se) can disrupt consciousness, while AI-C is non-committal but allows for indirect effects via precision estimation.
• Cortical Structure and Spatial Experience: IIT predicts that lesions or lacunae in grid-like cortical structures should result in contraction of phenomenal space, whereas NREP and AI-C predict perceptual 'filling-in' or reduced accuracy without systematic bias.
• Active Sampling and Conscious Content: AI-C uniquely predicts that active sampling (e.g., saccades) facilitates changes in conscious content, leading to faster subjective reappearance in motion-induced blindness paradigms. IIT and NREP can accommodate such effects only via auxiliary hypotheses.

These experiments are designed to maximize the evidential contrast between theories, but the review acknowledges the persistent challenge of theory underdetermination due to auxiliary assumptions and methodological limitations.

Bayesian Model Comparison Framework

To address the aggregation of evidence across multiple experiments and sites, the review proposes a Bayesian model comparison framework. Each theory is formalized as a generative model, and model evidence (log marginal likelihood) is computed for each dataset. This approach allows for principled evidence accumulation, explicit weighting of predictions by theorist confidence, and systematic handling of conflicting or non-replicating results. The framework incentivizes precise, testable predictions and mitigates post hoc rationalization.

Implications and Future Directions

The adversarial collaborative approach exemplified by the INTREPID Consortium represents a significant methodological advance in consciousness science, promoting rigorous, bias-minimized theory testing. The Bayesian evidence accumulation strategy provides a scalable, quantitative metric for tracking scientific progress and theory adjudication. However, the review highlights the persistent challenges of auxiliary hypothesis proliferation, operationalization of subjective phenomena, and the need for further refinement of model comparison metrics (e.g., distance from theory core, model families).

Theoretically, the review underscores the deep conceptual divergences between identity-based (IIT) and inferential (NREP, AI-C) accounts of consciousness, as well as the limitations of current empirical paradigms in definitively adjudicating between them. Practically, the proposed framework is extensible to other domains of cognitive neuroscience and may inform the development of consciousness indicators in artificial agents.

Conclusion

This adversarial collaborative review provides a rigorous, systematic comparison of IIT, NREP, and AI-C, clarifying their core claims, empirical predictions, and methodological strategies. The proposed Bayesian framework for evidence accumulation offers a principled approach to theory adjudication in consciousness science. While definitive resolution of theoretical disputes remains elusive, the structured adversarial model and formal analytic tools outlined here represent substantive progress toward an evidence-based, cumulative science of consciousness.