A Cortically Inspired Architecture for Modular Perceptual AI

Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ A CORTICALLY INSPIRED ARCHITECTURE FOR MOD- ULAR PERCEPTUAL AI Prerna Luthra Independent Researcher prerna@samvedna.ai ABSTRACT This paper bridges neuroscience and artificial intelligence to propose a cortically inspired blueprint for modular perceptual AI. While current monolithic models such as GPT-4V achieve impressive performance, they often struggle to explic- itly support interpretability, compositional generalization, and adaptive robustness - hallmarks of human cognition. Drawing on neuroscientific models of corti- cal modularity, predictive processing, and cross-modal integration, we advocate decomposing perception into specialized, interacting modules. This architecture supports structured, human-inspired reasoning by making internal inference pro- cesses explicit through hierarchical predictive feedback loops and shared latent spaces. Our proof-of-concept study provides empirical evidence that modular de- composition yields more stable and inspectable representations. By grounding AI design in biologically validated principles, we move toward systems that not only perform well, but also support more transparent and human-aligned inference. 1INTRODUCTION Modern perceptual AI systems spanning computer vision, speech recognition, and multimodal un- derstanding are overwhelmingly built as large monolithic models trained end-to-end (Krizhevsky et al., 2012; Devlin et al., 2019). This monolithic paradigm has delivered remarkable achievements, but it comes with well-documented limitations. Deep networks often operate as opaque “black boxes” that require enormous training data, and can struggle to generalize beyond the narrow sta- tistical patterns of their training distribution (Marcus, 2018; Lake et al., 2017). In particular, such models show brittleness in out-of-distribution scenarios and offer limited transparency or modularity of internal inference. As argued in a public exchange between Gary Marcus and Yoshua Bengio, “expecting a monolithic architecture to handle abstraction and reasoning is unrealistic”. However, much of today’s AI, including the latest foundation models (Bommasani et al., 2021), continues to pursue ever larger end-to-end networks as a one-size-fits-all solution. These limitations motivate the use of explicit cognitive organization as an architectural prior. We posit that the next generation of AI should move beyond unified black-box networks toward modular, cortically inspired systems. In biological brains, perception emerges from an interplay of specialized cortical regions (visual, auditory, language areas, etc.) organized into a deep hierarchy of process- ing stages (Felleman & Van Essen, 1991). Information flows not just bottom-up but also top-down, with higher cortical areas continuously generating predictions to inform lower-level processing—a principle known as predictive coding (Rao & Ballard, 1999; Friston, 2005). This predictive and hierarchical organization enables efficient data usage, generalization, and integration of multiple modalities. Indeed, the brain is often described as a “prediction machine” that actively anticipates sensory inputs (Clark, 2013). These insights from neuroscience motivate us to rethink AI archi- tectures. Instead of colossal homogeneous networks, we advocate modular designs composed of interacting components that mirror the brain’s division of labor and recurrent predictive loops. Some early steps toward modular perceptual AI have shown promise. For example, neural module networks for vision–language tasks demonstrate how separate learned sub-networks can be com- posed to solve complex queries (Andreas et al., 2016). Similarly, mixture-of-experts models and recurrent independent mechanisms introduce sparsely activated sub-networks aimed at capturing the benefits of specialization (Goyal et al., 2019). However, these approaches often remain trained 1 arXiv:2603.07295v1 [cs.AI] 7 Mar 2026 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ end-to-end and lack a unifying cognitively grounded framework that explains how specialization, coordination, and feedback should be structured. In contrast, cortical biology suggests a more principled architectural blueprint: a network of predictive, sparsely communicating components, each specialized for a particular modality or feature, yet operating under a shared predictive coding paradigm (Rao & Ballard, 1999; Friston, 2005). In human perception, multimodal processing arises from such an organization - for example, visual and auditory cortices process inputs separately but converge in higher association areas that resolve cross-modal predictions, allowing flexible integra- tion without collapsing all signals into a single undifferentiated representation (Beauchamp et al., 2004; Calvert et al., 2000) Building on these neuroscientific observations, we propose three core principles for designing cor- tically inspired perceptual AI systems: Modular Specialization and Hierarchical Organization. Perceptual AI should be factorized into semi-independent modules specialized for distinct tasks or modalities (e.g., vision, language, audio subsystems), analogous to functional cortical areas and organized hierarchically to support abstraction and contextual feedback (Mountcastle, 1997; Kanwisher, 2010; Felleman & Van Essen, 1991). Predictive Processing. Rather than purely feed-forward computation, systems should leverage recurrent predictive mechanisms in which components generate top-down expectations and update representations via prediction errors(Rao & Ballard, 1999; Friston, 2005), supporting robustness and uncertainty-sensitive inference (Clark, 2013). Cross-Modal Integration. While modular, perceptual AI should also be integrative. In the brain, distinct cortical regions interact through structured connectivity to produce coherent per- cepts (Sporns & Betzel, 2016). Analogously, specialized AI modules should communicate through well-defined interfaces, allowing distinct modalities to contribute complementary information to shared inference processes. Shared embedding spaces (Radford et al., 2021; Girdhar et al., 2023) or association-like representations can help align representations across modalities, enabling flexible multimodal reasoning while preserving modular structure. In this paper, we advance the case for cortically inspired modular perception and make four contribu- tions. (1) We synthesize neuroscientific and cognitive evidence motivating modular specialization, predictive feedback, and cross-modal integration. (2) We outline an architectural blueprint that op- erationalizes these principles in a modular perceptual system. (3) We present a focused diagnostic proof-of-concept study showing that explicit semantic modularization of internal representations within a large language model improves within-domain feature stability without sacrificing recon- struction fidelity. (4) Finally, we situate this proposal relative to monolithic, mixture-of-experts, and neuro-symbolic approaches, highlighting its implications for interpretability and robustness. 2NEUROSCIENTIFIC MOTIVATION To overcome the limitations of monolithic AI systems, such as brittleness, opacity, and inflexibility, we draw on foundational neuroscientific principles of cortical function. This section synthesizes empirical and conceptual evidence for three core design tenets: modular specialization, predictive feedback, and cross-modal integration. Each principle maps directly onto cortical organization and offers actionable guidance for developing interpretable, and adaptive AI systems. 2.1MODULAR SPECIALIZATION AND ISOLATION OF FUNCTION The mammalian cortex is composed of semi-autonomous modules. Seminal work in neuroscience demonstrates that different cortical regions specialize in distinct functional domains. For instance, the fusiform face area is dedicated to face perception, MT/V5 to motion processing, and V4 to color processing - each revealing finely tuned cortical maps (Kanwisher, 2010; Zeki, 1993). The concept of the cortical column as a canonical microcircuit was famously introduced by Mountcastle (1997), emphasizing the brain’s modular construction. Higher-order cortical association areas also show lateralized functional specialization - for instance, the left prefrontal cortex for language and the 2 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ right for visuospatial tasks. Functional connectivity studies further show that the cortex exhibits hierarchically modular organization: subnetworks for sensory, motor, and cognitive functions oper- ate largely independently but interact through structured interfaces (Sporns & Betzel, 2016). This kind of architectural separation in the brain is associated with robustness, interpretability, and con- tinual learning. Modular organization allows for task-specific updates without global interference, mitigating catastrophic forgetting - a chronic problem in monolithic models. For example, Ellefsen et al. (2015) demonstrated that modular artificial neural networks can outperform monolithic ones in learning new tasks while preserving previously acquired knowledge. In contrast, widely used mono- lithic architectures such as large transformers tend to represent multiple functions within a shared parameter space, leading to entangled internal representations. This limits transparency and makes targeted optimization difficult, as localized updates can produce unintended downstream effects. AI systems inspired by neural modularity, such as Neural Module Networks (Andreas et al., 2016), demonstrate that explicitly separating functional components can support greater interpretability and task efficiency. 2.2PREDICTIVE FEEDBACK AND ACTIVE INFERENCE A defining feature of cortical computation is pervasive recurrent feedback. Predictive coding the- ory posits that the brain continuously generates top-down predictions and updates them through bottom-up prediction errors (Rao & Ballard, 1999; Friston, 2005). Feedback loops refine percep- tion iteratively, producing a dynamic equilibrium between expectation and sensation. Experimental evidence supports this framework. For instance, visual cortex activity (e.g., V1) is suppressed for expected stimuli (Alink et al., 2010), and auditory cortex shows anticipatory activation in response to cued stimuli (Sohoglu et al., 2012). Clark (2013) famously described the brain as a “prediction machine” that reduces surprise via recursive inference. These circuits enable context-sensitive, re- silient interpretation of noisy or incomplete inputs. Most AI models today are feed-forward and lack the ability to refine hypotheses post hoc. This architectural limitation is thought to contribute to hallucinations: confident but incorrect outputs. Unlike the brain, which checks its internal genera- tive outputs against sensory input, AI models lack embedded feedback paths. Incorporating cortical feedback principles into AI could reduce these hallucinations. Feedback-enhanced architectures like recurrent vision transformers (Pan et al., 2022) and recurrent independent mechanisms (Goyal et al., 2019) show promise in handling ambiguity through iterative refinement. These observations moti- vate a more detailed treatment of hallucinations as emergent artifacts of predictive inference, which we revisit in the architectural section. 2.3CROSS-MODAL INTEGRATION AND SEMANTIC CONVERGENCE The brain integrates multimodal data into unified percepts. Association regions such as the superior temporal sulcus (STS) and posterior parietal cortex (PPC) bind information from vision, audition, and language into coherent representations (Beauchamp et al., 2004; Calvert et al., 2000). This fu- sion enhances perception, especially under uncertainty, and operates through temporal coincidence and semantic congruence. Examples include the McGurk effect, where visual cues alter auditory perception, and predictive cross-activation - e.g., speech primes activity in visual cortex (Summer- field & Egner, 2009). Such deep integration is present even in early sensory areas, suggesting an early and recursive multimodal architecture. Most current AI systems treat modalities in isolation or merge them through primarily static embedding alignment (e.g., CLIP-style contrastive objectives (Radford et al., 2021)). While recent systems such as ImageBind extend this alignment across many modalities (Girdhar et al., 2023), they lack explicit reciprocal predictive links across modalities. To more fully emulate brain-like perception, architectures should include modality-specific modules linked through a predictive workspace, allowing cross-modal error checking and redundancy, which are vital for grounding and interpretability. In systems without dynamic, inference-time grounding, hallucinations often stem from language priors overpowering perceptual evidence. Recent studies show that strengthening the influence of visual input during generation in vision–language models reduces hallucination rates (Favero et al., 2024), reinforcing the cortical principle that multimodal grounding serves as a safeguard against spurious outputs. 3 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ 3ARCHITECTURAL BLUEPRINT FOR MODULAR CORTICAL AI Inspired by neuroscience, we propose an AI architecture that decomposes perception and reason- ing into specialized, interacting modules. This design operationalizes three cortical principles – modular specialization, cross-modal integration, and predictive feedback – to enable more robust and interpretable intelligence than monolithic networks. The key components of this architectural blueprint, along with their functional roles and interactions, are described below. Figure 1 provides a schematic illustration of the proposed system. Figure 1: Architectural blueprint for modular perceptual AI, illustrating specialist encoders, a shared multimodal workspace, routing control, and predictive feedback loops. 3.1SPECIALIST ENCODERS FOR EACH MODALITY At the system’s periphery, we employ specialist encoder modules dedicated to particular input modalities or tasks, forming the perceptual front-end of the architecture. Each encoder can be instan- tiated as a pre-trained expert network that transforms raw sensory data into a latent representation optimized for its domain. For example, Whisper can serve as a speech-audio expert for mapping acoustic input into linguistic representations (Radford et al., 2022). Similarly, vision encoders (e.g., convolutional or ViT-based models) transform images into structured visual representations, while large language models such as LLaMA or Vicuna operate as text-based reasoning experts. This separation mirrors cortical specialization, where early sensory areas compute modality-specific rep- resentations prior to downstream integration. This modular approach mirrors the brain’s division into areas specialized for vision, audition, language, etc. Each module can be trained independently, improving robustness and interpretability: a failure in one module does not destabilize the whole, and each expert’s output can be inspected and debugged independently. New capabilities can be added by plugging in a new module without retraining the entire network. 4 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ 3.2SHARED CROSS-MODAL LATENT SPACE To enable coordination across specialized modules, encoder outputs are mapped into a shared latent space inspired by multimodal cortical association areas. This space supports semantic alignment across modalities while preserving modular processing at the periphery. Existing systems provide useful precedents for such shared representations. For example, CLIP jointly trains vision and text encoders to align embeddings of image–caption pairs within a common semantic space (Radford et al., 2021). ImageBind extends this idea to multiple modalities (e.g., depth and audio) by aligning them through shared visual supervision (Girdhar et al., 2023). In our framework, the shared latent space functions as a dynamic workspace for cross-modal semantic convergence rather than a static alignment layer. It enables zero-shot transfer across modalities, supports direct cross-modal compar- isons (e.g., matching sounds to images), and provides a locus for integrating top-down predictions and bottom-up evidence. This design mirrors the role of cortical association regions, which coordi- nate information across sensory systems while enabling flexible, context-dependent integration. 3.3ROUTING CONTROLLER FOR EXPERT SELECTION A routing controller governs which specialist modules are engaged for a given input, task, or in- termediate inference state. This component is inspired by the brain’s ability to flexibly allocate processing across cortical regions based on context and behavioral goals, rather than relying on a fixed processing pipeline. From a computational perspective, this controller may draw on mecha- nisms developed in Mixture-of-Experts (MoE) models, where a sparse subset of experts is activated per input (Shazeer et al., 2017; Lepikhin et al., 2020), or on efficient variants such as the Switch Transformer, which selects a single expert per token (Fedus et al., 2021). However, unlike standard MoE routing, the controller here operates at the level of modular reasoning and perception rather than purely for computational efficiency. Routing decisions can be informed by input modality, task context, or learned properties of the shared latent representation. For example, an image input may first activate a vision encoder, whose latent representation is then passed to a language reasoning module. The language module may, in turn, issue a query that recruits an audio expert or other spe- cialized component. By making such control decisions explicit, this modular orchestration exposes the reasoning chain for inspection while enabling scalability through selective activation rather than monolithic parameter growth. 3.4RECURRENT PREDICTIVE FEEDBACK LOOPS The final architectural ingredient is recurrent feedback across modules. Drawing on predictive cod- ing theories, the system incorporates top-down loops in which higher-level representations generate predictions that constrain lower-level processing (Lee & Mumford, 2003; Friston, 2005). For ex- ample, a language-based inference module may form expectations about semantic or contextual properties of an input (e.g., the likely language being spoken), which can bias downstream audi- tory processing toward compatible interpretations. Through such iterative refinement, modules can resolve ambiguity and improve robustness under noisy or incomplete sensory conditions. Recent architectures such as Recurrent Independent Mechanisms (RIMs) (Goyal et al., 2019) and Recurrent Vision Transformers (Pan et al., 2022) demonstrate how sparse, recurrent interactions allow modules to update their representations based on context and feedback. From this perspective, hallucinations can be reframed as provisional generative hypotheses rather than one-shot failures. Instead of emit- ting a single unconstrained output, a feedback-driven system can iteratively propose, evaluate, and revise predictions until they achieve consistency across modalities and levels of abstraction. This process resembles hypothesis testing in biological perception, where interpretations are continuously adjusted in light of new evidence. 3.5INFORMATION FLOW AND FEEDBACK-DRIVEN REASONING The system operates through an iterative loop: encode→ coordinate→ hypothesize→ feedback. Sensory data are first processed by specialist encoder modules and mapped into a shared latent workspace, where representations from different modalities can be coordinated without collapsing modular structure. A higher-level reasoning module interprets this joint state to form provisional hypotheses about the input and task context. These hypotheses are then broadcast as top-down con- textual signals, prompting lower-level modules to re-evaluate their representations in light of global 5 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ expectations. Each cycle propagates prediction errors across levels, enabling iterative refinement and increased cross-modal consistency. The routing controller facilitates this process by activating relevant modules and supporting repeated interaction until representations stabilize. Through this feedback-driven organization, the architecture supports interpretable and adaptive inference while remaining structurally modular and cognitively grounded, reflecting core principles of cortical in- formation processing. 3.5.1HALLUCINATIONS AS PREDICTIVE ARTIFACTS OF GENERATIVE INFERENCE From a cortical perspective, hallucinations can be understood as byproducts of generative inference rather than purely as errors. Predictive coding and the free-energy principle posit that perception op- erates by minimizing surprise through continuous prediction error reduction (Friston, 2010). When top-down predictions dominate weak or ambiguous sensory input, internally generated hypotheses may persist, manifesting as hallucinations. Closely related phenomena occur in biological systems, including dreaming, expectation-driven illusions, and perceptual filling-in (Lee & Mumford, 2003; Hobson & Friston, 2014), and clinical hallucinations associated with strong priors and weak sensory evidence (Sterzer et al., 2018). Under this view, hallucinations in AI emerge as a predictable consequence of adaptive inference under uncertainty rather than as pathological failures. In modular, feedback-driven systems, such generative hypotheses need not be emitted uncritically. Instead, they can be iteratively evaluated and constrained through recurrent feedback, cross-modal consistency checks, or external grounding sig- nals. Analogous to imagination in biological cognition, internally generated predictions can support creativity and hypothesis formation, while verification mechanisms ensure alignment with sensory evidence and task constraints. Because hypotheses and feedback pathways are modular and explicit, the sources and resolution of hallucinations become inspectable rather than opaque. 4RELATION TO EXISTING ARCHITECTURES 4.1HIGH-PERFORMING MONOLITHIC MODELS AND END-TO-END SUCCESSES Proponents of monolithic AI systems point to the remarkable performance of large end-to-end mod- els across a wide range of tasks. OpenAI’s GPT-4, for example, is a single transformer-based ar- chitecture that demonstrates strong reasoning and language capabilities on numerous benchmarks, including professional and academic examinations (OpenAI, 2023). Similarly, Google DeepMind’s Gemini, a multimodal successor to PaLM-2, achieves competitive results across reasoning, coding, and long-context understanding tasks. In the vision–language domain, models such as Flamingo show that unified architectures can achieve strong few-shot performance on multimodal benchmarks without explicit modular decomposition (Alayrac et al., 2022). Collectively, these systems highlight the empirical effectiveness of large, general-purpose models trained end-to-end. At the same time, despite their performance, such models are not explicitly designed around estab- lished cognitive or neuroscientific principles. Their internal computations remain highly entangled and opaque, motivating interest in complementary architectural paradigms that prioritize modularity, interpretability, and principled inference grounded in models of human cognition. 4.2HYBRID ARCHITECTURES AND NEURO-SYMBOLIC FRAMEWORKS Between fully monolithic and fully modular approaches, a range of hybrid architectures has emerged that blends specialization with end-to-end learning. Neural Module Networks (NMNs), for exam- ple, explicitly compose neural modules corresponding to sub-tasks such as filtering, counting, or relational reasoning. Early NMN systems for visual question answering demonstrated that parsing a query into a layout of sub-tasks enables the dynamic assembly of task-specific networks from a shared module library, yielding improved interpretability and compositional generalization (Andreas et al., 2016). However, these systems typically depend on external mechanisms, such as symbolic parsers, to determine module structure. Google’s Pathways architecture explores a different hybridization strategy by retaining a single large model while enforcing conditional sparse activation. In Pathways-enabled systems, only a subset of 6 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ parameters is activated per input, approximating aspects of specialized processing. Switch Trans- formers and related Mixture-of-Experts (MoE) models instantiate this idea via learned routing over expert sub-networks, achieving strong performance with reduced per-example computation (Fedus et al., 2021). In these models, modularity primarily emerges implicitly at the layer or expert level, while training and inference remain globally end-to-end. Unlike MoE systems, where specialization is largely motivated by computational efficiency, our architecture treats specialization as a represen- tational and inferential prior, with explicit semantic interfaces and recurrent hypothesis validation. A further class of hybrid approaches includes neuro-symbolic architectures, which integrate neu- ral perception with explicit symbolic reasoning or structured knowledge representations (Yi et al., 2018; Mao et al., 2019). By combining statistical learning with symbolic constraints, these sys- tems offer improved interpretability and leverage prior knowledge, aligning with cognitive theories of structured reasoning (Besold et al., 2017). However, they often rely on hand-designed symbolic components or task-specific supervision, limiting scalability. Relatedly, predictive representation learning frameworks such as Joint Embedding Predictive Archi- tectures (JEPA) emphasize learning abstract world models through prediction rather than reconstruc- tion (LeCun, 2022). While JEPA-style models are typically monolithic, their emphasis on internal predictive structure resonates with our focus on hypothesis-driven, feedback-mediated inference. Our proposal extends this predictive perspective by embedding it within an explicitly modular, cor- tically inspired architecture. 5DIAGNOSTIC PROOF-OF-CONCEPT STUDY 5.1OVERVIEW AND ARCHITECTURAL ALIGNMENT To provide an empirical anchor for the proposed cortically inspired modular architecture, we con- duct a focused proof-of-concept (PoC) study examining whether explicit modular decomposition sharpens latent semantic structure within an existing monolithic model. While the full architec- ture described in Section 3 involves specialist encoders, routing controllers, and recurrent predictive feedback, the present experiment isolates a single representational component: modular factoriza- tion of latent features. The goal is not to replicate the full architecture, but to test whether semantic partitioning alone alters the organization, concentration, and stability of internal representations. This experiment should therefore be interpreted as a diagnostic proxy for the modular specializa- tion principle of the broader architecture rather than a complete architectural instantiation. Please note complete methodology, statistical controls, and extended limitations appear in Appendix-A.1. Figures 2-4 in the appendix visualize training convergence, domain clustering, and modular im- provements. Experimental Design. We train sparse autoencoders (SAEs) to factorize Mistral-7B layer-15 ac- tivations (4096-dimensional residual stream, final-token hidden states) from 200 prompts spanning four semantic domains (vision, language, cross-modal, reasoning; 50 prompts each). While this dataset is small relative to typical SAE training regimes, the goal of this experiment is diagnostic rather than benchmark-oriented. Two conditions are compared: Monolithic. Single SAE (4096→1024→4096) trained on all prompts Modular. Four domain-specific SAEs (4096→256→4096 each) with ground-truth semantic rout- ing. We additionally train a capacity-matched monolithic control (256 features) to isolate modularity effects from representational budget constraints. Key Findings: Modular decomposition produces three interpretability-relevant effects: 1. Within-domain stability (+15.4p). Jaccard overlap of active features rises from 55.7% (monolithic, within-domain) to 71.1% (modular), indicating more consistent feature usage. This improvement is robust across all four domains: vision (+15.0p), language (+3.8p), cross-modal (+17.4p), reasoning (+25.4p). 2. Semantic clustering (modest). Observed feature-domain entropy (3.23) is significantly lower than the 100-run shuffled baseline (3.52± 0.01; p ¡ 0.01), indicating reliable concen- tration. However, capacity-matched control (entropy=2.70) reveals this effect is partially explained by representational budget rather than pure modularity. Critically, reduced ca- 7 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ pacity alone does not account for the substantial within-domain stability gains (+15.4p), which occur even at matched capacity. 3. Feature specialization (minimal): Only 6.2% of features are domain-exclusive vs. 5.0% ± 1.0% random baseline, suggesting features remain largely distributed. Reconstruction fidelity is preserved (MSE: 0.0026 vs. 0.0031). Interpretation. Modular decomposition primarily enhances within-domain consistency rather than enforcing hard feature partitioning. This aligns with neuroscientific accounts of cortical special- ization, where functional regions exhibit biased activation without strict exclusivity. The +15p stability improvement is the primary robust effect and suggests that explicit semantic decompo- sition can bias representations toward greater within-domain consistency even with ground-truth routing—motivating investigation of learned routing mechanisms and end-to-end architectural ben- efits. Scope and Limitations. This diagnostic tests only latent feature factorization, not the full proposed architecture (specialist encoders, learned routing, predictive feedback). Key constraints include: (1) capacity confound partially explains entropy effects, (2) small sample size (50/domain) limits generalization claims, (3) ground-truth semantic labels isolate decomposition effects from routing optimization. This choice intentionally upper-bounds the effect of idealized modular routing, estab- lishing whether modularization is even worth learning, and (4) effect sizes are modest for entropy and specialization. 6CONCLUSION We have argued that perceptual AI can benefit substantially from three core cortical princi- ples—modular specialization, predictive feedback, and shared latent spaces - which together support interpretable, adaptive, and robust systems. Decomposing perception into specialized expert units makes errors traceable and correctable at their source. Recurrent feedback loops allow what are tra- ditionally labeled “hallucinations” to function instead as provisional hypotheses, iteratively refined or rejected through internal validation. A shared cross-modal latent space further enables flexible, zero-shot reasoning across vision, language, and other modalities. Such an architecture is particularly well suited to real-world and embodied settings. When sensors fail or inputs are incomplete - as in autonomous driving, robotics, or environmental monitoring - other modules can compensate, while top-down priors support principled inference over missing information, much like human perception reconstructs coherent scenes from partial evidence. By enabling redundancy, introspection, and controlled inference, modularity transforms fragility into resilience. Realizing this vision will require new benchmarks that reward composability and feedback-driven refinement, tools for dynamic routing and module-level auditing, and deeper integration between AI research and neuroscience. Bridging these domains offers a path toward perceptual systems that are not only more capable, but also more transparent, robust, and cognitively grounded. REFERENCES Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Ivan Laptev, Josef Sivic, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. Arjen Alink, Caspar M. Schwiedrzik, Axel Kohler, Wolf Singer, and Lars Muckli. Stimulus pre- dictability reduces responses in primary visual cortex. Journal of Neuroscience, 30(8):2960–2966, 2010. Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Advances in Neural Information Processing Systems 29, p. 39–48. Curran Associates, Inc., 2016. Michael S. Beauchamp, Kathryn E. Lee, Brian D. Argall, and Alex Martin. Integration of auditory and visual information about objects in superior temporal sulcus. Neuron, 41(5):809–823, 2004. 8 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ Tarek R. Besold, Artur d. Garcez, Sebastian Bader, Howard Bowman, Pedro Domingos, Pascal Hitzler, and Kai-Uwe K ̈ uhnberger. Neural-symbolic learning and reasoning: A survey and inter- pretation. arXiv preprint arXiv:1711.03902, 2017. Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Sanjeev Arora, Sydney von Arx, and Percy Liang. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. Gemma A. Calvert, Ruth Campbell, and Michael J. Brammer. Evidence from functional magnetic resonance imaging of crossmodal binding in the human heteromodal cortex. Current Biology, 10 (11):649–657, 2000. Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3):181–204, 2013. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, p. 4171– 4186, 2019. Knut O. Ellefsen, Jean-Baptiste Mouret, and Jeff Clune. Neural modularity helps organisms evolve to learn new skills without forgetting old skills. PLoS Computational Biology, 11(4):e1004128, 2015. Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Poojan Perera, Alessan- dro Achille, Adith Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. arXiv preprint arXiv:2403.14003, 2024. William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961, 2021. Daniel J. Felleman and David C. Van Essen. Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex, 1(1):1–47, 1991. Karl Friston. A theory of cortical responses. Philosophical Transactions of the Royal Society B, 360 (1456):815–836, 2005. Karl Friston. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11 (2):127–138, 2010. Rohit Girdhar, Alaaeldin El-Nouby, Zhiwei Liu, Mannat Singh, K. Vijay Alwala, Armand Joulin, and Ishan Misra.Imagebind: One embedding space to bind them all.arXiv preprint arXiv:2305.05665, 2023. Anirudh Goyal, Alex Lamb, Jordan Hoffmann, Shagun Sodhani, Sergey Levine, Yoshua Bengio, and Bernhard Sch ̈ olkopf. Recurrent independent mechanisms. arXiv preprint arXiv:1909.10893, 2019. J. Allan Hobson and Karl J. Friston. Consciousness, dreams, and inference: the cartesian theatre revisited. Journal of Consciousness Studies, 21(1–2):6–32, 2014. Nancy Kanwisher. Functional specificity in the human brain: a window into the functional archi- tecture of the mind. Proceedings of the National Academy of Sciences, 107(25):11163–11170, 2010. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convo- lutional neural networks. In Advances in Neural Information Processing Systems 25, p. 1097– 1105, 2012. Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. Building machines that learn and think like people. Behavioral and Brain Sciences, 40, 2017. Yann LeCun. A path towards autonomous machine intelligence, 2022. Tai Sing Lee and David Mumford. Hierarchical bayesian inference in the visual cortex. Journal of the Optical Society of America A, 20(7):1434–1448, 2003. 9 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In Proceedings of ICML, p. 5893–5904, 2020. Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The neuro- symbolic concept learner. In Proceedings of ICLR, 2019. Gary Marcus. Deep learning: A critical appraisal. arXiv preprint arXiv:1801.00631, 2018. Vernon B. Mountcastle. The columnar organization of the neocortex. Brain, 120(4):701–722, 1997. OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. Zhiqin Pan, Zhiqiang Xu, Yuning Fang, and Gao Huang. Revisiting vision transformers for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), p. 14849–14859, 2022. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of ICML, p. 8748–8763, 2021. Alec Radford, Heewoo Jeong, Jong Wook Kim, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356, 2022. Rajesh P. N. Rao and Dana H. Ballard. Predictive coding in the visual cortex. Nature Neuroscience, 2(1):79–87, 1999. Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In Proceedings of ICLR, 2017. Elif Sohoglu, Jonathan E. Peelle, Robert P. Carlyon, and Matthew H. Davis. Predictive top-down integration of prior knowledge during speech perception. Journal of Neuroscience, 32(25):8443– 8453, 2012. Olaf Sporns and Richard F. Betzel. Modular brain networks. Annual Review of Psychology, 67: 613–640, 2016. Philipp Sterzer, Rick A. Adams, Paul Fletcher, Chris Frith, Stephen M. Lawrie, Lars Muckli, Predrag Petrovic, Peter Uhlhaas, Matthias Voss, and Philip R. Corlett. The predictive coding account of psychosis. Biological Psychiatry, 84(9):634–643, 2018. Christopher Summerfield and Tobias Egner. Expectation (and attention) in visual cognition. Trends in Cognitive Sciences, 13(9):403–409, 2009. Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Joshua B. Tenenbaum. Neural-symbolic vqa. In Advances in Neural Information Processing Systems 31, p. 1031–1042, 2018. Semir Zeki. A Vision of the Brain. Blackwell Scientific Publications, 1993. AAPPENDIX A.1PROOF-OF-CONCEPT FULL METHODOLOGY A.1.1DATASET: SEMANTIC PROMPT SET We construct a balanced prompt dataset of 200 short text inputs spanning four semantic domains - vision, language, cross-modal, and reasoning - with 50 prompts per domain. Each domain con- tains a mixture of concrete descriptive, abstract explanatory, procedural, and edge or unusual vari- ants in order to elicit heterogeneous internal activations rather than optimize for task performance. The dataset is designed to probe representational diversity and semantic clustering, not downstream accuracy. 10 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ A.1.2MODEL AND ACTIVATION EXTRACTION Model. We use the pretrained Mistral-7B-v0.1 causal language model in evaluation mode with FP16 precision and no fine-tuning. All random seeds are fixed (seed=42) for reproducibility. Activation Extraction. Internal representations are collected from the residual stream output of transformer block layer 15 using a forward hook on model.model.layers[layer idx]. We extract the final-token hidden state for each prompt. This captures the integrated block output following self-attention, MLP, and residual updates rather than a single sublayer write, providing a stable representational snapshot. If no activations are collected, execution halts to avoid silent fallbacks. A.1.3SPARSE AUTOENCODER FACTORIZATION We employ sparse autoencoders (SAEs) as lightweight interpretable factorization mechanisms over the 4096-dimensional activation space. Monolithic SAE. A single encoder–decoder pair with ReLU nonlinearity (4096→ 1024→ 4096) is trained on the full activation set. Capacity-Matched Control. To isolate modularity effects from representational budget, we addi- tionally train a reduced monolithic SAE (4096→ 256→ 4096). Per-Expert SAEs. For the modular condition, activations are partitioned by semantic domain and four independent SAEs are trained with reduced capacity (4096→ 256→ 4096) on∼50 samples each. Routing is based on ground-truth semantic labels rather than learned gating in order to isolate the effect of decomposition itself. Training Details. Optimizer: Adam; learning rate 10 −3 ; epochs: 20; batch size: 8; sparsitypenalty: L1 with λ = 0.01; gradient clipping norm 1.0; implementation: PyTorch on a single T4 GPU. A.1.4DETAILED METRIC DEFINITIONS • Sparsity Metrics: Mean active features per sample and sparsity fraction, measuring repre- sentational efficiency. • Domain–Feature Alignment: Frequency of top-activated features aggregated by semantic domain to identify domain-specific activation patterns. • Entropy Concentration: Shannon entropy of feature distributions across domains; lower entropy indicates features concentrate in fewer dimensions per domain. • Negative Control: Shuffled-label baseline repeated 100 times with different random seeds to estimate null entropy distribution and compute p-values. • Multi-Seed Stability: Jaccard overlap of top features across five independent SAE train- ings to verify features are not random artifacts. • Within-Domain Stability: Pairwise Jaccard overlap of active feature sets within each se- mantic domain; higher overlap indicates consistent feature usage across samples in the same domain. • Feature Specialization (Shared-Space Exclusivity): Fraction of top features (measured in a shared monolithic SAE feature space) that activate uniquely within a single domain; higher values indicate domain-exclusive representations. Measured in shared space to en- able fair comparison with random routing baseline. • Reconstruction Trade-off: Mean squared reconstruction error to ensure modular decom- position does not degrade representational fidelity. A.1.5RESULTS Modular decomposition substantially improves within-domain stability (+15.4p) while maintain- ing comparable reconstruction fidelity as indicated in Table 1 and visualized in Figures 2-4. Figure 2 demonstrates stable training convergence, Figure 3 shows clear domain-specific clustering pat- terns, and Figure 4 illustrates the primary stability improvement (+15.4p) alongside modest spe- 11 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ Table 1: Comparison of monolithic and modular SAE factorization conditions with statistical con- trols. ConditionStabilityEntropyMSE Monolithic (1024 features)0.5573.230.0026 Per-Expert (4×256, ground-truth routing)0.711–0.0031 Controls and Baselines Shuffled Labels (100 runs)– 3.52± 0.01– Capacity-Matched Monolithic (256)–2.700.0033 cialization gains. Capacity-matched control shows entropy reductions are partially explained by representational budget constraints. Note. Entropy is not reported for the per-expert condition because features are partitioned across four independent 256-dimensional SAEs. Within-domain stability is not measured for the capacity- matched control. The entropy reduction (2.70 vs. 3.23) indicates that reduced representational bud- get alone increases feature concentration, partially confounding entropy-based interpretations. Ran- dom routing baseline yields 5.0%± 1.0% specialization versus 6.2% under ground-truth semantic routing (measured in a shared monolithic feature space). Figure 2: SAE training convergence (MSE=0.0026 at epoch 20). Figure 3: Domain-feature clustering: real (left), shuffled control (center), sparsity metrics (right) . 12 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ Figure 4: Modular factorization: feature uniqueness (left), within-domain stability +15.4p (center), reconstruction fidelity (right). Sparse and Domain-Concentrated Representations. The monolithic SAE produces highly sparse encodings (∼25 active features per sample; ∼2.5% sparsity fraction). Feature activation patterns exhibit clear semantic clustering: entropy of the real domain–feature distribution (3.23) is significantly lower than the 100-run shuffled baseline (3.52± 0.01, p < 0.01; 100-run permutation test), indicating statistically reliable, though moderate, non-random concentration of features by do- main. However, the capacity-matched monolithic SAE (256 features) achieves entropy 2.70, lower than the standard condition, indicating that capacity reduction alone increases concentration inde- pendent of modularity. This partially confounds the entropy interpretation and suggests the effect reflects both semantic structure and representational budget constraints. Stability Across Training Runs. Multi-seed analysis yields a mean top-feature Jaccard overlap of 0.364± 0.014 across five independent trainings, suggesting moderate but consistent structural regularities rather than purely random feature allocations. Modular Factorization Effects. When encoded through a shared monolithic SAE feature space (to enable fair comparison with random routing), per-expert specialization is modest: 6.2% unique features versus 5.0% ± 1.0% random baseline (10 runs). This indicates that most features remain distributed rather than strictly partitioned by domain. However, modular decomposition substan- tially increases within-domain stability: average Jaccard overlap rises from 0.557 (monolithic, eval- uated within-domain) to 0.711 (per-expert), an absolute improvement of 15.4 percentage points (p). This improvement is consistent across all four domains: vision (+15.0p), language (+3.8p), cross-modal (+17.4p), and reasoning (+25.4p). Reconstruction loss remains comparable between conditions (monolithic MSE=0.0026, per-expert=0.0031), indicating that increased structural con- sistency does not meaningfully degrade fidelity. Interpretation. Taken together, these results suggest that modular decomposition primarily sharp- ens within-domain reliability and secondarily increases concentration of shared features rather than enforcing hard representational separation. Features remain largely distributed, but become more stable and semantically coherent under explicit partitioning. The primary robust effect is improved within-domain consistency (+15.4p Jaccard), while entropy and specialization effects are modest and partially confounded by capacity constraints. This pattern aligns with neuroscientific accounts of cortical specialization, where functional regions exhibit biased activation and recurrent stabiliza- tion without strict exclusivity. A.1.6DETAILED LIMITATIONS AND SCOPE This PoC study has several constraints that position it as exploratory rather than confirmatory. Capacity Confound. We reduce per-expert SAE capacity (256 vs. 1024 monolithic) to prevent overparameterization in the 50-sample regime. However, the capacity-matched control reveals that reduced capacity alone increases entropy concentration (2.70 vs. 3.23), indicating the entropy ef- fect is partially explained by representational budget constraints rather than pure modularity. While per-expert SAEs show superior within-domain stability at matched capacity, the entropy and special- 13 Accepted to the ICLR 2026 Workshop on “From Human Cognition to AI Reasoning: Models, Methods, and Applications (HCAIR)“ ization metrics are confounded. Future work should match per-sample capacity (e.g., 4×256-feature experts vs. 1024-feature monolithic per-sample capacity) to fully isolate modularity effects. Sample Size and Generalization. With only 50 samples per expert, overfitting risk exists despite capacity reduction and sparsity regularization. We evaluate SAEs on the same activations used for training—standard practice in interpretability research—but cannot validate whether learned fea- tures generalize to novel prompts within each domain. Cross-validation with larger prompt pools (200–500 per domain) is needed to confirm domain-general feature extraction. Ground-Truth Semantic Routing. We use ground-truth semantic labels rather than learned rout- ing to isolate decomposition effects from routing optimization. Real deployments require learn- ing routing functions, which may not achieve perfect semantic alignment and could introduce fail- ure modes not captured in this controlled setting. Investigating learned routing mechanisms (e.g., attention-based gating, gradient-based assignment) is essential for end-to-end architectural valida- tion. Effect Size Magnitude. While statistically robust (p < 0.01, validated across 100 shuffled base- lines), entropy differences are modest (∆ = 0.30, 9% reduction). Feature specialization shows minimal domain-exclusivity (6.2% vs. 5.0% random baseline). The primary robust effect is within- domain stability improvement (+15.4p), suggesting modularity enhances consistency more than exclusivity. This indicates features are biased toward domains rather than exclusive to them. Measurement Framework. Comparison metrics measure different structural properties. Jaccard stability assesses local consistency within domains, while specialization measures global exclusiv- ity across domains. These are complementary but not directly comparable. Additionally, feature interpretability to humans (e.g., via manual inspection or probe classifiers) remains unvalidated; automated metrics do not guarantee semantic meaningfulness or causal disentanglement. Architectural Scope. This tests only latent feature factorization. Specialist encoders, routing con- trollers, and predictive feedback mechanisms remain unvalidated. The PoC demonstrates that ex- plicit decomposition can improve representational properties under ground-truth routing but does not validate end-to-end architectural benefits such as emergent compositionality, transfer learning, or catastrophic forgetting mitigation. These constraints necessitate larger-scale replication with learned routing, held-out evaluation, di- verse prompt distributions, and capacity-matched comparisons before drawing architectural con- clusions. The present work establishes preliminary evidence that modular decomposition enhances within-domain feature consistency while highlighting methodological requirements for stronger val- idation. 14