arXiv subjects

Giovanni Pezzulo

Publications and source records attributed to Giovanni Pezzulo.

At least 19 recordsLinked to original sources

arXiv2026

Grounded world models in biological organisms and future embodied AI

Recent advances in generative and embodied AI have been driven by large-scale predictive learning over multimodal data. However, the resulting systems remain largely based on passive training regimes where linguistic regularities create the scaffold onto which information from other modalities is attached. Conversely, neuroscience and cognitive science suggest that biological intelligence is organized in the opposite way, where grounded world models acquired through interaction with the environment provide the semantic scaffold to which language is attached. Here, we illustrate five examples of neural circuits supporting grounded world modelling, which underlie navigation in physical and conceptual spaces, affordance-based perception and interaction with objects, active perception and exploratory learning, allostatic control and emotion, and the distinction between self- and world-generated outcomes. These examples highlight several features largely missing from current embodied AI, including the role of intrinsic dynamics as a foundation for learning, the centrality of action in aligning these dynamics with the external world, the prominence of autonomous experience and open-ended learning over passive assimilation of externally provided data, and the fact that early predictive and control mechanisms scaffold higher cognitive abilities such as reasoning, conceptual navigation, planning, imagination, understanding others' minds, and communication. Finally, we discuss whether and how principles derived from biological systems may inform future embodied AI, including training regimes based on social interaction to construct world models that are not only grounded but also socially shared and aligned with human norms and values.

q-bio.NC↗

arXiv2026

Multi-Agent Digital Twins for Strategic Decision-Making using Active Inference

Active Inference is an emerging framework providing a quantitative account of behavioral processes in neuroscience and a principled approach to decision-making under uncertainty. Its application to agency problems is natural, offering an autopoietic interpretation of action while addressing classical challenges such as the exploration-exploitation trade-off. Recently, Active Inference has been applied to digital twin scenarios for adaptive and predictive modeling of complex systems. In this work, we extend Active Inference to multi-agent digital twins in which agents interact within a shared environment while maintaining decentralized generative models. Our multi-agent framework features two innovations: (i) contextual inference to improve adaptability in dynamic environments, and (ii) the integration of streaming machine learning within agents' generative structures, enabling tunable goal-oriented behavior while preserving efficiency and scalability. The framework is illustrated through a Cournot competition example, providing a digital twin representation of a socio-economic system and highlighting its potential for coordinated decision-making in multi-agent contexts.

cs.CE↗

arXiv2026

Planning difficulty and temporal constraints differently shape the level of detail and ordering of visual exploration

Planning entails identifying sequences of actions to reach a goal, but how much of a future plan is constructed before movement and how visual exploration supports this process remain poorly understood. Here, we investigated how visual exploration before and during action relates to the executed path and how this relationship is shaped by task difficulty and temporal constraints. Participants solved a path-finding task while their gaze and navigation movements were recorded. We manipulated planning difficulty using misleadingness, which captures the mismatch between the perceptual layout of a problem and the structure of its solution, and manipulated temporal constraints by either allowing immediate movement or imposing a temporal delay. We quantified how visual exploration covered the executed path (spatial coverage) and retained its sequential organization (temporal coherence). Greater difficulty increased visual exploration, repeated sampling, and spatial coverage, but was associated with lower temporal coherence with the upcoming path. When participants could act immediately, pre-movement visual exploration showed lower spatial coverage than during execution. Imposing a delay increased pre-movement spatial coverage, bringing it closer to the level observed during execution, but did not increase temporal coherence. The delay was also associated with fewer pauses and backtracks, suggesting that plans were better consolidated before movement, reducing the need for online revision. Individual differences reflected continuous variation in the balance between pre-movement planning and online revision. Together, these findings suggest that participants flexibly allocate visual exploration before and during plan execution, adjusting its level of detail and ordering to problem difficulty and temporal constraints.

q-bio.NC↗

arXiv2026

Bootstrapping Life-Inspired Machine Intelligence: The Biological Route from Chemistry to Cognition and Creativity

Achieving advanced machine intelligence remains a central challenge in AI research, often approached through scaling neural architectures and generative models. However, biological systems offer a broader repertoire of strategies for adaptive, goal-directed behavior - strategies that emerged long before nervous systems evolved. This paper advocates a genuinely life-inspired approach to machine intelligence, drawing on principles from biology that enable robustness, autonomy, and open-ended problem-solving across scales. We frame intelligence as flexible problem-solving, following William James, and develop the concept of "cognitive light cones" to characterize the continuum of intelligence in living systems and machines. We argue that biological evolution has discovered a scalable recipe for intelligence - and the progressive expansion of organisms' "cognitive light cone", predictive and control capacities. To explain how this is possible, we distill five design principles - multiscale autonomy, growth through self-assemblage of active components, continuous reconstruction of capabilities, exploitation of physical and embodied constraints, and pervasive signaling enabling self-organization and top-down control from goals - that underpin life's ability to navigate creatively diverse problem spaces. We discuss how these principles contrast with current AI paradigms and outline pathways for integrating them into future autonomous, embodied, and resilient artificial systems.

q-bio.NC↗

arXiv2026

Schema-based active inference supports rapid generalization of experience and frontal cortical coding of abstract structure

Schemas -- abstract relational structures that capture the commonalities across experiences -- are thought to underlie humans' and animals' ability to rapidly generalize knowledge, rebind new experiences to existing structures, and flexibly adapt behavior across contexts. Despite their central role in cognition, the computational principles and neural mechanisms supporting schema formation and use remain elusive. Here, we introduce schema-based hierarchical active inference (S-HAI), a novel computational framework that combines predictive processing and active inference with schema-based mechanisms. In S-HAI, a higher-level generative model encodes abstract task structure, while a lower-level model encodes spatial navigation, with the two levels linked by a grounding likelihood that maps abstract goals to physical locations. Through a series of simulations, we show that S-HAI reproduces key behavioral signatures of rapid schema-based generalization in spatial navigation tasks, including the ability to flexibly remap abstract schemas onto novel contexts, resolve goal ambiguity, and balance reuse versus accommodation of novel mappings. Crucially, S-HAI also reproduces prominent neural codes reported in rodent medial prefrontal cortex during a schema-dependent navigation and decision task, including task-invariant goal-progress cells, goal-and-spatially conjunctive cells, and place-like codes at the lower level. Taken together, these results provide a mechanistic account of schema-based learning and inference that bridges behavior, neural data, and theory. More broadly, our findings suggest that schema formation and generalization may arise from predictive processing principles implemented hierarchically across cortical and hippocampal circuits, enabling the generalization of experience.

q-bio.NC↗

arXiv2025

A Rate-Distortion Perspective on the Emergence of Number Sense in Unsupervised Generative Models

Number sense is a core cognitive ability supporting various adaptive behaviors and is foundational for mathematical learning. Here, we study its emergence in unsupervised generative models through the lens of rate-distortion theory (RDT), a normative framework for understanding information processing under limited resources. We train $\beta$-Variational Autoencoders -- which embody key formal principles of RDT -- on synthetic images containing varying numbers of items, as commonly used in numerosity perception research. We systematically vary the encoding capacity and assess the models' sensitivity to numerosity and the robustness of the emergent numerical representations through a comprehensive set of analyses, including numerosity estimation and discrimination tasks, latent-space analysis, generative capabilities and generalization to novel stimuli. In line with RDT, we find that behavioral performance in numerosity perception and the ability to extract numerosity unconfounded by non-numerical visual features scale with encoding capacity according to a power law. At high capacity, the unsupervised model develops a robust neural code for numerical information, with performance closely approximating a supervised model explicitly trained for visual enumeration. It exhibits strong generative abilities and generalizes well to novel images, whereas at low capacity, the model shows marked deficits in numerosity perception and representation. Finally, comparison with human data shows that models trained at intermediate capacity levels span the full range of human behavioral performance while still developing a robust emergent numerical code. In sum, our results show that unsupervised generative models can develop a number sense and demonstrate that rate-distortion theory provides a powerful information-theoretic framework for understanding how capacity constraints shape numerosity perception.

q-bio.NC↗

arXiv2025

Gravity Prior and Temporal Horizon Shape Interceptive Behavior under Active Inference

Accurate interception of moving objects, such as catching a ball, requires the nervous system to overcome sensory delays, noise, and environmental dynamics. One key challenge is predicting future object motion in the presence of sensory uncertainty and inherent neural processing latencies. Theoretical frameworks such as internal models and optimal control have emphasized the role of predictive mechanisms in motor behavior. Active Inference extends these ideas by positing that perception and action arise from minimizing variational free energy under a generative model of the world. In this study, we investigate how different predictive strategies and the inclusion of environmental dynamics, specifically an internal model of gravity, influence interceptive control within an Active Inference agent. We simulate a simplified ball-catching task in which the agent moves a cursor horizontally to intercept a parabolically falling object. Four strategies are compared: short temporal horizon prediction of the next position or long horizon estimation of the interception point, each with or without a gravity prior. Performance is evaluated across diverse initial conditions using spatial and temporal error, action magnitude, and movement corrections. All strategies produce successful interception behavior, but those that incorporate gravity and longer temporal horizons outperform others. Including a gravity prior significantly improves spatial and temporal accuracy. Predicting the future interception point yields lower action values and smoother trajectories compared to short-horizon prediction. These findings suggest that internal models of physical dynamics and extended predictive horizons can enhance interceptive control, providing a unified computational account of how the brain may integrate sensory uncertainty, physical expectations, and motor planning.

q-bio.NC↗

arXiv2025

What the flock knows that the birds do not: exploring the emergence of joint agency in multi-agent active inference

Collective behavior pervades biological systems, from flocks of birds to neural assemblies and human societies. Yet, how such collectives acquire functional properties -- such as joint agency or knowledge -- that transcend those of their individual components remains an open question. Here, we combine active inference and information-theoretic analyses to explore how a minimal system of interacting agents can give rise to joint agency and collective knowledge. We model flocking dynamics using multiple active inference agents, each minimizing its own free energy while coupling reciprocally with its neighbors. We show that as agents self-organize, their interactions define higher-order statistical boundaries (Markov blankets) enclosing a ``flock'' that can be treated as an emergent agent with its own sensory, active, and internal states. When exposed to external perturbations (a ``predator''), the flock exhibits faster, coordinated responses than individual agents, reflecting collective sensitivity to environmental change. Crucially, analyses of synergistic information reveal that the flock encodes information about the predator's location that is not accessible to every individual bird, demonstrating implicit collective knowledge. Together, these results show how informational coupling among active inference agents can generate new levels of autonomy and inference, providing a framework for understanding the emergence of (implicit) collective knowledge and joint agency.

nlin.AO↗

arXiv2025

GENESIS: A Generative Model of Episodic-Semantic Interaction

A central challenge in cognitive neuroscience is to explain how semantic and episodic memory, two major forms of declarative memory, typically associated with cortical and hippocampal processing, interact to support learning, recall, and imagination. Despite significant advances, we still lack a unified computational framework that jointly accounts for core empirical phenomena across both semantic and episodic processing domains. Here, we introduce the Generative Episodic-Semantic Integration System (GENESIS), a computational model that formalizes memory as the interaction between two limited-capacity generative systems: a Cortical-VAE, supporting semantic learning and generalization, and a Hippocampal-VAE, supporting episodic encoding and retrieval within a retrieval-augmented generation (RAG) architecture. GENESIS reproduces hallmark behavioral findings, including generalization in semantic memory, recognition, serial recall effects and gist-based distortions in episodic memory, and constructive episodic simulation, while capturing their dynamic interactions. The model elucidates how capacity constraints shape the fidelity and memorability of experiences, how semantic processing introduces systematic distortions in episodic recall, and how episodic replay can recombine previous experiences. Together, these results provide a principled account of memory as an active, constructive, and resource-bounded process. GENESIS thus advances a unified theoretical framework that bridges semantic and episodic memory, offering new insights into the generative foundations of human cognition.

q-bio.NC↗

arXiv2025

Sensory robustness through top-down feedback and neural stochasticity in recurrent vision models

Biological systems leverage top-down feedback for visual processing, yet most artificial vision models succeed in image classification using purely feedforward or recurrent architectures, calling into question the functional significance of descending cortical pathways. Here, we trained convolutional recurrent neural networks (ConvRNN) on image classification in the presence or absence of top-down feedback projections to elucidate the specific computational contributions of those feedback pathways. We found that ConvRNNs with top-down feedback exhibited remarkable speed-accuracy trade-off and robustness to noise perturbations and adversarial attacks, but only when they were trained with stochastic neural variability, simulated by randomly silencing single units via dropout. By performing detailed analyses to identify the reasons for such benefits, we observed that feedback information substantially shaped the representational geometry of the post-integration layer, combining the bottom-up and top-down streams, and this effect was amplified by dropout. Moreover, feedback signals coupled with dropout optimally constrained network activity onto a low-dimensional manifold and encoded object information more efficiently in out-of-distribution regimes, with top-down information stabilizing the representational dynamics at the population level. Together, these findings uncover a dual mechanism for resilient sensory coding. On the one hand, neural stochasticity prevents unit-level co-adaptation albeit at the cost of more chaotic dynamics. On the other hand, top-down feedback harnesses high-level information to stabilize network activity on compact low-dimensional manifolds.

q-bio.NC↗

arXiv2025

Active Digital Twins via Active Inference

Digital twins are transforming engineering and applied sciences by enabling real-time monitoring, simulation, and predictive analysis of physical systems and processes. However, conventional digital twins rely primarily on passive data assimilation, which limits their adaptability in uncertain and dynamic environments. This paper introduces the active digital twin paradigm, based on active inference. Active inference is a neuroscience-inspired Bayesian framework for probabilistic reasoning and predictive modeling that unifies inference, decision-making, and learning under a single free energy minimization objective. By modeling the dynamics of the coupled physical--digital system as a partially observable Markov decision process, active digital twins autonomously balance pragmatic exploitation (maximizing goal-directed utility) and epistemic exploration (actively resolving uncertainty). As action becomes an integral part of the inference process, active digital twins actively seek information to maintain synchronization with, and learn from their physical counterparts. The proposed framework is assessed through virtual experiments of structural health monitoring and predictive maintenance of a railway bridge. The application showcases the step-by-step construction of a generative model enabling bidirectional perception--action interaction. The results demonstrate that active digital twins exhibit superior exploration capabilities compared to traditional reactive approaches, enabling enhanced autonomy and resilience.

cs.CE↗

arXiv2025

Structuring the Environment Nudges Participants Toward Hierarchical Over Shortest Path Planning

Effective planning is crucial for navigating complex environments and achieving goals efficiently. In this study, we investigated how environmental structure influences the selection of planning strategies. Forty-two participants navigated a space station to collect colored spheres, with environments either structured (spheres grouped by color) or unstructured (spheres scattered randomly). We tested three types of plans: hierarchical (grouping spheres by color), shortest path (minimizing travel distance), and neutral (none of the above). By manipulating environmental structure, we were able to nudge participants toward a preference for hierarchical planning in structured environments, while shortest path plans were favored in unstructured environments. A mismatch between self-reported preferences and actual choices indicated that participants often adopted implicit strategies, unaware of their decision-making processes. These findings highlight the powerful effect of environmental cues on planning and suggest that even subtle changes in structure can guide the selection of planning strategies.

q-bio.NC↗

arXiv2024

Predictive Strategies for the Control of Complex Motor Skills: Recent Insights into Individual and Joint Actions

Humans perform exquisite sensorimotor skills, both individually and in teams, from athletes performing rhythmic gymnastics to everyday tasks like carrying a cup of coffee. The "predictive brain" framework suggests that mastering these skills relies on predictive mechanisms, raising the question of how we deploy predictions for real-time control and coordination. This review highlights two research lines, showing that during the control of complex objects people make the interaction with 'tools' predictable; and that during dyadic coordination people make their behavior predictable and legible for their partners. These studies demonstrate that to achieve sophisticated motor skills, we play "prediction tricks": we select subspaces of predictable solutions and make sensorimotor interactions more predictable and legible by and for others. This synthesis underscores the critical role of predictability in optimizing control strategies across contexts. Furthermore, it emphasizes the need for novel studies on the scope and limits of predictive mechanisms in motor control.

q-bio.NC↗

arXiv2024

Human foraging strategies flexibly adapt to resource distribution and time constraints

Foraging is a crucial activity, yet the extent to which humans employ flexible versus rigid strategies remains unclear. This study investigates how individuals adapt their foraging strategies in response to resource distribution and foraging time constraints. For this, we designed a video-game-like foraging task that requires participants to navigate a four-areas environment to collect coins from treasure boxes within a limited time. This task engages multiple cognitive abilities, such as navigation, learning, and memorization of treasure box locations. Findings indicate that participants adjust their foraging strategies -- encompassing both stay-or-leave decisions, such as the number of boxes opened in initial areas and behavioral aspects, such as the time to navigate from box to box -- depending on both resource distribution and foraging time. Additionally, they improved their performance over time as an effect of both enhanced navigation skills and adaptation of foraging strategies. Finally, participants' performance was initially distant from the reward-maximizing performance of optimal agents due to the learning process humans undergo; however, it approximated the optimal agent's performance towards the end of the task, without fully reaching it. These results highlight the flexibility of human foraging behavior and underscore the importance of employing optimality models and ecologically rich scenarios to study foraging.

q-bio.NC↗

arXiv2024

The geometry of efficient codes: how rate-distortion trade-offs distort the latent representations of generative models

Living organisms rely on internal models of the world to act adaptively. These models, because of resource limitations, cannot encode every detail and hence need to compress information. From a cognitive standpoint, information compression can manifest as a distortion of latent representations, resulting in the emergence of representations that may not accurately reflect the external world or its geometry. Rate-distortion theory formalizes the optimal way to compress information while minimizing such distortions, by considering factors such as capacity limitations, the frequency and the utility of stimuli. However, while this theory explains why the above factors distort latent representations, it does not specify which specific distortions they produce. To address this question, here we investigate how rate-distortion trade-offs shape the latent representations of images in generative models, specifically Beta Variational Autoencoders ($\beta$-VAEs), under varying constraints of model capacity, data distributions, and task objectives. By systematically exploring these factors, we identify three primary distortions in latent representations: prototypization, specialization, and orthogonalization. These distortions emerge as signatures of information compression, reflecting the model's adaptation to capacity limitations, data imbalances, and task demands. Additionally, our findings demonstrate that these distortions can coexist, giving rise to a rich landscape of latent spaces, whose geometry could differ significantly across generative models subject to different constraints. Our findings contribute to explain how the normative constraints of rate-distortion theory shape the geometry of latent representations of generative models of artificial systems and living organisms.

q-bio.NC↗

arXiv2023

Transitive inference as probabilistic preference learning

Transitive Inference (TI) is a cognitive task that assesses an organism's ability to infer novel relations between items based on previously acquired knowledge. TI is known for exhibiting various behavioral and neural signatures, such as the Serial Position Effect (SPE), Symbolic Distance Effect (SDE), and the brain's capacity to maintain and merge separate ranking models. We propose a novel framework that casts TI as a probabilistic preference learning task, using one-parameter Mallows models. We present a series of simulations that highlight the effectiveness of our novel approach. We show that the Mallows ranking model natively reproduces SDE and SPE. Furthermore, extending the model using Bayesian selection showcases its capacity to generate and merge ranking hypotheses as pairs with connecting symbols are encountered. Finally, we employ neural networks to replicate Mallows models, demonstrating how this framework aligns with observed prefrontal neural activity during TI. Our innovative approach sheds new light on the nature of TI, emphasizing the potential of probabilistic preference learning for unraveling its underlying neural mechanisms.

q-bio.NC↗

arXiv2023

Neural representation in active inference: using generative models to interact with -- and understand -- the lived world

This paper considers neural representation through the lens of active inference, a normative framework for understanding brain function. It delves into how living organisms employ generative models to minimize the discrepancy between predictions and observations (as scored with variational free energy). The ensuing analysis suggests that the brain learns generative models to navigate the world adaptively, not (or not solely) to understand it. Different living organisms may possess an array of generative models, spanning from those that support action-perception cycles to those that underwrite planning and imagination; namely, from "explicit" models that entail variables for predicting concurrent sensations, like objects, faces, or people - to "action-oriented models" that predict action outcomes. It then elucidates how generative models and belief dynamics might link to neural representation and the implications of different types of generative models for understanding an agent's cognitive capabilities in relation to its ecological niche. The paper concludes with open questions regarding the evolution of generative models and the development of advanced cognitive abilities - and the gradual transition from "pragmatic" to "detached" neural representations. The analysis on offer foregrounds the diverse roles that generative models play in cognitive processes and the evolution of neural representation.

q-bio.NC↗

arXiv2023

Modeling motor control in continuous-time Active Inference: a survey

The way the brain selects and controls actions is still widely debated. Mainstream approaches based on Optimal Control focus on stimulus-response mappings that optimize cost functions. Ideomotor theory and cybernetics propose a different perspective: they suggest that actions are selected and controlled by activating action effects and by continuously matching internal predictions with sensations. Active Inference offers a modern formulation of these ideas, in terms of inferential mechanisms and prediction-error-based control, which can be linked to neural mechanisms of living organisms. This article provides a technical illustration of Active Inference models in continuous time and a brief survey of Active Inference models that solve four kinds of control problems; namely, the control of goal-directed reaching movements, active sensing, the resolution of multisensory conflict during movement and the integration of decision-making and motor control. Crucially, in Active Inference, all these different facets of motor control emerge from the same optimization process - namely, the minimization of Free Energy - and do not require designing separate cost functions. Therefore, Active Inference provides a unitary perspective on various aspects of motor control that can inform both the study of biological control mechanisms and the design of artificial and robotic systems.

q-bio.NC↗